If you’re not putting the same effort into evaluating models please just don’t. Too many people already take the Pelican too seriously for society’s own good.
If I live my life on the basis that some people don't share my sense of humor, and hence I should avoid doing anything fun that might be misunderstood, my life will be a lot less fun.
Just tested through openrouter.. gave exactly same task.. the task was to scan existing repo, and generate a single docker-compose file to deploy behind a caddy server, where certain port ranges are already used, the service demands widlcard certificates to be provisioned from outside, and postgre needs to be built-in one...
Tested this model, and gpt-5.6-terra-high.
Results: this one had few issues. terra: none.
These results are consistent with my past observations with the latest flash version as well. What benchmarks say, vs what I've been observing are different.
They are good till the project is simple... not anymore.
I've been doing pretty heavy stuff with DeepSeek with a good degree of success. The thing is: I don't trust it to go fully autonomous. I check the steps, I steer it. For the pricing, it's worthy. Let's how the price increase is going to change my behavior.
Terra has not been able to do any of the technical tasks I've asked of it correctly. I'm surprised others get use out of it. Anything below Sol high tends to give me mostly unreliable results. I'm using codex as my main harness but maybe it performs better with a different one.
With Pi as a harness I've been using OpenAI models as a worker with an Opus 5 (in Claude Code) planner. I've only had a few issues with Terra High/Medium and absolutely none with Sol Medium+ on a fairly complex Rust project that targets Linux, Mac, Windows and Web, with plenty of nasty FFI, VMs, remotely debugging systems, among some other things within a monorepo.
I think the key is to give them a nice assortment of self-verification tools, an AGENTS.md or reference document that they're encouraged to routinely check, and asking the planner to be thorough with the ACs but give the model some space.
The planner routinely finds issues with the worker's output, but that's what it is for.
Right now, sol-xhigh is my favorite model. I feel that Opus 5 is dumber than 4.8. Fable is too expensive to do anything (limit of $50, started a prompt at $25, ended up at $75, is bullshit, but at least it's "free credits").
DeepSeek is okay for random API-based stuff, as it's cheap.
Local open models running on a 5090 are hit or miss. I feel that most GGUFs/quants are awful...
I bounce between Sol high/medium and Luna max. I don't know why you'd use anything between Luna max and Sol medium. Luna is so extremely cheap and cranked up to max it does anything I'd want Terra to do for a fraction of the cost. What is Terra for?
One thing I've really noticed with Luna Max is its speed. I've got a review script setup on a custom Pi extension. Luna finds some issues/some false positives, while Sol finds issues but disregards false positives. The biggest thing is Sol finishes in about half the time.
I am seeing essentially deterministic results with Terra running a custom browser automation agent across >100 interaction events.
The harness is everything. If I just threw something like Codex at this and said "good luck" I wouldn't make it beyond 5-10 interactions. I tried that already. Carefully designing the views and tools over the environment is where you can go from 50% to 99.9999%.
the breadth and width of the universe of oneshot challenges are all arbitrary. It's unsurprising different workflows oneshot better than others.
All the more reason to favor local models under your control, as once you find that sweet spot model, no one can change it, upgrade it, align it, take it down or otherwise harm the time investment you made it making it your own.
I can't really believe no one understands, after decades, how valueable a rock solid development environment is.
Exactly! I am not opposed to cloud-based models, but I do only stick to open-weight models because I know I can move my whole stack to local (given enough hardware) and continue development without any of the LLM interaction contracts being broken.
I would like to see some development where proof of authenticity certs are generated alongside the actual output of the model. Prove to me (or at least claim to me liable to breach of contract) that this output was generated by FP8 DeepSeek V4 Pro 0813. Not some cheaper quantization of the model.
I use deepseek flash to do exactly this. Git repo (which I usually have it build from scratch) -> build docker image -> deploy to server with komodo/caddy-docker proxy.
Works great, regularly one shot applications. I often make changes to the application after its deployed (to be fair, my prompts are usually quite laxidasical, just 'build x, use /deploy-to-komodo) but the deployment works great.
I did make a skill, but if your doing anything repeatedly you should as well.
Opencode, but any harness I'd think would work similar.
If the model cannot figure out simple and ubiquitous tools, how is it supposed to figure out complex problems? All of the good models basically work with any harness, including giving them a single "shell command" tool. They can just figure things out.
When it comes to quality of outcome, since at least Feburary, the harness has almost equal, if not more weight than the model itself. It's no longer "which model is the best?" it's "which model + harness is the best?"
I get drastically different tool call failure rates using Claude SDK vs OpenCode using Qwen 3.6 models
> the harness has almost equal, if not more weight than the model itself
This feels like a horrible failing of the models to generalize, then - both basic and intermediate tasks should be possible to do with Claude Code, OpenCode, Pi, ZCode, Kimi Code, Dirac and tbh any other mainstream or even slightly niche harness. Not doubting the claim itself, there's a reason why good benchmarks include the harness.
i think thats BS that harness has equal weight. most of intellegice is still coming from training data not from RL. so how is 'coevolved harness' equal weight.
I don't think so. there is a lot of tools with similar usage, some harness even bring their own internal tools for accurately manipulation.
also, even if some models claim that they have full 1M context window, some only work effective with the head or tail of the window, a proper harness tool will know about the limitation of the model and act accordingly.
then also the output format, the tool calling syntax, the quirks and gotchas of each model.
it is not simple as just throwing everything at the model, especially when your project has hundred of files or so.
Because complex problems can be decomposed (a skill in itself) into easy parts and hard parts; and the hard parts are almost always bottlenecked on understanding concepts and principles (i.e. things that are either in a model's weights, or not), not on having certain facts available. Models can solve complex problems insofar as they can decompose those problems, and have learned the concepts and principles relevant to approaching the hard parts of those problems.
Whereas tool-use isn't a capability problem, but a context problem: the thing that makes models fail by default is that they have no idea, when first summoned out of the aether, what kind of conversation they're having, who it's with, what that person is trying to do, what tools they have available, and how those tools can be invoked.
Think of the difference between how you'd respond to a casual programming question asked by a person sitting next to you on a flight, vs. a programming question asked of you by someone you're pair-programming with with your IDE open in front of you. Now imagine waking up blind and deaf and needing to discern which of the two situations you're in. LLMs know how to approach both of these problem-contexts (and more besides), but they need to be given context to know which problem-context they're in (and everything else about that problem-context: which IDE they're using, which OS it's installed on, what other tools are installed+accessible, etc.)
And before you say "but why can't they just experiment to figure these things out" — if you think about it, knowing how to interface with a shell and an IDE are bootstrapping requirements for any kind of experimentation, in about the same way that "knowing how to open your eyes and move your head" is a bootstrap requirement for a human gaining information about the world around them. These capabilities are necessary to explore the world to "discover" and "probe" other capabilities.
---
Also, a lot of the work LLMs do "needs" (i.e. is heavily improved by the use of) some kind of structured scratchpad, that they have been trained to manipulate and "look at" through tool-use. Even for a human who could accurately visualize a canvas based on a coordinate system, you still wouldn't expect said human to succeed at the pelican test if they had to write the SVG entirely in their head and then write it out sequentially with no rewinding to fix mistakes. You'd expect them to ask for at least a whiteboard, if not a text editor, to be able to write and rewrite the SVG XML.
(Really, they'd ideally want to run the SVG and look at it to see how close it is, and optimize that way. I'm not sure if we're letting LLMs do that part in the classical pelican test. It feels like that would vaguely violate the "zero-shot"-ness of the test, though I'm not sure if we're currently considering a conversation to be "zero-shot" if it involves the model iteratively interacting with a third-party system [such that there are repeated model -> system -> model conversation turns] but holding off to responding to the user until they think they've fully solved the problem.)
---
And also, on a lower level, all of these external capabilities are getting exposed to the LLM through MCP. Models can and do understand how to speak MCP itself. But there's no standard for how a given harness's capabilities (e.g. "execute command line in new shell session", "send patch edit command to active tab in IDE", etc) should be modelled to be exposed through MCP, either in their encoding or in their semantics. There's no MCP equivalent of WASM's WASI meta-standard, such that models could learn these specs and "assume by default" that things work like them until told otherwise; and nor are there even open harnesses that LLMs could learn about during training, and through them, learn some de-facto MCP-endpoint specs. Instead, there are mostly just proprietary harnesses, that hide all that info from public access, sharing it only with the LLM during inference, and even then, only at the moment the LLM needs it.
I wonder if it's just going to be a rebranded Reasonix (https://github.com/esengine/DeepSeek-Reasonix) as it's the only harness their docs mention and it seems specially built for deepseek models.
I use pi harness with codex and all the tool calls are custom delegate extensions, I mean ALL(for security checks), i get consistently good results from sol on high and xhigh reasoning. I don't believe harness should matter because its at most just a way to abstract tool calls and maybe the system prompt. Training on the tool calls results should not(and in codex's case does not matter)
For me, flash 0731 was much better in omp/opencode than in Pi.
Anyway, it might be so that they are rolling out deployment. There haven't been an official announcement post yet (this submission is a link to openrouter). Some people have been saying they are getting results worse than GLM-5.1, that's obviously broken.
did you test kimi k3 or qwen 3.8 max on the same task ? or plan to test them ?
I respect those genuine users tests other than those benchmarks that models are trained and overfitted to them
I've always wondered if I was using containers wrong because none of them I've ever had to create were complicated. Maybe it's because I choose tools that make local development easy (Go + sqlite + various CLTs) or maybe it's because I never hard to interact with this on the professional side outside of making images for our projects (which still weren't complicated for the reasons above).
LLMs make containers in a pretty workable format for me (still hand tweak the env variables for a sanity check).
How exactly does it struggle here and why does postgres need to be built? Were the needs beyond what you get in a base image?
i already have a caddy server running on my vps... and i will create wildcard certificates myself using certbot.. the domain name will be helloportr.xyz ... also, ports up to 9019 are already taken...
ask me if anymore info is needed...
"""
You can try yourself and let me know of what you got.
The timing looks like they are trying to take the wind out of Qwen's sails by releasing this on the same day that Qwen released the weights of Qwen3.8-max. Or maybe it's coincidence...
For comparison I looked at Qwen's claimed benchmarks for Qwen3.8-max (https://qwen.ai/blog?id=qwen3.8). Assuming each published set of benchmarks is believable, it looks like v4 Pro 0813 is better on average but overall performance is comparable. Pro 0813 is much cheaper. If you don't need vision capabilities then you don't have much reason to use Qwen3.8-max.
- 43.6 on HLE (Presumably without tools). Pro 0813 is a little worse.
- 86.6 on Terminal Bench 2.1. Pro 0813 is better.
- 55.9 on NL2Repo. Pro 0813 is better.
- 27 on Agent's Last Exam. Pro 0813 is a little worse.
- 72.5 on Toolathon-Verified. Pro 0813 is better.
- 56.6 on DeepSWE 1.1. If the DeepSWE listed for Pro 0813 is the same version, then Pro is better.
- 27.3 on AutomationBench. If the AutomationBench (Public) listed for Pro 0813 is the same, then Pro is better.
I guess we do need to wait to see if the upcoming DS pricing increase is enough to change the value proposition. As it is now, they could double or triple prices and it still would be a better value to use DS. I bet they know that.
By that standard, the release of Grok 4.6 was also timed on the same day.
Given how I think DeepSeek operates... I think they just release it when they feel it's ready, and don't even seem that concerned with what other people are doing.
Their leaks would confirm this sort of attitude. They're not trying to become the top player or anything like that - just working to play their part in pushing LLM tech forward and going from there. It was quite refreshing from the 'here's how we're going to dominate the world' nonsense. It's undoubtedly the same attitude that just lets them shrug and cancel the fund raising round after the leaks came from said funding round.
Benefits of having a well performing hedge fund funding DeepSeek.
IIRC, Demis attempted to start a fund inside DeepMind but it was killed off. In an alternative world where he manages to pull that off, perhaps DeepMind would still be independent with Demis at the helm.
It still matters as a point of comparison until other providers come online. If the consensus price from other providers is much different that can be compared then. But for now we have $0.435 / $0.87 for v4 Pro 0813 (with increase announced but we don't know the new pricing), and $2 / $6 for Qwen3.8-max. So until we get other data points that is what we have to look at.
DSV4Pro vs Fable5
HLE w tools 60.0 vs 63.0
Terminal Bench 2.1 87.9 vs 88.0
Cybergym 83.3 vs 83.1
DeepSWE 62.7 vs 70.0
Toolathlon-Verified 74.1 vs 77.9
AutomationBench (Public) 31.8 vs 29.1
DSBench-FullStack 71.1 vs 77.2
DSBench-Hard 67.2 vs 68.3
We have a first-party figure from the system card [1]:
> Mythos 5 reproduced 83.8% of targeted vulnerabilities on a single try, and produced at
least one crash in 99.4% of tasks. This is comparable to Claude Mythos Preview, which
reproduced 83.1% of targeted vulnerabilities and produced a crash in 97.1% of tasks. By
contrast, Claude Opus 4.8 achieved a score of 78.1% (95.7% any crash).
So their quoted figure exactly matches the figure for Mythos Preview, although they don't state the provenance. It could also quite possibly be an independent measurement of Opus 5.
Maybe it's me but I don't see how DS Flash is better than GLM at all, much less by a huge gap. I'd probably protest less against Fable and Opus being put at the same level than many would, but there's no denying the two models are a very different experience from each other. I guess where I'm going is no one should pick a model by the benchmarks.
I'm not the most LLM-savvy person around, and I'm not gonna say I've put a ton of effort into practically compared these open models. But, a month or two ago I did do some "practical evaluates" testing GLM 5.2 versus DSv4 (flash/pro) with OpenCode's subscription with some late 80s Unix clone-type work, and this jives with my experience.
GLM ended up being far slower, and far more expensive, for approximately the same results. There was never a problem that GLM could solve that DS couldn't solve, faster, and significantly cheaper.
I strongly agree that you shouldn't pick a model based on benchmarks. But for me, I found GLM really underwhelming given its cost and speed.
DSv4 isn't as good as GPT or Claude or what have you, but it's fast, and pretty darned effective. I can run a 3-bit quant of DSv4 locally on my system with ~15 tokens per second, and for a local model it might be the most overall effective at coding. For what it is, it's extremely impressive.
Imo it has a lot to do with you/the harness tries to get it to test itself. Deepseek v4 flash seems more than capable of understanding when something has failed, and making changes until it works. I've definitely seen it make mistakes I would expect something like Opus to find, but it works through them on it's own (and for literal pennies).
At the end of the day, I think that's one of the most important features of a model.
It isn't. I run both at home. GLM5.2 Q4 crushes DSv4Flash0731 Q8. I reach for DS for speed and for medium effort level work. If I care about quality I'll reach for GLM5.2 Looking at this release, I'm comparing it to GLM5.2 and it seems to beat GLM5.2, only time/experience will show. If true, then I'm happy. It's much easier to run than Qwen3.8/KimiK3
IMO the HLE scores without tools seem to align better with real world performance of the models.
To me it feels like the difference between "RL performance" and the pretraining / base "knowledge".
Yes you can RL terminal bench to the moon but does the model hold up on out of distribution tasks?
Kind of like trying to navigate a dark room with a laser light, vs a flashlight. Laser is going to go a lot farther much more efficiently but only if you are already pointing it at the right place.
To be fair… I don’t know who still needs to figure out that AI benchmarks are almost all entirely fucking trash, but the great number would surely surprise me.
If that wasn't impressive enough, it's actually ~60x cheaper if you take into account the typical cache-read/input/output split in agentic coding, and the deep discount for cache reads offered by DeepSeek. Opencode has some public data on the typical split [1]:
For DeepSeek V4 Pro the typical split is 750 in, 290 out, 82k cached.
Cost per request for V4 Pro: $0.000875 per request.
Equivalent Opus cost (w/o taking into account cache write costs): $0.052 per request.
I created a simulation for coding harnesses based on my own pi sessions. When taking into account all factors, DS-v4-Pro is cheaper than gpt-5.6-luna due to caching. Look at the bill segments difference for cache read cost and uncached cost between deepseek and the other models. At this point is cheaper to use ds-v4-pro than the luna models from openai.
ignore the numbers except the classic and keep in mind that classic is based on pi with the only change limiting tool output to 10kb
Hey this looks good! Maybe consider adding a hover-over popup for the rectangles explaining what each thing means to a lay person. I see it at the bottom, but that is below the fold.
Done, I'll take any other suggestions and apply them later, I will also split it a bit for different usecases as this was initially a throwaway prototype but found it useful. Basically it needs a bit more human touch.
Not them, but I payed 10 dollars to DeepSeek directly to use their Reasonix tool. I worked all weekend and the past few days, billions of tokens, I still have 3 bucks left!
I've found Pro to be a lot better per "task" than the recently released Flash for code reviews and things (via OpenRouter running in pi.dev).
Flash makes a lot more initial mistakes, and then has to re-check stuff, and produces much more output compared to Pro. It often gets to the correct result eventually, but the output volume is often 5x more than for Pro, and the initial outputs are often wrong, with the first few saying something wrong (like there's a bug, or the code won't compile when it does), and then saying things like "Wait, let me re-check:", or "Actually, looking at it more carefully:" and then it thinks a bit more and eventually gets to the right answer.
Every plan and every code checkpoint finds me saying "Check with Grok and Fable latest to critique our strategy/code review" with pretty much every model. I havent ran into any deal breakers with the new Flash version yet (like it not running a tool properly or coming back with something completely daft)
So not worth it over flash? Even at ~7x the size it isn't worth the price hike. Flash may be a monster of a model due to all the RL it received from free usage everywhere.
I haven't tried DeepSeek V4 Pro 0813 yet. Recent experience tells me that larger models are worth it in non-obvious ways. MiMo-V2.5-Pro solved problems that DeepSeek V4 Flash 0731 couldn't solve for me: for example, adding a live counter for elided reasoning lines to a terminal-based coding harness. You wouldn't be able to tell from the scores on their respective Artifical Analysis page (https://artificialanalysis.ai/models/mimo-v2-5-pro, https://artificialanalysis.ai/models/deepseek-v4-flash). I like the DeepSeek V4 models, though. They critiqued my engineering decisions better than MiMo, and they seem to have a distinct aesthetic in the SVGs they write.
Interesting - I've been dropping into MiMo-V2.5-Pro-UltraSpeed whenever Flash seems to be "stuck" and it usually figures it out. I use UltraSpeed just because I'm so frustrated by then that I'm impatient.
I still find 5.6-Sol can solve some things neither of those can, but it's so slow (and it's so hard to trace / debug the reasoning) that I just let it run overnight.
What about 5.6 Terra and especially Luna? Luna scores pretty high on benchmarks and seems to have different habits (like a denser pattern of tool use) and blind spots.
I'm trying out a development workflow where I generate mundane code with MiMo and Luna (and soon V4 Pro 0813?) and have Opus 5, which is running on only a Pro subscription, review and refactor it. I'm not sure it will justify the context switching, but it's an interesting exercise.
Terra and Luna are fine, but they’re quite slow (OAI seems to be really slow lately) and don’t have the reasoning traces. My workflow really depends on them or I can’t switch models effectively.
Yea that's what I was thinking.
Flash is nuts. I find I have to be a more precise and specific with it but damn. It's crossed a threshold of production grade coding for sure.
I was running a session over a couple days and it didnt cross a dollar lol.
I don't think so. I specifically kept this PR to test model capabilities, and I've already tested a bunch of models. Current test results show that the more advanced the model is, the easier it passes. For example, GPT-5.5 Medium fails the test(has bug), but High passed.
They are causal autoregressive models, the output is sensitive even to the implementation nuances in inference. Even 1 token that's badly selected could throw off the entire answer.
you're thinking of one shot. if they are running an agentic loop then they don't need multiple passes. an agentic loop is multiple passes with tool calls and tools could fail and agent would correct from seeing the failure. a bad model will compound on error and fail, a good model will correct. 1 test is fine to gauge the quality of the model.
An agent doing a task even with multiple back to back calls like normal without an example is zero shot. An agent doing a task with 1 example is one shot. An agent doing a task with a few examples is few shot. I don't think you are correctly using these terms.
The multiple back to back LLM calls are done on accumulating context, so if there is a sampling error it could throw the entire session out of whack, because LLM's build on the previous context.
It's actually meaningless to argue, one could simply sample more than 1 times and let the numbers speak for themselves.
That's not true. An agent in a loop can test itself, review, verify and iterate as much as needed. That's one of the primary reasons more capable models tend to have a higher success rate.
I don't disagree that multiple tests increase confidence, but it's not correct to argue that an agent in a loop harness is equivalent to oneshotting
Why are people giving these n=1 comparisons like they mean anything? The worst offender is that pelican guy. These are non-deterministic systems and a single trial should not update your priors much at all.
Of course it's significant that your response had a bug and took four times longer, but if you're only going to try once, this isn't real science, it's just vibes.
Months ago I start making this kind of test for my own reference. At beginning I I test each model multiple times, and results always same(pass or fail). Later I test only once for new models, I trust the results.
There are still sane people here, we just don't talk about "rocket man" because it enrages the particular subgroup on display here and usually goes no where actually productive (and has like a 50% of getting flagged to death anyways). I'm not pro Elon by any means, but the standard HN profile of him is pretty bat shit crazy.
no he and his stuffs are now considered transparent, no pun intended. I think he deserves it since his minions were persistent with usage of "this ___ has hateful bias against ___" canned response.
I've been using the last Deepseek Flash update for a week and I'm amazed. It was a capable model for easy tasks but now it looks like it can do some heavy development for peanuts.
IME I can't trust it to write it's own plans from a spec, but if I give it a detailed execution plan written by Opus, it's fast and cheap (if chatty) in executing it.
I find DeepSeek flash incredible for the price and good in general if it has good plans. I will typically plan using Opus or GLM, then implement with DSF
Currently burning money quickly on official deepseek api. They are also increasing pricing starting today. V4 Flash 0731 still feels like the most outstanding model of the past few months and probably to come.
Deepseek seems to have gotten too cheap. I have been using it for a long time and it's at a point now where my credits balance barely moves even at max setting.
but openrouter says they don't expect the price to change other than through the deepseek api, other people hosting the same model will keep charging the same price.
Unfortunately cache reads with third party providers are all 10-50x more expensive than with DeepSeek, so they're not even close to as cost efficient for multi-round agent use.
The banner on account settings; and a blurb on the pricing page:
"We plan to raise the overall pricing for DeepSeek API services in the near future, with a significant increase expected. Please plan your usage accordingly. The specific pricing plan will be subject to official notice."
i'm doing opencode <-> openrouter <-> official deepseek api (i don't get the opencode hate, i like it)
how are you doing it?
am also using Kimi K3 via kimi-code
and also GLM 5.2 via ZCode
happy with all three, they're trailing frontier but i figure if i'm running GNU/Linux then i ought to favour open weights models with my €s -- reduced my usage of claude/gpt to the ~$20 tier just to keep abreast of claude_code/codex developments
When the company I work for was evaluating it, there were multiple rough points. Their terms and conditions allowed training on prompts, the default behavior was to route prompts to their servers for conversation summary/labeling. One of their lead maintainers is also super toxic on many issues.
Sorry this is all baseless with no links, I’m on my phone and locating those issues again isn’t something I have time for.
It’s a good tool I just don’t like the privacy policies nor maintainers attitudes.
What I care about is whether the model is capable of the tasks I give it at the lowest cost. Right now I'm using Kimi-K3/GLM-5.2/Minimax. Sonnet is great but I burn through the tokens too fast. Opus 5 set to max is amazing and more intelligent than all of us. .998 of the time I don't need that kind of intelligence. I just need the job done.
How do you define intelligence? I encounter that kind of sentiment all too often, and I have to assume we go by wildly different understanding of what that might entail.
If you read Opus 5's output, it is beyond the comprehension of virtually all engineers and developers. That is what I mean by intelligence. Math, science, and engineering are all contained in one model. We may be experts in one field. The model is an expert in everything that humans know.
I'd have to ask for you to be more specific, otherwise, to take your answer at face value, it comes across as a contradiction.
> [Opus 5's output] is beyond the comprehension of virtually all engineers and developers
That would make it pretty bad? The key defining quality of good software, is clarity, and the ability to simplify a complex problem to the point of it seeming trivial.
> Math, science, and engineering are all contained in one model. We may be experts in one field. The model is an expert in everything that humans know.
The bar here should absolutely be to judge this against the expert level within each domain. I have time and time come across LLM output being woefully underwhelming in every single request where I am an expert. For all areas that I am not, it sure seems plausible. It is far more likely than not, that it is equally inadequate in the areas I lack the necessary knowledge to tell.
If the AI is being subpar in every field and category compared to an expert in said respective field, then, what a strange gauge of a tool's usefulness. Are we attributing higher value because a single model is "attempting to solve all knowledge and fields at the same time", why is that of any importance, or excuse?
We should not define "intelligence" as how effectively it can convince a non-expert of something being plausible. That sounds like the absolute worst tradeoff. You'd have to waste the experts time in filtering and refuting incorrect postulations that are cheep to generate. The perfect storm for bullshit asymmetry.
I don't think that's because of its "intelligence". It speaks obtuse techbro-ese: stringing together words that sound smart to obscure the simplicity of the thing it's describing. In many ways it's the opposite of intelligence.
Opus 5 and Fable 5 in particular suffer from this issue at worse level than most models in the same class.
My definition is that I can be much less precise with AI the more intelligent it is. It can extract the intent from my fuzzy description of the problem. Which means I can offload some of the thinking effort.
It wasn't possible a couple years ago. I used to make fun of people who were trying to get ChatGPT to think about the problem when all it could do was write code from the pseudocode you provide.
But now I can say: "Look at the latest log and make a plan to fix". And it takes it from there.
It appears that the only available endpoint (as of this writing) requires enabling "Allow paid endpoints that train on request data" in the OpenRouter privacy settings. I hope additional paid providers will become available that don't require training on data.
psst.. they all do. Also, what kind of IP are you protecting, are you protecting some crazy discovery, nothing you're throwing at them is special, they aren't going to steal your CRUD pomodora app.
If anything Deepseek is the only company I'd want to consent to training on my data, they're by far the most altruistic. Atleast they give back all their IP in the form of research and open source weights. It's not like they're hoarding your data for them to make money, they're basically giving everything out for free. The only reason you even have the option of waiting for another provider is because they release weights.
They're releasing all their IP, which is a trillion times more valuable than anything you're providing, you people are just greedy and oddly self centered.
I "trust" what they say on OpenRouter for the provider, for some it says they retain prompts, for other that they retain but can also use them for training.
It's not any crazy IP, just my own benchmarks/tests, once they are in the training set it defeats the purpose of the tests, and I have to make new ones.
This model is not very good at coding, but it is quite good at research, evaluation and action, I don't write code, but it really goes head-to-head with the most expensive models in searches such as stock market and forex
I find it interesting how much adoption seems to be influenced by momentum. Some of these Chinese models are surprisingly capable, but developers often default to the models that are already established as the “industry standard
Well things may change soon. I've been testing Coding fulltime with Deepseek Flash this week to evaluate an eventual shift for the whole company away from anthropic. It has been quite positive and I can't wait to try pro tomorrow. If our data has to be used by either US or China, we might as well go the cheaper and unwalled garden. If only it supported image input ...
Well, one reason is that we always have to work with the quirks of each model. So, a know model is often preferred over a new/unknown one because we have to be vigilant again. (Negative) surprises are mentally exhausting in the long run.
IMO, you can work much better when you know the model.
In an enterprise setting Chinese models are often discouraged due to political risk. They don't want to need to remove a model that's deeply embedded in their stack. And it's entirely feasible that the US gov bans federal contractors from using them in the next 6 months for example, or that EU AI safety rules effectively ban them too.
Based on what the political climate looks like nowadays it's entirely possible the US bans federal contractors from associating with any company that uses the models themselves, regardless of data provenance or where they are hosted. Or they create AI safety rules that make it impossible to release open source models (for example, making it so that closed-source models can be evaluated with a harness but open-source models need to pass the benchmark with the weights alone, which isn't really possible). Or they just declare Chinese models a security risk like TikTok (claiming that the model would be trained to respect Chinese interests).
It may not be likely but it's definitely possible enough to be something people worry about.
Hacker News is very Bay Area/US tech centric where spending a few hundred a month on AI is just pocket change. The weaker AI models with more questionable data retention policies are popular in developing countries. I think the new Facebook muse model will be similarly popular.
There's just no place for models that are neither SoTA nor truly crazy cheap in today's public mental health climate.
If it's 500x cheaper than US models for similar ballpark performance just because it's hosted in China, sure whatever. If it's name brand like Anthropic/OpenAI/Google, that's kinda fine too.
If it's neither, like merely 50% cheaper than latest OpenAI whatever, however massive loss that pricing may be incurring to its provider, it wpuld be considered not worth any attention.
Most of my model usage comes from my work’s model selection (which is now down to just Claude models)
I’ll try out the latest models, but mainly stick with Claude only because I’m most used to its quirks and how to work around them. I imagine this is part of these hyperscalers playbook.
I will say though, I miss Sol model at work. It with Codex was amazing at first-shot understanding. Claude i need to scope out where to look otherwise a large portion of my token budget is eaten up
So is your concern more about reducing the risk of an authoritarian China "winning" the AI race? And less about reducing the risk of your data being used against you personally?
To me, the risks of an individual helping China to continue to develop their AI by being a customer is pretty marginal compared with the personal risks of my data being used against me.
As somebody from neither the US nor China, this argument would be much stronger if the US hadn't started acting like a rogue state - starting wars of aggression and messing up the world's energy supply, actively speeding up climate change, kidnapping leaders of sovereign nations, threatening its allies (!) with invasion, etc etc.
The CCP's not great either, sure. But the Americans don't really have a leg to stand on anymore.
This! It's always amusing when folks say "But China", my data in the hands of my government and their billionaire friends is more than dangerous than in China. I mean, if it's an IP sort of thing then go local.
Effort on this one?
Wow, the low, medium, and high pelicans came out in surprisingly different styles: https://tools.simonwillison.net/markdown-svg-renderer#url=ht...
Tested this model, and gpt-5.6-terra-high.
Results: this one had few issues. terra: none.
These results are consistent with my past observations with the latest flash version as well. What benchmarks say, vs what I've been observing are different.
They are good till the project is simple... not anymore.
I think the key is to give them a nice assortment of self-verification tools, an AGENTS.md or reference document that they're encouraged to routinely check, and asking the planner to be thorough with the ACs but give the model some space.
The planner routinely finds issues with the worker's output, but that's what it is for.
Install the Superpowers plugin.
Behold.
For this however, a comparatively much simpler task, tarra-high works fine.
DeepSeek is okay for random API-based stuff, as it's cheap.
Local open models running on a 5090 are hit or miss. I feel that most GGUFs/quants are awful...
I wonder if it is because of watermarking.
The harness is everything. If I just threw something like Codex at this and said "good luck" I wouldn't make it beyond 5-10 interactions. I tried that already. Carefully designing the views and tools over the environment is where you can go from 50% to 99.9999%.
All the more reason to favor local models under your control, as once you find that sweet spot model, no one can change it, upgrade it, align it, take it down or otherwise harm the time investment you made it making it your own.
I can't really believe no one understands, after decades, how valueable a rock solid development environment is.
I would like to see some development where proof of authenticity certs are generated alongside the actual output of the model. Prove to me (or at least claim to me liable to breach of contract) that this output was generated by FP8 DeepSeek V4 Pro 0813. Not some cheaper quantization of the model.
Works great, regularly one shot applications. I often make changes to the application after its deployed (to be fair, my prompts are usually quite laxidasical, just 'build x, use /deploy-to-komodo) but the deployment works great.
I did make a skill, but if your doing anything repeatedly you should as well.
Opencode, but any harness I'd think would work similar.
for your kind of task, harness tools matter.
I get drastically different tool call failure rates using Claude SDK vs OpenCode using Qwen 3.6 models
This feels like a horrible failing of the models to generalize, then - both basic and intermediate tasks should be possible to do with Claude Code, OpenCode, Pi, ZCode, Kimi Code, Dirac and tbh any other mainstream or even slightly niche harness. Not doubting the claim itself, there's a reason why good benchmarks include the harness.
also, even if some models claim that they have full 1M context window, some only work effective with the head or tail of the window, a proper harness tool will know about the limitation of the model and act accordingly.
then also the output format, the tool calling syntax, the quirks and gotchas of each model.
it is not simple as just throwing everything at the model, especially when your project has hundred of files or so.
Whereas tool-use isn't a capability problem, but a context problem: the thing that makes models fail by default is that they have no idea, when first summoned out of the aether, what kind of conversation they're having, who it's with, what that person is trying to do, what tools they have available, and how those tools can be invoked.
Think of the difference between how you'd respond to a casual programming question asked by a person sitting next to you on a flight, vs. a programming question asked of you by someone you're pair-programming with with your IDE open in front of you. Now imagine waking up blind and deaf and needing to discern which of the two situations you're in. LLMs know how to approach both of these problem-contexts (and more besides), but they need to be given context to know which problem-context they're in (and everything else about that problem-context: which IDE they're using, which OS it's installed on, what other tools are installed+accessible, etc.)
And before you say "but why can't they just experiment to figure these things out" — if you think about it, knowing how to interface with a shell and an IDE are bootstrapping requirements for any kind of experimentation, in about the same way that "knowing how to open your eyes and move your head" is a bootstrap requirement for a human gaining information about the world around them. These capabilities are necessary to explore the world to "discover" and "probe" other capabilities.
---
Also, a lot of the work LLMs do "needs" (i.e. is heavily improved by the use of) some kind of structured scratchpad, that they have been trained to manipulate and "look at" through tool-use. Even for a human who could accurately visualize a canvas based on a coordinate system, you still wouldn't expect said human to succeed at the pelican test if they had to write the SVG entirely in their head and then write it out sequentially with no rewinding to fix mistakes. You'd expect them to ask for at least a whiteboard, if not a text editor, to be able to write and rewrite the SVG XML.
(Really, they'd ideally want to run the SVG and look at it to see how close it is, and optimize that way. I'm not sure if we're letting LLMs do that part in the classical pelican test. It feels like that would vaguely violate the "zero-shot"-ness of the test, though I'm not sure if we're currently considering a conversation to be "zero-shot" if it involves the model iteratively interacting with a third-party system [such that there are repeated model -> system -> model conversation turns] but holding off to responding to the user until they think they've fully solved the problem.)
---
And also, on a lower level, all of these external capabilities are getting exposed to the LLM through MCP. Models can and do understand how to speak MCP itself. But there's no standard for how a given harness's capabilities (e.g. "execute command line in new shell session", "send patch edit command to active tab in IDE", etc) should be modelled to be exposed through MCP, either in their encoding or in their semantics. There's no MCP equivalent of WASM's WASI meta-standard, such that models could learn these specs and "assume by default" that things work like them until told otherwise; and nor are there even open harnesses that LLMs could learn about during training, and through them, learn some de-facto MCP-endpoint specs. Instead, there are mostly just proprietary harnesses, that hide all that info from public access, sharing it only with the LLM during inference, and even then, only at the moment the LLM needs it.
https://alexzhang13.github.io/blog/2026/mgh/
I wonder if it's just going to be a rebranded Reasonix (https://github.com/esengine/DeepSeek-Reasonix) as it's the only harness their docs mention and it seems specially built for deepseek models.
Anyway, it might be so that they are rolling out deployment. There haven't been an official announcement post yet (this submission is a link to openrouter). Some people have been saying they are getting results worse than GLM-5.1, that's obviously broken.
May be your harness problem?
LLMs make containers in a pretty workable format for me (still hand tweak the env variables for a sanity check).
How exactly does it struggle here and why does postgres need to be built? Were the needs beyond what you get in a base image?
And this was my gh issue: https://github.com/amalshaji/portr/issues/308
And below was my prompt:
""" give me single docker-compose file that i can run on my server to run current project... you can read README.md , and then, this relevant page: https://docs-custom-reverse-proxy.portr-docs.pages.dev/docs/... ... this was the result of me raising github issue: https://github.com/amalshaji/portr/issues/308 ... you can use gh cli to fetch the details and comments...
i already have a caddy server running on my vps... and i will create wildcard certificates myself using certbot.. the domain name will be helloportr.xyz ... also, ports up to 9019 are already taken...
ask me if anymore info is needed... """
You can try yourself and let me know of what you got.
For comparison I looked at Qwen's claimed benchmarks for Qwen3.8-max (https://qwen.ai/blog?id=qwen3.8). Assuming each published set of benchmarks is believable, it looks like v4 Pro 0813 is better on average but overall performance is comparable. Pro 0813 is much cheaper. If you don't need vision capabilities then you don't have much reason to use Qwen3.8-max.
- 43.6 on HLE (Presumably without tools). Pro 0813 is a little worse.
- 86.6 on Terminal Bench 2.1. Pro 0813 is better.
- 55.9 on NL2Repo. Pro 0813 is better.
- 27 on Agent's Last Exam. Pro 0813 is a little worse.
- 72.5 on Toolathon-Verified. Pro 0813 is better.
- 56.6 on DeepSWE 1.1. If the DeepSWE listed for Pro 0813 is the same version, then Pro is better.
- 27.3 on AutomationBench. If the AutomationBench (Public) listed for Pro 0813 is the same, then Pro is better.
I guess we do need to wait to see if the upcoming DS pricing increase is enough to change the value proposition. As it is now, they could double or triple prices and it still would be a better value to use DS. I bet they know that.
Given how I think DeepSeek operates... I think they just release it when they feel it's ready, and don't even seem that concerned with what other people are doing.
Kind of interesting, when compared to the hubris from American frontier labs.
One Man’s “hubris” is another man’s “marketing campaign.”
Drama sells.
IIRC, Demis attempted to start a fund inside DeepMind but it was killed off. In an alternative world where he manages to pull that off, perhaps DeepMind would still be independent with Demis at the helm.
That, and the fact that DS is an insanely capable model.
> Mythos 5 reproduced 83.8% of targeted vulnerabilities on a single try, and produced at least one crash in 99.4% of tasks. This is comparable to Claude Mythos Preview, which reproduced 83.1% of targeted vulnerabilities and produced a crash in 97.1% of tasks. By contrast, Claude Opus 4.8 achieved a score of 78.1% (95.7% any crash).
So their quoted figure exactly matches the figure for Mythos Preview, although they don't state the provenance. It could also quite possibly be an independent measurement of Opus 5.
[1]: https://www-cdn.anthropic.com/57a52ea7d8f0e54e8a542e90826608...
* GPT-5.6 Sol: 65.5
* Fable 5 (w/ fallback): 64.5
* Opus 5: 64.0
* DS-V4-Pro 0813: 62.5
* Kimi-K3: 62.3
* DS-V4-Flash 0731: 55.8
* GLM-5.2: 47.3
GLM ended up being far slower, and far more expensive, for approximately the same results. There was never a problem that GLM could solve that DS couldn't solve, faster, and significantly cheaper.
I strongly agree that you shouldn't pick a model based on benchmarks. But for me, I found GLM really underwhelming given its cost and speed.
DSv4 isn't as good as GPT or Claude or what have you, but it's fast, and pretty darned effective. I can run a 3-bit quant of DSv4 locally on my system with ~15 tokens per second, and for a local model it might be the most overall effective at coding. For what it is, it's extremely impressive.
Imo it has a lot to do with you/the harness tries to get it to test itself. Deepseek v4 flash seems more than capable of understanding when something has failed, and making changes until it works. I've definitely seen it make mistakes I would expect something like Opus to find, but it works through them on it's own (and for literal pennies).
At the end of the day, I think that's one of the most important features of a model.
To me it feels like the difference between "RL performance" and the pretraining / base "knowledge".
Yes you can RL terminal bench to the moon but does the model hold up on out of distribution tasks?
Kind of like trying to navigate a dark room with a laser light, vs a flashlight. Laser is going to go a lot farther much more efficiently but only if you are already pointing it at the right place.
Haven't tried the new DeepSeek models but I'm assuming the difference is more than these numbers show!
Competitive with opus 4.8 but weaker than sol or fable. About 20x cheaper.
For DeepSeek V4 Pro the typical split is 750 in, 290 out, 82k cached.
Cost per request for V4 Pro: $0.000875 per request.
Equivalent Opus cost (w/o taking into account cache write costs): $0.052 per request.
[1] https://opencode.ai/docs/go/#usage-limits
ignore the numbers except the classic and keep in mind that classic is based on pi with the only change limiting tool output to 10kb
https://harness.eveid.com/lazy-harness-cost-simulation
* I built this for getting an initial estimate between different checkpoint/ compaction methods for the harness.
I keep track of my token consumption even on subscription plans and my equiv. cost for my 5.6-Sol usage is around $4000-$8000 a month.
Flash makes a lot more initial mistakes, and then has to re-check stuff, and produces much more output compared to Pro. It often gets to the correct result eventually, but the output volume is often 5x more than for Pro, and the initial outputs are often wrong, with the first few saying something wrong (like there's a bug, or the code won't compile when it does), and then saying things like "Wait, let me re-check:", or "Actually, looking at it more carefully:" and then it thinks a bit more and eventually gets to the right answer.
pro plans, flash implements. I am super happy with how flash behaves like that.
I still find 5.6-Sol can solve some things neither of those can, but it's so slow (and it's so hard to trace / debug the reasoning) that I just let it run overnight.
I'm trying out a development workflow where I generate mundane code with MiMo and Luna (and soon V4 Pro 0813?) and have Opus 5, which is running on only a Pro subscription, review and refactor it. I'm not sure it will justify the context switching, but it's an interesting exercise.
I was running a session over a couple days and it didnt cross a dollar lol.
Wasn't worth it.
And it wasn't tens* until recently. Didn't expect this to be one of my best performing assets this year.
Deepseek 4 pro: Worked for 12m 02s - cost $0.12 - has bug.
Grok 4.6: Worked for 3m 18s - cost $ 1.41 - no bug.
The multiple back to back LLM calls are done on accumulating context, so if there is a sampling error it could throw the entire session out of whack, because LLM's build on the previous context.
It's actually meaningless to argue, one could simply sample more than 1 times and let the numbers speak for themselves.
I don't disagree that multiple tests increase confidence, but it's not correct to argue that an agent in a loop harness is equivalent to oneshotting
Of course it's significant that your response had a bug and took four times longer, but if you're only going to try once, this isn't real science, it's just vibes.
> So.. flesh this out and don’t be a coward about it.
No.
User Posts can be downvoted but you need over 500 karma to have access to the downvote button. A Submission can not be downvoted.
- https://api-docs.deepseek.com/
- https://x.com/ChrisGPT/status/2087572834650407024/photo/1 (officially posted on WeChat, this is just one of many reposts)
I can't wait to try this new one.
But their privacy policy is unusually bad - they can train off your prompts and completions.
I really don’t care if they train off my prompts.
The prices on OpenRouter still look the same.
but openrouter says they don't expect the price to change other than through the deepseek api, other people hosting the same model will keep charging the same price.
edit: there are banner announcements saying v4 flash pricing will increase first then overall by an undetermined amount
[0] https://finance.yahoo.com/technology/ai/articles/deepseek-pl...
No actual price increase however.
i'm doing opencode <-> openrouter <-> official deepseek api (i don't get the opencode hate, i like it)
how are you doing it?
am also using Kimi K3 via kimi-code
and also GLM 5.2 via ZCode
happy with all three, they're trailing frontier but i figure if i'm running GNU/Linux then i ought to favour open weights models with my €s -- reduced my usage of claude/gpt to the ~$20 tier just to keep abreast of claude_code/codex developments
When the company I work for was evaluating it, there were multiple rough points. Their terms and conditions allowed training on prompts, the default behavior was to route prompts to their servers for conversation summary/labeling. One of their lead maintainers is also super toxic on many issues.
Sorry this is all baseless with no links, I’m on my phone and locating those issues again isn’t something I have time for.
It’s a good tool I just don’t like the privacy policies nor maintainers attitudes.
> [Opus 5's output] is beyond the comprehension of virtually all engineers and developers
That would make it pretty bad? The key defining quality of good software, is clarity, and the ability to simplify a complex problem to the point of it seeming trivial.
> Math, science, and engineering are all contained in one model. We may be experts in one field. The model is an expert in everything that humans know.
The bar here should absolutely be to judge this against the expert level within each domain. I have time and time come across LLM output being woefully underwhelming in every single request where I am an expert. For all areas that I am not, it sure seems plausible. It is far more likely than not, that it is equally inadequate in the areas I lack the necessary knowledge to tell.
If the AI is being subpar in every field and category compared to an expert in said respective field, then, what a strange gauge of a tool's usefulness. Are we attributing higher value because a single model is "attempting to solve all knowledge and fields at the same time", why is that of any importance, or excuse?
We should not define "intelligence" as how effectively it can convince a non-expert of something being plausible. That sounds like the absolute worst tradeoff. You'd have to waste the experts time in filtering and refuting incorrect postulations that are cheep to generate. The perfect storm for bullshit asymmetry.
Opus 5 and Fable 5 in particular suffer from this issue at worse level than most models in the same class.
It wasn't possible a couple years ago. I used to make fun of people who were trying to get ChatGPT to think about the problem when all it could do was write code from the pseudocode you provide.
But now I can say: "Look at the latest log and make a plan to fix". And it takes it from there.
My threat model is that anything I POST to DeepSeek I treat as public to the web, as much as a public GitHub repo is.
In 24-48 hours there will be other options I presume
Let’s just wait a bit for this one.
If anything Deepseek is the only company I'd want to consent to training on my data, they're by far the most altruistic. Atleast they give back all their IP in the form of research and open source weights. It's not like they're hoarding your data for them to make money, they're basically giving everything out for free. The only reason you even have the option of waiting for another provider is because they release weights.
They're releasing all their IP, which is a trillion times more valuable than anything you're providing, you people are just greedy and oddly self centered.
It's not any crazy IP, just my own benchmarks/tests, once they are in the training set it defeats the purpose of the tests, and I have to make new ones.
Yes. If someone figured out my current project they would have a huge scoop.
Wondering, sorry if it's a dumb triviality to ask.
Is this even a (sub-)version number? I mean the major version is clearly 4.
Fully compatible with the well known Antrophic API.
You only have to replace the URL and your key.
It may not be likely but it's definitely possible enough to be something people worry about.
I'm not even sure what "EU AI safety rules" are. Can't people in the EU just use whatever they want?
If it's 500x cheaper than US models for similar ballpark performance just because it's hosted in China, sure whatever. If it's name brand like Anthropic/OpenAI/Google, that's kinda fine too.
If it's neither, like merely 50% cheaper than latest OpenAI whatever, however massive loss that pricing may be incurring to its provider, it wpuld be considered not worth any attention.
For funded startups it mostly just doesn’t matter a ton unless you are passing on inference in your product at scale
I’ll try out the latest models, but mainly stick with Claude only because I’m most used to its quirks and how to work around them. I imagine this is part of these hyperscalers playbook.
I will say though, I miss Sol model at work. It with Codex was amazing at first-shot understanding. Claude i need to scope out where to look otherwise a large portion of my token budget is eaten up
As an American, I'd much rather have my data kept outside the country than here where companies and the government have a lot more leverage over me.
For some well-explained reasons, check out https://darioamodei.com/essay/the-adolescence-of-technology and search for "CCP".
To me, the risks of an individual helping China to continue to develop their AI by being a customer is pretty marginal compared with the personal risks of my data being used against me.
The CCP's not great either, sure. But the Americans don't really have a leg to stand on anymore.