we've invested heavily into infra
one of the non-trivial things we did there is 2 levels of isolation instead of one
there are VMs with fast NVMe disks and then inside VMs lighter rootless podman sandboxes - and each organisation gets a fleet of N VMs dedicated, N>=1
this way the VMs serve as a "soft" concurrency limit and all sessions are still isolated, guaranteed to run on your VM, and start fast
"Premium markup." If you go past the usage included in a paid plan, extra usage is billed at API pricing. In September Devin also got 30–40% cheaper in Fusion and Normal mode and 15–20% cheaper in Ultra while quality held or improved at every mode.
"Cheap, low-quality default model." Devin isn't tied to one model. Our normal router uses Opus 5.5 at the moment (or GPT-6.2 Sol high on GPT plan), and Ultra uses Fable 5.2 (or GPT-6.2 Astra high on GPT plan), only when someone chooses the light mode specifically are they routed to SWE-2 which is actually a great model trained specifically for software engineering that benches really well and developers really love (and SWE-2 inference / usage is currently free).
"Only the bad engineers use it." Our users include engineers from Nvidia, Exa, Modal, OpenRouter, Mercor, Ramp, and hundreds of other companies with a high bar for engineering talent, and the best ones seem to really love and appreciate the quality and shipping velocity they get with Devin.
I encourage anyone reading this to consider asking someone they know who's used Devin within the past few months for their opinion.
> Devin charges a premium markup and has switched their default model to a very cheap, low quality model.
I work at Cognition. This is not true. The picker currently has:
- Normal: Balances cost and intelligence with models like Opus 5.5 and GPT-6.2 Sol
- Ultra: Highest intelligence available with Fable 5.2 and GPT-6.2 Astra
- Lite: Fast lightweight models
- Fusion: Multi-model harness that pairs frontier models like Opus 5.5, Fable 5.2, and GPT-6.2 Astra with a cost efficient sidekick for up to 30% cost savings
And then you can select any of the "preview" models like SWE-2, GPT 6.2, Gemini 4 Neon, Fable 5.2, etc.
The built-in modes are basically just "best model at xyz on cost curve" and then we have individual models people can select.
I didn't realize they both came from the same company. Apparently Windsurf was renamed Devin Desktop.
What I associate Devin with, and what this seemingly competes with, is the cloud hosted version. The UI and harness as well as runners are hosted, and multiple people can prompt in the same session.
The main difference from our pov is that OpenHands is an agent with its own harness. We’re unopinionated on the harness and more focused on the infrastructure it runs in. In fact we started as one of the first “persistent” sandbox providers [1].
OpenComputer runs the user’s Claude Code or Codex (you pick per thread), and you can bring your own ChatGPT/Codex subscription.
Our focus is predominantly on building high perf specialised infra for cloud coding agents. eg: each Slack thread gets its own rootless Podman container, with state written to local disk and then persisted to a git backed store. That’s why it feels like Claude Code on your laptop: in our benchmark it opened PRs 3.3x faster than Devin, with first push in 24s vs a 62s average across the 4 most popular cloud agents [2]. It's also cheaper at at scale, we've found, than an ephemeral sandbox per session [3].
hey! founder here. We have a web ui and a cli as alternatives, you can also integrate the cli with local coding agents or things like openbot. Discord can be on the roadmap! Hope you can try it and file feature request for us
I see the word self-hosted but don't see any mention of Ollama or dedicated home self-hosting scenario.
What if somebody wants to just self-host on a home pc, say using Ollama to serve the models? Does OC work w/ Ollama directly, providing "its own agent capabilities" rather than relying on other agents e.g. ClaudeCode, Codex, OpenCode, etc?
About a year ago I think, yes. But I've heard of at least 100 other agentic tools in that time and I don't recall what most of them do. What's windsurf?
we've invested heavily into infra one of the non-trivial things we did there is 2 levels of isolation instead of one there are VMs with fast NVMe disks and then inside VMs lighter rootless podman sandboxes - and each organisation gets a fleet of N VMs dedicated, N>=1
this way the VMs serve as a "soft" concurrency limit and all sessions are still isolated, guaranteed to run on your VM, and start fast
At my organization, all of the skilled engineers use Claude Code and the bad ones use Devin cause it’s “easier”.
It shows in the code thats written
"Premium markup." If you go past the usage included in a paid plan, extra usage is billed at API pricing. In September Devin also got 30–40% cheaper in Fusion and Normal mode and 15–20% cheaper in Ultra while quality held or improved at every mode.
"Cheap, low-quality default model." Devin isn't tied to one model. Our normal router uses Opus 5.5 at the moment (or GPT-6.2 Sol high on GPT plan), and Ultra uses Fable 5.2 (or GPT-6.2 Astra high on GPT plan), only when someone chooses the light mode specifically are they routed to SWE-2 which is actually a great model trained specifically for software engineering that benches really well and developers really love (and SWE-2 inference / usage is currently free).
"Only the bad engineers use it." Our users include engineers from Nvidia, Exa, Modal, OpenRouter, Mercor, Ramp, and hundreds of other companies with a high bar for engineering talent, and the best ones seem to really love and appreciate the quality and shipping velocity they get with Devin.
I encourage anyone reading this to consider asking someone they know who's used Devin within the past few months for their opinion.
I work at Cognition. This is not true. The picker currently has:
- Normal: Balances cost and intelligence with models like Opus 5.5 and GPT-6.2 Sol
- Ultra: Highest intelligence available with Fable 5.2 and GPT-6.2 Astra
- Lite: Fast lightweight models
- Fusion: Multi-model harness that pairs frontier models like Opus 5.5, Fable 5.2, and GPT-6.2 Astra with a cost efficient sidekick for up to 30% cost savings
And then you can select any of the "preview" models like SWE-2, GPT 6.2, Gemini 4 Neon, Fable 5.2, etc.
The built-in modes are basically just "best model at xyz on cost curve" and then we have individual models people can select.
If you invested in the company, or are affiliated somehow, consider disclosing..
And yes please give it a shot and let us know what you think!
A little harsh, but at least it's leagues ahead of Devin.
What I associate Devin with, and what this seemingly competes with, is the cloud hosted version. The UI and harness as well as runners are hosted, and multiple people can prompt in the same session.
The main difference from our pov is that OpenHands is an agent with its own harness. We’re unopinionated on the harness and more focused on the infrastructure it runs in. In fact we started as one of the first “persistent” sandbox providers [1]. OpenComputer runs the user’s Claude Code or Codex (you pick per thread), and you can bring your own ChatGPT/Codex subscription.
Our focus is predominantly on building high perf specialised infra for cloud coding agents. eg: each Slack thread gets its own rootless Podman container, with state written to local disk and then persisted to a git backed store. That’s why it feels like Claude Code on your laptop: in our benchmark it opened PRs 3.3x faster than Devin, with first push in 24s vs a 62s average across the 4 most popular cloud agents [2]. It's also cheaper at at scale, we've found, than an ephemeral sandbox per session [3].
[1] https://x.com/utpalnadiger/status/2047436177016537251 [2] https://opencomputer.dev/blog/cloud-agents-as-fast-as-local/ [3] https://opencomputer.dev/blog/opencomputer-open-source-devin...
You need one VM in your cloud plus Cloudflare Workers (the runtime of which is also open source as workerd) for Slack events.
What if somebody wants to just self-host on a home pc, say using Ollama to serve the models? Does OC work w/ Ollama directly, providing "its own agent capabilities" rather than relying on other agents e.g. ClaudeCode, Codex, OpenCode, etc?
Will any of the above be supported in the future?
right? right..?