What is SWE-2? And what does it mean for your AI stack?

SWE-2 is Cognition's new coding model, post-trained with reinforcement learning from Kimi K3, Moonshot AI's 2.8-trillion-parameter open model. Cognition reports it scores within one point of a frontier model on its FrontierCode benchmark at 64% lower cost, and it runs only inside Devin. For AI buyers, it's more evidence that models change fast and context should stay portable.

What did Cognition actually release?

Cognition, the company behind the Devin coding agent, released SWE-2 on 12 September 2026, as MarkTechPost reported. It's post-trained with reinforcement learning from Kimi K3, Moonshot AI's 2.8-trillion-parameter open model, and Cognition's own announcement reports 50.0% on FrontierCode 1.1 Main, within a point of a frontier model's 50.9%, at 64% lower cost. It's also Cognition's first model with selectable effort levels, all trained in a single RL run.

Does SWE-2 really match the frontier?

On Cognition's numbers, nearly, and those two words are doing real work. FrontierCode is Cognition's own benchmark, and the rival scores in its table come from Cognition's own evaluation. The same table shows an honest weak spot: on Terminal-Bench 4, SWE-2 trails the frontier models by roughly 30 points.

So read it as a vendor-reported result, not a settled ranking. But even heavily discounted, the direction is the story. A model post-trained from an open base is landing near frontier performance on real coding work for a fraction of the price, and this keeps happening. Every quarter, the price of frontier-level capability drops.

Can you run SWE-2 yourself?

No. SWE-2 has no open weights and no standalone API. It runs only inside Devin: Desktop and CLI today, with the web version rolling out.

Sit with that combination for a second. The open-weight base made the model cheap to build. The product it landed in is closed. That's not a criticism of Cognition (it's a reasonable commercial choice), but it means SWE-2 isn't infrastructure you own. It's a feature of someone else's product, and everything you build around it lives on their terms.

What does this mean for your AI stack?

The cheaper frontier-level models get, the more expensive welded context becomes.

Here's the mechanism. A better, cheaper model landing every quarter is only good news if you can actually move to it. If your context, workflows and guardrails live inside one vendor's product, every leap is a migration you'll quietly skip: the pricing rules someone finally taught the old tool in March, the edge cases, the tone your sales lead insisted on. None of it moves with you, so you stay. That's capability left on the table, again and again, to avoid a rebuild.

I think cost parity, not benchmark supremacy, is the real story of this release cycle. The engines are commoditising. The compounding asset is the layer that knows your business, and the only question that matters is who owns it. That holds whether you buy the agent or build it yourself.

How do you keep your context portable?

Keep the engine rented and swappable, and keep the layer it acts through in your own hands, expressed over an open protocol like MCP so it isn't welded to any one model or tool.

That's what Oi is: a shared brain that connects once through MCP and travels across models and tools. (It's the same layer that makes multiplayer AI work from the model's side.) When the next SWE-2-shaped release lands, you point the same contexts and guardrails at the new engine and keep everything the old one learned. Switching stays a configuration change, so the quarterly price drops become pure upside instead of a migration backlog.

If you'd rather own that layer than rebuild it every time the frontier moves, book a demo.

FAQ

What is Kimi K3?

Kimi K3 is Moonshot AI's 2.8-trillion-parameter open model. Cognition used it as the base for SWE-2, adding reinforcement learning on top; Cognition reports the RL adds 5 to 6 points on many benchmarks over the base.

Is SWE-2 open source?

No. Its base model, Kimi K3, has open weights, but SWE-2 itself ships with no open weights and no standalone API. It's available only inside Cognition's Devin product.

What is FrontierCode?

FrontierCode is Cognition's own coding benchmark. SWE-2's headline result (within one point of a frontier model at 64% lower cost) comes from it, with rival numbers produced by Cognition's own evaluation, so treat cross-vendor comparisons as vendor-reported.

Why does model portability matter?

Because the model layer keeps changing: prices drop, capabilities jump, and access can be renegotiated. If your context, workflows and guardrails sit in a portable layer over an open protocol like the Model Context Protocol, each change is an upgrade. If they're welded to one product, each change is a migration.