SWE-2 (Cognition)

Learning Objectives

• Understand what SWE-2 is, and why a US lab post-trained its flagship on Chinese open weights
• Read Cognition's benchmark claims correctly, including the one where SWE-2 loses badly
• Judge whether a model with no standalone API belongs in your toolchain

What Is SWE-2?

SWE-2 is the coding model released by Cognition on September 10, 2026. Cognition's framing is about the cost-performance curve rather than raw capability: the pitch is not that SWE-2 is the best coding model available, but that it lands close to the frontier for a fraction of the price.

The headline figure is 50.0 percent on FrontierCode 1.1 Main, which Cognition says is within one point of Claude Fable 5.1 while costing 64 percent less.

The base model is the story

SWE-2 is post-trained from Kimi K3, the 2.8 trillion parameter mixture-of-experts model published by Moonshot AI in July 2026 — a Chinese lab's open-weight release. Cognition says K3 had already undergone extensive reinforcement learning for agentic coding, and that its own reinforcement learning found substantial headroom on top of that, adding five to six points on many benchmarks.

That a well-funded US company building a flagship developer product chose a Chinese open-weight base is the most consequential fact on this page, and it is worth stating plainly rather than leaving as an aside. Open weights travel regardless of where they were trained, and a permissive-enough license makes the origin an engineering detail rather than a procurement one.

The benchmark table

Every number below is published by Cognition, and FrontierCode is Cognition's own leaderboard. Read them as vendor claims pending independent replication.

Where SWE-2 loses, and it is not close

Three of those four rows support Cognition's argument. The fourth does not, and it is the one worth dwelling on.

On Terminal-Bench 4, SWE-2 scores 27.3 percent against 55.8 percent for Fable 5.1 and 57.9 percent for GPT-6 Astra — roughly half. The same model that leads the table on Terminal-Bench 2.1 finishes last among the frontier models on Terminal-Bench 4.

The honest reading is that SWE-2 is strong on the benchmarks closest to what it was trained for and much weaker on the newest and hardest terminal tasks, where the two closed frontier models hold a large lead that price does not offset. If your work looks like the hard end of Terminal-Bench 4, the cheaper model is not the better buy.

Efficiency gains over SWE-1.7

The clearest improvement is not the score but the path to it. Against its own predecessor on FrontierCode 1.1 Main, Cognition reports SWE-2 medium scoring higher than SWE-1.7 while taking 58 percent fewer turns and costing 81 percent less.

Mean steps per run fall from 127 for SWE-1.7 to 53 for SWE-2 medium, and the model makes its first real edit after a median of 18 steps against 48 for SWE-1.7. Cognition attributes this to focused exploration — SWE-1.7 drew user complaints for over-exploring simple tasks, and better judgment about which parts of a codebase matter is what closes that gap.

The training change behind it: a reinforcement learning algorithm that trains all reasoning-effort levels in a single run, applying a linear cost penalty per effort level tuned to the local slope of the base model's curve, so the whole frontier advances rather than one operating point.

Pricing and access

This matters more than it would for a chat model. The efficiency claims above are measured inside Cognition's own agent harness, so they describe SWE-2 as deployed in Devin rather than SWE-2 as a model you could run somewhere else and compare.

Strengths

• A real cost-performance argument, not a capability boast — close to frontier scores at a quarter to a third of frontier price is the kind of claim that changes what a team can afford to run
• Large efficiency gain over its predecessor — 58 percent fewer turns and 81 percent lower cost against SWE-1.7, with the first edit arriving at a median of 18 steps rather than 48
• Leads the table on Terminal-Bench 2.1 at 92.8 percent, above both closed frontier models
• Built on open weights, which makes the base independently inspectable even though the post-trained result is not
• A training result worth knowing — advancing an entire cost-performance frontier in one reinforcement learning run, rather than tuning one operating point, is a genuinely useful technique

Limitations & Considerations

• Terminal-Bench 4 is a large, unexplained loss — 27.3 percent against 55.8 for Fable 5.1 and 57.9 for GPT-6 Astra. Cognition publishes the number without accounting for it, and it is the benchmark most likely to resemble hard real work
• Every figure here is vendor-reported, and the headline benchmark is Cognition's own leaderboard. No independent replication exists yet
• No standalone API and no published model pricing — you cannot evaluate it against your own tasks outside Devin, which is exactly the comparison the cost argument invites
• The efficiency numbers are harness-bound, measured inside Devin rather than for the model in isolation
• Still behind GPT-6 Astra on three of four benchmarks, and behind Fable 5.1 on two — the claim is proximity at lower cost, not parity
• Base-model provenance is an open question for the category generally, following Anthropic's allegations against Moonshot published the same day

Getting Started

1. SWE-2 is reached through Devin, so start at devin.ai — there is nothing to install or call directly
2. Use Devin Desktop or the CLI, which have SWE-2 today; Devin Web and Fusion are still rolling out
3. Pick your effort level deliberately — medium, high and max differ substantially in steps and cost, and medium is where the efficiency claims are measured
4. Benchmark it against your own repository rather than the published table, since the vendor's numbers come from its own leaderboard and its own harness
5. Try a hard terminal task early. That is where the published gap against Fable 5.1 and GPT-6 Astra is widest, and where you will find out fastest whether the cost saving holds for your work
6. Compare total task cost, not per-token price — fewer turns is where the saving actually comes from

Key Takeaways

• SWE-2 is Cognition's coding model, released September 10, 2026 and post-trained from Moonshot AI's open-weight Kimi K3
• The claim is cost-performance, not capability — 50.0 percent on FrontierCode 1.1 Main, within one point of Fable 5.1 at 64 percent less cost, and within a few points of GPT-6 Astra at roughly a quarter of the cost
• It loses badly on Terminal-Bench 4, scoring 27.3 percent against 55.8 for Fable 5.1 and 57.9 for GPT-6 Astra, and Cognition does not explain the gap
• It is not a product you can buy — no standalone API, no published model price, available only inside Devin
• Every benchmark figure is vendor-reported on a leaderboard Cognition operates
• A US lab built its flagship on Chinese open weights, which is the strategically interesting fact and a reminder that open weights travel regardless of origin