# Cognition's SWE-2 Coding Model Undercuts Rivals, but Not on Every Test

Page: https://hirantha.me/blog/cognitions-swe-2-coding-model-undercuts-rivals-but-not-on-every-test
Published: 2026-09-14
Written by: hirantha.me Newsroom (AI-generated). An AI model wrote this from the sources listed below; how it is made: https://hirantha.me/about/newsroom
Topics: cognition, swe-2, devin, ai coding tools, kimi k3

> Cognition's new SWE-2 coding model matches top rivals on most benchmarks at far lower claimed cost, but lags badly on one key test.

## Three speeds from one training run

Cognition, the company behind the autonomous coding agent Devin, released a new coding model called SWE-2 on September 10. It replaces SWE-1.7 as the model doing the actual code-writing inside Devin's desktop app, command-line interface, web app, and its multi-model "Fusion" workflow. Rather than one fixed model, SWE-2 ships in three effort levels, medium, high and max, meant to trade cost against thoroughness depending on the task.

The model is not built from scratch. Cognition took Kimi K3, a 2.8-trillion-parameter mixture-of-experts model from Moonshot AI that had already been through substantial reinforcement learning (RL, a training method where a model is rewarded or penalized based on the outcome of its actions) for coding tasks, and post-trained it further for use inside Devin's agent loop.

## Where it actually lands against the competition

On Cognition's own published benchmark table, SWE-2 scores 50.0% on FrontierCode 1.1 Main, a coding benchmark, putting it within one point of Anthropic's Fable 5.1 (50.9%) and ahead of xAI's Grok 4.6 (48.0%), OpenAI's GPT-5.6 Sol (47.5%), and its own predecessor SWE-1.7 (42.0%). On a second benchmark, DeepSWE 1.1, SWE-2 scores 73.0%, ahead of Fable 5.1's 67.4% and close behind OpenAI's newest model, GPT-6 Astra, at 74.1%. On Terminal-Bench 2.1, a benchmark for command-line agent tasks, SWE-2 leads the whole field at 92.8%.

That is three benchmarks where Cognition's claim of near-frontier performance holds up. The fourth does not. On Terminal-Bench 4, SWE-2 scores just 27.3%, less than half of Fable 5.1's 55.8% and GPT-6 Astra's 57.9%, though still well ahead of SWE-1.7's 7.6%. Independent analysis site CellCog flagged this gap directly, noting that Cognition's own published table contains the number but its promotional material does not mention it. In CellCog's words, the "on par with the frontier" framing "only holds on three benchmarks and not on the fourth."

## The cost claim, and what is not published

Cognition's headline pitch is that SWE-2 matches Fable 5.1 at 64% lower cost on FrontierCode 1.1 Main. That figure is real, but it comes with a gap: Cognition has not published a per-token or per-task price for SWE-2, and there is no standalone application programming interface (API) for the model outside Devin. The 64% figure is stated only as a comparison to Fable 5.1's list price, not as an absolute number.

CellCog worked backward from Fable 5.1's published rate of $3.28 per FrontierCode task to estimate that SWE-2 medium costs roughly $1.18 per task, but was explicit that this is an inference from Cognition's percentage claim, not a number Cognition has confirmed. Readers comparing coding models on price should treat that figure as an estimate, not a quote.

## The training method behind the three effort levels

The more interesting part of the release, technically, is how Cognition trained three effort levels at once instead of building three separate models. Earlier approaches typically require a separate training run, or a separate model, for a fast-cheap mode versus a slow-thorough one. Cognition instead ran a single reinforcement learning pass using what it calls a Pareto-informed cost penalty: the reward given to the model during training equals its success on the task minus a cost penalty, and that penalty is tuned separately for each effort level to match the local slope of the cost-versus-performance curve at that point.

In practice, Cognition says this made the medium effort level far more efficient. Where SWE-1.7 needed a mean of 127 steps per task and a median of 48 steps before making its first code edit, SWE-2's medium setting needs a mean of 53 steps and a median of just 18 steps to first edit. Cognition also states that medium effort matches SWE-1.7's performance using 58% fewer turns and at 81% lower average cost per task. Cognition attributes this to better engineering judgment early in a task: in the company's own words, "stronger engineering judgment allows the agent to write more complete solutions alongside fewer detours and redundant reads."

## What this means for developers

SWE-2 is not available as a general-purpose model. It only runs inside Cognition's Devin product line, so developers who want it have to adopt Devin's agent workflow rather than plugging the model into their own tooling through an API, which is how Fable 5.1, GPT-6 Astra and Kimi K3 are all also obtainable directly. For teams already using Devin, the practical upshot is a cheaper default for routine code-writing and editing tasks, with the option to step up to high or max effort when a task turns out to be harder than expected.

For anyone choosing a coding model based on benchmarks alone, the release is a reminder to check which specific benchmark a headline claim is measured against. SWE-2 is a genuinely strong, cheaper option for the kinds of tasks FrontierCode, DeepSWE and Terminal-Bench 2.1 test. It is a poor fit, at least for now, for whatever harder, longer-horizon terminal tasks Terminal-Bench 4 is designed to catch.

## Questions

**What is Cognition's SWE-2?**
It is Cognition's newest coding model, released September 10, 2026, post-trained with reinforcement learning on top of Moonshot AI's 2.8-trillion-parameter Kimi K3 model, and it powers Cognition's Devin coding agent.

**Is SWE-2 better than Anthropic's Fable 5.1?**
It closely matches or beats Fable 5.1 on three of the four benchmarks Cognition published, at a claimed 64% lower cost, but it scores far lower on Terminal-Bench 4 (27.3% versus 55.8%).

**Can I use SWE-2 outside of Devin?**
No. Cognition has not released open weights or a standalone API for SWE-2; it is only available inside Devin's desktop app, command-line interface, web app and Fusion workflow.

**How much does SWE-2 cost to run?**
Cognition has not published per-token or per-task pricing; its 64% savings claim is relative to Fable 5.1's list price, and outside analysts have only been able to estimate an absolute figure.

**What is new about how SWE-2 was trained?**
Cognition trained three effort levels (medium, high, max) in a single reinforcement learning run using a cost penalty tuned to each level, instead of training separate models for fast-cheap versus slow-thorough responses.

## Sources

- [Introducing SWE-2: Pushing the Pareto Frontier](https://cognition.com/blog/swe-2), Cognition
- [Cognition Releases SWE-2, Says It Performs Close To Frontier At 70% Lower Cost](https://officechai.com/ai/cognition-releases-swe-2-says-it-performs-close-to-frontier-at-70-lower-cost/), OfficeChai
- [Cognition Releases SWE-2: A Kimi K3 Post-Trained Coding Model That Matches Fable 5.1 on FrontierCode at 64% Lower Cost](https://www.marktechpost.com/2026/09/12/cognition-releases-swe-2-a-kimi-k3-post-trained-coding-model-that-matches-fable-5-1-on-frontiercode-at-64-lower-cost/), MarkTechPost
- [Cognition SWE-2: Benchmarks, the 64% Cost Claim, and the Row It Loses](https://cellcog.ai/blog/cognition-swe-2/), CellCog