SWE-2, from Cognition, promises to rival top-tier models at a quarter of the cost

Cognition, the company behind the Devin engineering agent, announced SWE-2, which it describes as its most advanced coding model to date. The central argument isn't being the best at everything, but pushing the Pareto frontier between capability and cost: according to the company, the model reaches 50.0% on FrontierCode 1.1 Main, less than one point behind Fable 5.1 (50.9%), while costing 64% less.

For the developer who today pays a high price to run a coding agent in production, this is the angle that matters. It isn't about swapping out the top-ranked model, it's about finding a similar quality tier while spending a fraction of the cost per task.

What Cognition is promising

SWE-2 is post-trained from Kimi K3, a 2.8-trillion-parameter model that, according to Cognition, had already gone through extensive RL for agentic coding. On top of that base, the company applied its own reinforcement learning and claims gains of 5 to 6 points across several benchmarks. It's the first time the company has scaled RL to the multi-trillion-parameter regime, reusing SWE-1.7's infrastructure.

The numbers released by the company, comparing SWE-2 with competitors:

| Benchmark | SWE-2 | Kimi K3 | Grok 4.6 | Fable 5.1 | GPT-5.6 Sol | GPT-6 Astra | SWE-1.7 | |---|---|---|---|---|---|---|---| | FrontierCode 1.1 Main | 50.0% | 44.2% | 48.0% | 50.9% | 47.5% | 53.3% | 42.0% | | DeepSWE 1.1 | 73.0% | 68.5% | 67.5% | 67.4% | 72.7% | 74.1% | 37.7% | | Terminal-Bench 2.1 | 92.8% | 88.3% | 88.4% | 91.4% | 88.8% | 89.9% | 81.5% | | Terminal-Bench 4 | 27.3% | 21.5% | 20.3% | 55.8% | 37.3% | 57.9% | 7.6% |

The company's own reading: SWE-2 beats SWE-1.7 and Grok 4.6 on score and cost, ties with GPT-5.6 Sol and Fable 5/5.1 for a fraction of the price, and comes within a few points of GPT-6 Astra for a quarter of the cost.

Less back-and-forth, more code

The point that hits closest to the daily experience of anyone using an agent isn't in the benchmark table, it's in the behavior. One of the complaints reported about SWE-1.7 was its tendency to over-explore the codebase before touching any file, which inflated time and cost on simple tasks.

According to Cognition, SWE-2 fixes this with more focused exploration. The numbers released:

• On FrontierCode 1.1 Main, SWE-2 medium scores higher than SWE-1.7 while taking 58% fewer turns and costing 81% less on average.
• The first real code edit happens after a median of 18 steps, versus 48 for SWE-1.7.
• The average number of steps per task drops from 127 (SWE-1.7) to 53 (SWE-2 medium).

The company offers three effort levels (medium, high, and max), trained in a single RL run with cost penalties calibrated for each level. In the practice proposed by Cognition, medium kicks in faster and solves simple and intermediate tasks at low cost, while high and max plan and explore more to handle complex problems. For those building agent pipelines, this explicit control over effort per task is the kind of lever that changes the bill at the end of the month.

The point the community didn't let slide

The gap between the two Terminal-Bench scores drew attention right away on the Hacker News thread. SWE-2 scores 92.8% on Terminal-Bench 2.1, but drops to 27.3% on Terminal-Bench 4, behind Fable 5.1 (55.8%), GPT-6 Astra (57.9%), and even GPT-5.6 Sol (37.3%).

The user postalcoder sums up why this matters:

"If you're looking for reason to be skeptical, look no further than the massive delta between the Terminal Bench 2.1 (92.8%) and the Terminal Bench 4 score (27.3%). Terminal Bench 4 was released a couple weeks ago, so the difference you're seeing between the two scores can be interpreted as 'how well does this model generalize to new problems'?"

>

-- postalcoder

His reading: since Terminal-Bench 4 is recent, the difference suggests how much the model might be benchmaxxed, that is, optimized for known benchmarks rather than real generalization. Along the same lines, mydreamof asks: "Seems like benchmaxing? For example for Terminal-Bench 4 it doesn't have great results. And why not show other benchmarks?" It's the usual caveat for anyone evaluating a model by blog-post numbers: benchmark is a starting point, not a verdict. For the Brazilian developer, the practical recommendation is to test it on your own codebase before swapping out whatever is already in production.

Another point raised in the thread is distribution. The user scronkfinkle complains that the model seems to require the company's proprietary platform: "this appears to require Devin to use? I'm disappointed to see I need to use a bespoke platform to interact with this agent, to the point that I probably won't be trying it." Indeed, according to the announcement, SWE-2 is available starting today on Devin Desktop and CLI, with rollout also planned for Devin Web and Fusion. There's no mention of an open API or integration with third-party editors, which limits quick testing for those who don't want to adopt the Devin ecosystem.

What remains open

SWE-2 being based on Kimi K3 sparked an interesting reaction among developers. TheJCDenton comments that they expected a fully new model, but sees value in the path chosen: "it's an RL-ed K3 go Fable 5 capabilities, which demonstrate that this is probably possible, which is nice." The demonstration that you can take a strong open model and, via RL, bring it closer to frontier models is relevant for anyone tracking the cost of running coding assistants.

There's also track record in its favor. The user _doctor_love notes that "SWE-1.5 was surprisingly good when I used it last" and that Cognition is "one of the solid players that's flying a bit under the radar while Anthropic and OpenAI race to IPO". It's the kind of reputation that carries weight when cost-effectiveness is the selling point.

What we still don't know, and what will determine whether SWE-2 is worth switching to: how it behaves outside benchmarks, on real and legacy codebases; whether there will be a way to use it without depending entirely on Devin; and what the final price per task looks like in a currency you can actually put into a spreadsheet. In the meantime, Cognition details in the technical post the RL cost-penalty methodology and the length-weighted reward baseline, dense reading for anyone interested in the training side.