Cognition just dropped SWE-2, and the headline number is wild: within one point of the best coding model on the planet, at 64% less cost. If you pay for AI coding help by the task, that math matters. Even if you’ve never written a line of code.
Cognition SWE-2 scored 50.0% on FrontierCode 1.1 Main, the industry’s toughest agentic coding benchmark. Claude Fable 5.1 sits at 50.9%. The gap is basically a rounding error. The price gap is not, and that’s why developers are paying attention this morning.
What Cognition SWE-2 actually is
SWE-2 is the new coding model from Cognition, the company behind Devin, the AI software engineer. It’s built for agentic coding, meaning it doesn’t just answer questions about code. It plans, edits files, runs tests, and fixes what breaks, mostly on its own.
Under the hood, Cognition post-trained SWE-2 from Kimi K3, a massive 2.8-trillion-parameter open model. Their reinforcement learning pass added another 5 to 6 points on most benchmarks. Sounds incremental, until you see where it landed.
Here’s the part beginners should care about. Cheaper frontier-adjacent coding AI means the tools you use (or will soon) get faster and less expensive. The AI features inside freelance platforms, no-code app builders, and dev agencies all get their engines from models like this.
The benchmark numbers, translated
Cognition published its own results, so take them with the usual grain of salt until independent labs replicate. But the table tells a clear story:
| Benchmark | SWE-2 | Fable 5.1 | GPT-6 Astra | Kimi K3 |
|---|---|---|---|---|
| FrontierCode 1.1 Main | 50.0% | 50.9% | 53.3% | 44.2% |
| DeepSWE 1.1 | 73.0% | 67.4% | 74.1% | 68.5% |
| Terminal-Bench 2.1 | 92.8% | 91.4% | 89.9% | 88.3% |
| Terminal-Bench 4.0 | 27.3% | 55.8% | 57.9% | 21.5% |
Translation for normal humans. On everyday coding tasks, SWE-2 basically matches the most expensive models on Earth. On Terminal-Bench 2.1, which tests real terminal work, it posted the highest score in the published table. It even beats the previous version, SWE-1.7, while taking 58% fewer turns and costing 81% less per task.
But there’s a weak spot, and Cognition doesn’t hide it. On Terminal-Bench 4.0, which measures long-horizon work (multi-hour projects with lots of moving parts), SWE-2 scores 27.3% against 55.8% for Fable 5.1. Big jobs still favor the premium models. Quick jobs are where SWE-2 eats their lunch.
The speed story nobody expected
Raw scores only tell half the story. The other half is patience, yours and the model’s.
SWE-1.7 averaged 127 steps per run because it over-explored. It would read file after file, double-checking everything before touching anything. Thorough, yes. But users complained it overthought simple tasks, and they were right.
SWE-2 medium averages 53 steps. It makes its first real edit after a median of 18 steps, where SWE-1.7 waited until step 48. Same quality bar on medium-difficulty work, less than half the waiting. Anyone who has watched an AI assistant “think” for five minutes before doing something useful knows exactly why this matters.
There’s a nice human detail in Cognition’s post, too. In one internal test, an MCP integration SWE-2 needed was unavailable. So the model dug through the Slack channel history it already had access to and reconstructed the missing data itself. That’s resourcefulness, and it’s the kind of behavior that separates a tool from an assistant.
How to try it today
Getting your hands on Cognition SWE-2 is simple, and you don’t need to be a developer to poke at it:
- Devin Desktop: download from devin.ai/desktop and SWE-2 is the built-in brain. This is the easiest path for non-engineers.
- Devin CLI: for terminal people, devin.ai/cli gets the same model.
- Devin Web and Fusion: rolling out now, so check your workspace if you’re already a Devin user.
One important catch: there’s no public API and no downloadable weights. You can’t run SWE-2 locally, and you can’t plug it into your own scripts the way you can with open models like Qwen. We covered that trade-off in our Qwen 3.8 review, which is the opposite bet: slightly weaker performance, fully free and local.
Cognition prices SWE-2 per task inside Devin instead of per token, so your bill depends on how much you use it. The company’s claim is that medium effort hits near-frontier quality at 64% less cost than Fable 5.1. If your usage is a handful of tasks a day, the savings are small. If you’re running an agency or shipping software daily, they compound fast.
Should you care if you’re not a coder?
Short answer: yes, indirectly. Here’s why this launch reaches beyond engineering teams.
First, price pressure. When a frontier-adjacent model costs a quarter of the competition, every AI-powered tool built on top of coding engines gets cheaper. The website builders, automation platforms, and app generators you use subscribe to models like this under the hood. Their costs are your subscription prices.
Second, the “agents doing real work” trend keeps accelerating. We tracked it in our piece on OpenAI Codex persistent agents, where background coding agents grind through tasks while you sleep. SWE-2 makes that style of working cheaper, which means more products will offer it.
Third, quality floor. When near-frontier coding becomes cheap, small businesses can afford custom software fixes that used to cost four figures. A broken integration, a clunky spreadsheet-to-database move, a simple internal tool: these become reasonable one-off purchases instead of projects.
Honestly, the most interesting angle is what SWE-2 says about the market. Two years ago, frontier quality came from exactly three companies. Now a post-trained open model lands within a point of the leader. Competition like this is why AI tools keep getting better and cheaper at the same time.
The bigger story: open models are catching the frontier
Zoom out, because Cognition SWE-2 is part of a pattern that got very obvious this week.
The recipe here matters. Kimi K3 is an open-weights model anyone can download. Cognition took it, ran serious reinforcement learning on top, and landed within a point of the closed-model leader. Meanwhile, Sakana AI released Fugu Max today too, an orchestration system that routes work across a pool of open models and claims frontier-grade results at two to six times lower cost. Different approaches, same conclusion.
The gap between “expensive closed models” and “everything else” is collapsing from both directions. For buyers, that’s the best possible news. Prices fall while quality holds steady, and you’re never stuck with one vendor’s bill.
Where SWE-2 fits against the alternatives
Quick orientation if you’re choosing a coding AI right now:
- SWE-2 (via Devin): best value for everyday coding tasks and automated fixes. Weak on very long projects.
- Fable 5.1: the premium pick for complex, multi-hour builds where budget is secondary.
- Qwen 3.8: the free local option, if you have the hardware and want zero API bills.
For bug fixes, small features, and routine maintenance, SWE-2’s numbers say it handles the job at a fraction of the price. We also put its bug-hunting cousins through real tests in our Claude Code debugging guide, which is worth a read if debugging is your main use case.
The takeaway
Cognition SWE-2 just made near-frontier coding AI dramatically cheaper, with the only real sacrifice being long-horizon projects. If you already use Devin, you have it today. If you don’t, and you’ve been curious about AI coding help, this launch plus Devin’s free tier is a reasonable starting point. Try it on one small, real task this week and judge the output yourself.
