Quick take: SWE-1.7 lands within a hair of GPT-5.5 on coding benchmarks at roughly $1.97 a task, that's the whole pitch. On FrontierCode 1.1 it scored 42.3, just 0.7 behind GPT-5.5. Every number comes from Cognition's own testing, unverified by an independent lab. It's tuned for long-horizon agentic work inside Devin, not inline completions. Pilot it on your own tickets first.
What Is Cognition's SWE-1.7?
SWE-1.7 is a coding-focused large language model that Cognition, the company behind the Devin coding agent, shipped on July 8, 2026, built to run inside the same Devin harness their production customers already use for long-horizon agentic work. In a year where a new frontier-ish model lands roughly every fortnight, most of them blur together. This one has a sharper pitch than usual: near-frontier coding quality at about $1.97 per task, which is a fraction of what the top proprietary models cost for comparable agentic work.
So what did they actually build, and should you care? Let me lay out the claims, then the caveats.
Cognition isn't a research lab bolting a demo onto a paper. Devin has been running in production for real customers since 2025, so SWE-1.7 is a model shipped by a team that already knows what breaks when an agent works unsupervised for an hour. That context matters more than the benchmark table does, honestly, because it's the difference between a model tuned to win a leaderboard and one tuned to survive a real workday.
That context matters more than the benchmark table does, honestly, because it's the difference between a model tuned to win a leaderboard for one afternoon and one tuned to survive a real workday, unsupervised, inside a production coding agent that customers already depend on daily.
What Benchmark Numbers Is Cognition Quoting?
SWE-1.7 was RL post-trained on Kimi K2.7 Code, run inside the real Devin harness, and it's serving at around 1000 tokens per second on Cerebras. On Cognition's own FrontierCode 1.1 benchmark it scored 42.3, landing just 0.7 points behind GPT-5.5 and 4.2 behind Claude Opus 4.8. On SWE-Bench Multilingual it actually beat GPT-5.5, 77.8 to 76.8, though both sat behind Opus 4.8 at 84.4. On the independent Terminal-Bench 2.1 it managed 81.5, behind both frontier models but not by a mile.
The pattern is consistent: SWE-1.7 gets within a few points of the frontier on agentic coding while claiming a much lower cost per task. If those numbers hold in your workload, that's a real efficiency story, not a marketing one.
Worth noting what these three benchmarks actually measure, since they're not interchangeable. FrontierCode 1.1 is Cognition's own suite, built to stress the kind of multi-file, multi-step changes Devin handles daily. SWE-Bench Multilingual widens the net past Python-heavy repos into the languages most real codebases actually mix. Terminal-Bench 2.1 is the odd one out here, and the important one, because it's independently run rather than vendor-authored. A model that does well on a vendor's own suite but stumbles on the outside test is telling you something about where the training budget went.
SWE-1.7 vs GPT-5.5 vs Claude Opus 4.8
| SWE-1.7 | GPT-5.5 | Claude Opus 4.8 | |
|---|---|---|---|
| FrontierCode 1.1 | 42.3 | 43.0 | 46.5 |
| SWE-Bench Multilingual | 77.8 | 76.8 | 84.4 |
| Terminal-Bench 2.1 | 81.5 | not disclosed here | not disclosed here |
| Cost per task | ~$1.97 | not disclosed | not disclosed |
Where Should You Be Careful With These Claims?
Here's the honest part. Every benchmark above is Cognition's own reporting. And when has a vendor ever published an eval that made its own model look bad? No independent lab has reproduced them yet, and vendor-run evals have a way of flattering the vendor. GPT-5.5 and Opus 4.8 still lead outright on the harder tests, so "matches the frontier" is a cost-adjusted claim, not a quality-parity one. I'd read "near-frontier at low cost" as the real message, and I'd verify it on my own tasks before trusting it in production.
The model is also tuned for long-horizon, asynchronous work inside Devin. That's a specific shape of problem, autonomous agent runs rather than quick inline completions, so it matters whether that's how you actually use a coding model day to day.
Long-horizon and asynchronous also means fewer natural checkpoints for a human to catch a bad decision before it compounds. An inline completion you reject costs you a keystroke. An agent that spends twenty minutes building on a wrong assumption costs you the twenty minutes plus the cleanup. Whatever you pilot this against, build in a review step at the pull request boundary, not just a pass or fail on the final diff.
Should You Pilot SWE-1.7?
If your AI spend is dominated by long agentic coding runs, SWE-1.7 is worth a pilot on the strength of the cost claim alone. A vendor-run benchmark is a performance test designed, executed, and reported by the company selling the model, which is useful directional evidence but not a substitute for testing on your own repository. Run a pilot this way:
- Pick two or three tickets you'd normally hand to a mid-level engineer.
- Run each ticket through SWE-1.7 and whatever model you use today.
- Compare cost per completed task, not headline benchmark scores.
- Review the actual diffs a human has to read, not just whether the tests went green.
For the wider picture of where models like this fit alongside the editors and agents you already use, our guide to practical AI coding tools is the place to start.
One thing worth flagging if you point SWE-1.7 at a TypeScript codebase specifically: an agent running unattended for a while is exactly the scenario where loose types bite you. Don't relax strict mode to make it happier. The type system is what catches a bad autonomous edit before it ships rather than after, and that matters more here than with a quick inline completion, see our breakdown of why TypeScript makes AI code reliable for the mechanics.
The short version: a genuinely interesting release with a sharp cost angle, and a set of benchmarks that deserve your own verification before you believe the headline. Run it on a real ticket before you run it on a real deadline.