What changed
On 08-07-2026 Cognition shipped SWE-1.7, an agentic coding model that lands within a few points of the closed frontier on software-engineering benchmarks 1. The lineage is the interesting part: it was trained from a Kimi K2.7 base that had already gone through extensive RL post-training, then Cognition ran its own additional RL on top 12. It is available today only inside Devin (Web, Desktop, and CLI), served via Cerebras at 1000 tokens per second 12.
The moat moved to the RL harness
The value here did not come from a new base model. Cognition trained SWE-1.7 from a Kimi K2.7 base that was already RL-heavy, and its second RL pass still bought large gains, which the team reads as evidence against a “post-training ceiling” 1. For any team that cannot afford to pre-train, that is the load-bearing signal: returns still live in the harness and the post-training, not only in the weights.
Keep the scores honest 1:
| Benchmark | SWE-1.7 | GPT-5.5 | Opus 4.8 |
|---|---|---|---|
| FrontierCode 1.1 Main | 42.3% | 43.0% | 46.5% |
| Terminal-Bench 2.1 | 81.5% | 84.2% | 86.9% |
| SWE-Bench Multilingual | 77.8% | 76.8% | 84.4% |
This is near-frontier, not frontier. SWE-1.7 trails both GPT-5.5 and Opus 4.8 on FrontierCode Main and Terminal-Bench, and only edges GPT-5.5 on SWE-Bench Multilingual (77.8% versus 76.8%). The pitch is cost and 1000-TPS throughput on the Pareto curve, not a new capability crown.
Impact on your team
If you are choosing a cheap coding model to drop into your own agent stack, SWE-1.7 gives you nothing today: it is Devin-only, with no API to call 12. If you already run Devin, the decision is a dollars-per-task and throughput trade on the Pareto curve, not a capability upgrade, since it still trails Opus 4.8 and GPT-5.5 on two of three benchmarks 1.
The result worth watching is the method, not the model. The cheapest path to near-frontier agentic coding this cycle was a second RL pass on someone else’s open base plus a proprietary harness, not a fresh pre-train 1. Until a callable API shows up, treat the cost story as Devin-internal economics rather than your invoice, and hold any bake-off in your own stack.