Cognition released SWE-1.7 on Wednesday, a coding model the startup says reaches near-frontier intelligence while costing far less to serve than its closest rivals. Cognition, the company behind the autonomous coding agent Devin, built the model on top of Kimi K2.7, an open-weight base model from Moonshot AI, rather than pretraining from scratch, then extended it through its own reinforcement learning pipeline.
That detail matters more than it looks. Kimi K2.7 had already gone through extensive post-training before Cognition touched it, and the company says its additional RL work still produced large capability gains. That result cuts against the assumption that a heavily post-trained model has little headroom left to extract, and it suggests smaller labs can keep improving borrowed base models rather than needing to train one from zero.
On Cognition’s three-benchmark suite, the bet mostly pays off. SWE-1.7 scores 42.3% on FrontierCode 1.1 Main, 81.5% on Terminal-Bench 2.1, and 77.8% on SWE-Bench Multilingual, well ahead of the Kimi K2.7 Code base it started from (30.1%, 72.7%, and 73.5%) and far past Cognition’s prior model, SWE-1.6, which scored 9.4% on FrontierCode 1.1 Main. SWE-1.7 also edges past OpenAI’s GPT-5.5 on the multilingual benchmark. It does not catch Anthropic’s Opus 4.8, which leads on all three tests by four to seven points. FrontierCode is Cognition’s own benchmark, introduced in June, so that particular comparison is worth reading with that ownership in mind.
The gains came from infrastructure work as much as new algorithms. Cognition trained SWE-1.7 across four data centers spanning three continents, keeping only the trainer on a single high-bandwidth cluster while treating each inference cluster as disposable and self-contained. Weight updates move between clusters as compressed deltas instead of full model copies, which the company says cuts transfer size by more than 99%. Cognition also targeted a known reinforcement-learning failure mode, where a capable model stops trying new approaches and improvement flatlines after a few hundred steps, by restricting rollout sampling to higher-probability tokens.
Long-horizon coding work introduces a separate problem: agent sessions run longer than a model’s context window allows. Cognition trained SWE-1.7 to summarize its own working state and resume from that summary, a technique it calls self-compaction, which let training sessions run as long as six hours. On the data side, the company said it built automated checks to catch both false-positive and false-negative task grading, dropped tasks the model always solved or always failed, and restricted network access inside its training sandboxes to prevent the model from gaming its own reward.
SWE-1.7 is live now inside Devin’s web, desktop, and CLI interfaces, running on Cerebras inference at 1,000 tokens per second. That pairing, a company selling both the coding agent and the model underneath it, is the more consequential part of this release than any single benchmark line. A year ago, most coding agents were thin wrappers around a frontier lab’s API. Cognition is training a model specifically for the Devin harness it runs, which means its RL pipeline, not just the product on top, becomes the thing competitors have to replicate.
That points to where competition in coding models is heading. If a startup can take someone else’s open-weight base model and push it into GPT-5.5 territory using its own post-training pipeline, frontier-level coding capability stops being a scarce input controlled by three or four labs. The differentiator shifts from who trains the smartest model to who can serve it fastest and cheapest inside a product engineers already trust with their codebase.
Teams evaluating coding agents this quarter should test SWE-1.7 inside Devin against Opus 4.8 and GPT-5.5 on their own repositories rather than on Cognition’s published scores, and weigh the 1,000-token-per-second Cerebras inference speed against the modest accuracy gap before locking in a 2026 vendor.
Cognition detailed SWE-1.7’s training pipeline, benchmark results, and Devin availability in a company blog post published July 8, 2026.
