SWE-1.7 Brings Near-Frontier Coding Power to Devin

SWE-1.7 Pushes Near-Frontier Coding Into Devin at a Fraction of the Cost

On July 8, 2026, Cognition released SWE-1.7, its most capable coding model yet, and the headline is genuinely exciting: frontier-class coding assistance is arriving at a price that puts it within reach of far more developers. SWE-1.7 was built through large-scale reinforcement learning on top of the already RL-trained Kimi K2.7 Code base — an approach Cognition calls "RL on RL" — and it ships directly inside Devin across Web, Desktop, and CLI. The result is a model that scores near the very top of today's coding benchmarks while completing tasks for roughly the price of a cup of coffee.

Key Takeaways

• 42.3% on FrontierCode 1.1 Main, up from just 9.4% for the predecessor SWE-1.6 — roughly a 4x jump in a single generation.
• ~$1.97 per task on FrontierCode Main, keeping frontier-class help remarkably affordable.
• ~1000 tokens/sec served on Cerebras hardware, so responses arrive fast inside Devin.
• Built via "RL on RL" — reinforcement learning layered on the RL-trained Kimi K2.7 Code base.

What Makes SWE-1.7 Different?

The most interesting part of this release is the training recipe. Rather than starting from a conventional pretrained base, Cognition applied its own large-scale reinforcement learning on top of Kimi K2.7 Code — a model that had *already* been shaped by reinforcement learning. Stacking a second RL stage on an RL-trained foundation is what "RL on RL" refers to, and the payoff shows up clearly in the numbers.

That 42.3% score on FrontierCode 1.1 Main is the standout figure. FrontierCode is designed to be hard, and the predecessor SWE-1.6 managed 9.4% on the same test. Quadrupling a benchmark result within one model generation is a strong signal that reinforcement learning still has meaningful headroom for coding tasks — a genuinely encouraging finding for the whole field. For more context on where models like this fit, see our broader AI coverage.

How Does RL-on-RL Work in Practice?

At a high level, reinforcement learning trains a model by rewarding good outcomes — in coding, that means solutions that compile, pass tests, and complete real engineering tasks. The Kimi K2.7 Code base already had this treatment, so it arrived with strong coding instincts. Cognition then ran another RL pass focused on the kinds of multi-step, tool-using work that Devin performs day to day.

Because the reward signal is grounded in real task completion, the second stage sharpens exactly the behaviors that matter for an autonomous coding agent: reading a repository, running commands in a terminal, and iterating toward a working fix. That focus is reflected in SWE-1.7's 81.5% on Terminal-Bench 2.1, a test of command-line competence, and 77.8% on SWE-Bench Multilingual, which measures fixing real issues across many programming languages.

Benchmarks in Context

SWE-1.7 lands impressively close to the frontier. On SWE-Bench Multilingual it edges ahead of GPT-5.5, and it trails Anthropic's Opus 4.8 by only a few points — a narrow gap for a model priced this accessibly. The constructive way to read these results is that the distance between the very best coding models and the broadly affordable ones keeps shrinking.

Why Cost and Speed Matter

At roughly $1.97 per task, SWE-1.7 makes sustained, iterative agent work economical rather than a luxury. Pair that with the ~1000 tokens/sec throughput from Cerebras, and Devin can plan, edit, and test with little waiting. Affordable plus fast is the combination that turns a capable model into a practical daily tool — the same democratizing trend we highlighted with Grok 4.5, an Opus-class coding model and Meta Muse Spark 1.1.

What It Means for Developers

For the growing community of engineers using autonomous coding agents, SWE-1.7 is a clear step forward. Delivered inside Devin on Web, Desktop, and CLI, it slots into existing workflows without new tooling, and its low per-task cost invites experimentation — running more agents, tackling bigger refactors, and letting the model iterate freely.

The bigger story is one of momentum. A 4x benchmark leap in one generation suggests the reinforcement-learning playbook for coding is far from exhausted, and each gain that arrives this affordably widens access to serious engineering help. SWE-1.7 is a bright, constructive marker of where AI-assisted software development is heading.

Sources: Cognition — SWE-1.7 blog — July 8, 2026; TechTimes — July 9, 2026; Winbuzzer — July 9, 2026.