Cognition has launched SWE-1.7, its most capable coding model to date, scoring 42.3% on FrontierCode 1.1 Main — within striking distance of GPT-5.5 (43.0%) and Anthropic’s Opus 4.8 (46.5%). It’s available today in Devin via Cerebras at 1,000 tokens per second.
🔍 THE BOTTOM LINE
SWE-1.7 was trained from Kimi K2.7’s base — an open-source model that had already undergone extensive RL post-training. Cognition then applied its own RL pipeline on top, achieving large additional gains. This directly challenges the idea of a “post-training ceiling” and suggests RL can push capabilities far further than previously believed. The cost-performance gap between frontier labs and well-engineered RL shops is shrinking fast.
The Numbers That Matter
| Benchmark | SWE-1.7 | Kimi K2.7 Code | GPT-5.5 | Opus 4.8 | GLM-5.2 |
|---|---|---|---|---|---|
| FrontierCode 1.1 Main | 42.3% | 30.1% | 43.0% | 46.5% | 24.5% |
| Terminal-Bench 2.1 | 81.5% | 72.7% | 84.2% | 86.9% | 81.0% |
| SWE-Bench Multilingual | 77.8% | 73.5% | 76.8% | 84.4% | 74.5% |
The jump from SWE-1.6 (9.4% on FrontierCode) to SWE-1.7 (42.3%) is massive — a 4.5x improvement. And the gap to GPT-5.5 is now 0.7 percentage points. The gap to Opus 4.8 is 4.2 points. Both are close enough that cost differences become the deciding factor.
How They Did It: Four Breakthroughs
Cognition’s blog post details four technical components that drove the improvement:
1. Entropy preservation and training stability. Long RL runs face entropy collapse — the model stops exploring and reward plateaus. Cognition implemented top-p sampling with “sampling distribution replay,” recording which tokens were available for sampling at rollout time and renormalizing probabilities with those masks in the trainer. This keeps entropy roughly constant throughout training, where naive implementations collapse within hundreds of steps.
2. Multi-cluster training across three continents. RL decomposes naturally — only the trainer needs a single high-bandwidth cluster. Inference engines can run anywhere. Cognition trained across four datacenters on three continents, using compressed weight deltas through cloud object storage. Cross-continental weight updates for a trillion-parameter model complete in 1-2 minutes with only 3-4 seconds of inference downtime.
3. Self-compaction for long-horizon tasks. When an agent approaches the context limit, it summarizes its own working state and resumes from that summary. During training, the model learns both to write better summaries and to work from them more effectively. Rollouts during SWE-1.7 training reached up to six hours in duration.
4. Data quality at scale. Cognition built extensive pipelines to filter out tasks with low learning signal, harden tasks against reward-hacking, and isolate grading from the agent. Cheating attempts — regardless of success — received zero reward.
The Post-Training Ceiling Is a Myth
The most provocative claim in Cognition’s announcement: “Since SWE-1.7 was trained from a Kimi K2.7 base, which had already undergone extensive RL post-training, the large additional gains from our own training challenge the idea of a ‘post-training ceiling’ and suggest that RL can push capabilities much further than previously believed.”
This is a direct challenge to the narrative that models plateau after initial training. The open-weights community has been closing the gap — we covered GLM 5.2 terrifying frontier labs, DeepSeek V4 matching GPT-5.5, and Kimi K2.6 beating Claude and GPT on coding. SWE-1.7 takes an open-source base model and, through RL engineering alone, nearly matches the frontier.
Behavioral Shift: Explore Before You Act
SWE-1.7 exhibits noticeably different behavior from its Kimi K2.7 base. It explores codebases more thoroughly before acting — more tool calls, more file reads, more searches. It probes edge cases, adversarial inputs, and hypotheticals. Its chain-of-thought is condensed: nearly half the function-word ratio and half the average sentence length of K2.7 Code.
This “explore first” behavior is a direct consequence of the data quality pipeline. By stripping out false positives and false negatives in training data, the model was forced to produce complete, end-to-end solutions — which requires understanding the full context before making changes. The trade-off is increased change scope: SWE-1.7 touches more files than strictly necessary, writing additional test cases and exploring more code. Cognition acknowledges this as an axis to improve.
The SpaceX-Cursor Context
This launch comes weeks after SpaceX acquired Cursor for $60 billion, consolidating the AI coding market. The competitive landscape now has Cognition (Devin), SpaceX-Cursor (Composer), Anthropic (Claude Code), and OpenAI (Codex) as the major players. SWE-1.7 narrows the gap to the top two — and does it with a fraction of the compute budget.
❓ FAQ
Q: Is SWE-1.7 available outside Devin? Currently, it’s available in Devin (Web, Desktop, and CLI) via Cerebras at 1,000 TPS. No API release was announced.
Q: How does it compare to Claude Code? Claude’s Opus 4.8 scores 46.5% on FrontierCode 1.1 vs SWE-1.7’s 42.3%. On Terminal-Bench 2.1, Opus 4.8 leads 86.9% to 81.5%. SWE-1.7 is close but not ahead — the advantage is cost, not raw performance.
Q: What is FrontierCode 1.1? Cognition’s own agentic coding benchmark, evaluating models on real-world software engineering tasks. Full methodology is in their FrontierCode blog post.
Q: Can I use SWE-1.7’s training approach on other base models? The techniques (entropy preservation, multi-cluster RL, self-compaction) are described in detail. Cognition trained from Kimi K2.7’s open-weights base, but the RL pipeline is their proprietary infrastructure.
Q: What does “trained from Kimi K2.7 base” mean for safety? Cognition published a companion post on “Measuring the Trustworthiness of Open-Source-Derived Models.” They report SWE-1.7 is “significantly more aligned and trustworthy than K2.7 or other frontier open-source models.”
🔍 THE BOTTOM LINE
SWE-1.7 proves that a well-engineered RL pipeline applied to an open-source base model can nearly match the frontier — 42.3% vs GPT-5.5’s 43.0% on FrontierCode 1.1. The “post-training ceiling” is a myth. The implication for the industry is clear: the moat isn’t the base model anymore. It’s the RL training infrastructure, the data quality pipeline, and the agent harness. Frontier labs betting on model size alone should be worried.
