Cognition launched SWE-1.7 on Wednesday with a claim that will reshape how labs think about reinforcement learning: a model already extensively post-trained with RL by another team can be trained significantly further, producing large additional gains rather than the diminishing-returns wall the field has assumed was fundamental.
The core finding challenges conventional wisdom. Moonshot AI's Kimi K2.7 Code—the base Cognition started from—had already undergone extensive RL post-training. Conventional thinking says once you've done that, you're done. Additional RL should yield little; the policy becomes too deterministic, the model stops exploring, training stalls. Cognition extracted 12.2 additional percentage points on FrontierCode 1.1 Main anyway, climbing from Kimi's 30.1% to SWE-1.7's 42.3%. That's not diminishing returns. That's a recalibration.
The Methodology Matters More Than the Benchmarks
SWE-1.7 sits 0.7 points behind GPT-5.5 (43.0%) on FrontierCode and 4.2 points behind Claude Opus 4.8 (46.5%). On cross-vendor Terminal-Bench 2.1, it reaches 81.5% versus GPT-5.5's 84.2%. Those are solid positions, not headline positions—and Cognition is explicit about that. What matters is not that SWE-1.7 won the leaderboard. It's that it suggests the binding constraint in long RL runs is not some fundamental property of the base model. It may be infrastructure quality.
If that holds under independent replication, labs sitting on powerful open-weight base models can extract substantially more capability without waiting for a new pre-trained generation. That's a meaningful shift in the feasibility calculus for smaller teams.
Four Engineering Moves That Made It Possible
Cognition documented four techniques that kept the RL run from collapsing—each addressing a distinct failure mode that typically terminates long training runs.
Entropy preservation via top-p sampling replay. Long RL runs fail through entropy collapse: the model stops exploring, probability mass narrows, reward plateaus within a few hundred steps. The culprit is low-probability tokens. When a model samples an off-distribution token and earns low reward, the resulting gradient update disproportionately sharpens the overall token probability distribution, accelerating collapse. Cognition's fix: top-p sampling restricts rollout sampling to the highest-probability subset. But naive top-p creates a training-inference mismatch—the trainer sees all tokens, rollouts only see top-p tokens. Cognition's solution was sampling distribution replay: during rollouts, record which tokens were available (the "kept-set"), then use those same masks to renormalize probabilities in the trainer. Entropy stayed roughly constant across the run—the first published methodology to report this property at this scale, to Cognition's knowledge. They also credit the Muon optimizer and eliminating non-deterministic operations.
Multi-cluster RL across three continents. Large RL runs require enormous inference compute to generate training rollouts. Single-network cluster access is scarce. Cognition exploited a structural property: only the trainer needs to live on a single high-bandwidth cluster. Inference engines are self-contained—they need only current model weights. SWE-1.7 trained across four datacenters on three continents, combining Cognition's GPU clusters with Fireworks AI's multi-cluster approach. The synchronization challenge was keeping inference engines aligned with the trainer's latest weights. Solution: every K gradient steps, the trainer computes and transmits only the weight delta—the difference between current and previous weights—not the full model. The delta reduces transfer size by more than 99%. For a trillion-parameter model, cross-continental weight updates complete in one to two minutes end-to-end with just three to four seconds of inference downtime.
Fault tolerance at scale. Hardware failures are continuous at this scale. Inference engines are stateless: a dead node costs only its in-flight sessions, which NVIDIA Dynamo's router replays on healthy workers. The trainer—the tightly coupled piece—checkpoints asynchronously to local disk on every step, replicating shards to peers so a dead node's state rebuilds in seconds. If capacity shrinks, the run scales down by full data-parallel replicas and regrows when nodes return.
Self-compaction for long-horizon tasks. Devin tasks extend well beyond raw context windows. Cognition trained the model to summarize its working state when approaching the limit, then resume from its own summary. During SWE-1.7 training, rollouts reached up to six hours. To counter RL's documented tendency to produce progressively longer outputs, Cognition alternated between unconstrained phases optimizing for task success and budget phases penalizing solutions exceeding a weighted cost function of tokens, turns, and tool-call time.
Access and Cost
SWE-1.7 runs at approximately $1.97 per task on FrontierCode Main. It is accessible exclusively through Devin (Web, Desktop, and CLI) at 1,000 tokens per second via Cerebras. There is no direct API for calling SWE-1.7 outside Devin. Custom workflows must route through Devin's agentic environment.
Cognition grew from $37 million ARR in May 2025 to $492 million in May 2026—a 13-fold jump in 12 months—with enterprise usage up more than tenfold since January 2026. The company counts Goldman Sachs, Mercedes-Benz, NASA, and Santander among customers.
What Independent Research Will Actually Test
Every benchmark cited here—SWE-1.7, GPT-5.5, Claude Opus 4.8—is self-reported by the respective labs. FrontierCode 1.1 is Cognition's own benchmark. No independent third-party audit exists. That matters. The benchmark contamination research is extensive: OpenAI abandoned SWE-bench Verified after finding roughly 60% of a sampled problem set had material flaws.
But the RL-on-RL finding is falsifiable. Cognition made a specific claim: domain-specific RL stacking on top of an already heavily RL-trained base consistently produces large additional gains. The research community will now run similar experiments. Over the next few quarters, we'll know whether this is a generalizable property of RL infrastructure quality or specific to Cognition's task distribution and training environment. That's the real test, not the leaderboard position.