SWE-1.7: Cognition's RL-on-RL Training Lands at 81.5% Terminal-Bench — And Why You Can't Call It Directly

Cognition shipped SWE-1.7 on July 8, 2026 — “the most capable model we’ve trained so far,” in the company’s own words, available today in Devin at 1,000 tokens per second via Cerebras. The headline is benchmark performance at a fraction of frontier cost. But the more interesting story is how they got there: reinforcement learning on top of a model that was already heavily RL-trained. That breaks with the conventional assumption that RL has a ceiling.

Here’s what builders need to know.

Where SWE-1.7 Sits in the July 2026 Model Landscape

The benchmark that matters for direct comparison is Terminal-Bench 2.1. Cognition’s own release post puts SWE-1.7 against its own base model and its closest frontier competitors — not against Grok or Sonnet, which don’t appear in Cognition’s comparison at all:

(All scores as self-reported in Cognition’s SWE-1.7 announcement.)

SWE-1.7 lands just above GLM-5.2 and well above its own Kimi K2.7 Code base, Composer 2.5, and Cognition’s prior model SWE-1.6 — but it’s still behind Opus 4.7, GPT-5.5, and Opus 4.8. The story isn’t “beats frontier models,” it’s “closes most of the gap to frontier models at a fraction of the cost.”

Cognition’s own FrontierCode 1.1 benchmark, also from the same release post, gives a fuller picture of agentic coding specifically:

SWE-1.7 sits 0.7 points behind GPT-5.5 and 4.2 points behind Opus 4.8. More strikingly, it’s 12.2 points above its own starting point — the Kimi K2.7 Code base model it was trained on.

It also scores 77.8% on SWE-Bench Multilingual, Cognition’s cross-language coding benchmark.

The RL-on-RL Breakthrough

This is the technical claim worth taking seriously.

The conventional wisdom has been that once a base model is extensively RL post-trained, you approach a ceiling. Additional RL training on an already-RL-optimized model yields diminishing returns, partly because the model’s entropy collapses — the policy becomes too deterministic and stops exploring new strategies.

Kimi K2.7 Code — Moonshot AI’s 1-trillion-parameter MoE model with 32 billion active parameters, released as open weights on June 12, 2026 — had already undergone extensive RL post-training before Cognition touched it. That makes SWE-1.7 a direct test of whether additional RL training stacks.

The result: a 12.2-point jump on FrontierCode 1.1 over the base model. Cognition’s own framing: “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.”

How They Avoided Entropy Collapse

The key technical innovation is entropy preservation using top-p sampling with sampling distribution replay. During RL training, as the model improves, it tends to become overconfident — repeatedly choosing the same high-probability actions and losing the diversity that drives further learning. Cognition’s approach replays the sampling distribution from earlier in training, keeping policy entropy high enough to keep learning.

Self-Compaction for Long Agentic Tasks

Agentic coding tasks can run for hours. SWE-1.7’s training rollouts ran up to six hours — far beyond what fits in a raw context window. Cognition solved this with self-compaction: the model periodically summarizes its own working state and picks up from that summary, using an alternating length penalty to drive efficient compression. This means SWE-1.7 can maintain coherent task progress across very long horizons without hitting context limits.

Multi-Cluster Training Infrastructure

To run RL at this scale, Cognition trained on clusters across three continents, shipped weight updates through object storage, and built fault tolerance so that hardware failures never stalled the run.

Data Curation

Cognition built an automated data-quality pipeline that runs each task through execution tests, filters out tasks with low learning signal, and hardens tasks against reward-hacking — preventing the model from gaming its training rewards rather than learning genuine task completion.

Cost: $1.97 Per Task

On FrontierCode’s Main evaluation set, Cognition reports SWE-1.7 costs roughly $1.97 per task on average — a figure attributed to Cognition but presented as a chart data point in their release post rather than a plain-text price. Cognition frames this as advancing the cost-performance Pareto curve, meaning you’d pay significantly more per task with GPT-5.5 or Opus 4.8 for a small additional benchmark gain.

At 1,000 tokens/second via Cerebras, the model is also fast enough for interactive coding workflows rather than just batch background work.

The Caveat Builders Need to Know

SWE-1.7 is not an API model. You cannot call it directly.

It’s exclusively available through Devin — Cognition’s agentic coding product (web, desktop, and CLI). This is a deliberate product decision: Cognition is not selling model API access, they’re selling a coding agent service.

What that means for your stack:

• If you’re already using Devin for agentic coding tasks, you get SWE-1.7 automatically — it’s the new default model
• If you want to build agents with this architecture yourself, you need to wait for the techniques (RL-on-RL, self-compaction, entropy preservation) to appear in open-weights or API-accessible models
• Kimi K2.7 Code itself is open weights, so the base model is available — but SWE-1.7’s trained weights are not

What This Means for Builders Not on Devin

Even if you’re not a Devin user, this release matters for a few reasons:

The RL-on-RL insight will spread. If Cognition’s entropy preservation technique works, other labs will replicate it. Expect to see “extended RL post-training” appear as a differentiator in future model releases from Mistral, DeepSeek, and others working with open base models.

Self-compaction is an architecture builders should watch. The idea of a model that summarizes its own working state to extend effective task horizon is directly applicable to agentic system design — not just as a training technique but as an inference-time strategy. If you’re building long-horizon agents today, you’re doing a manual version of this with prompt engineering; future models may do it natively.

Benchmark positioning matters. SWE-1.7 at 81.5% Terminal-Bench makes the case that frontier-competitive agentic coding performance is no longer exclusive to the top-tier API models. If Cognition eventually opens an API (or if the techniques appear in open models), the competitive dynamics of the model routing market change.

Bottom Line for Builders

SWE-1.7 is a legitimate technical achievement. The RL-on-RL result challenges conventional assumptions about post-training ceilings, the self-compaction technique is novel, and the Terminal-Bench position (81.5%, within a few points of GPT-5.5 and Opus 4.8) is credible at a claimed $1.97/task cost.

The access constraint is real: you get SWE-1.7 by using Devin, not by calling an API. If Devin fits your workflow, this upgrade is meaningful. If you’re building directly on model APIs, track the techniques rather than the model.

ChatForest covers AI model releases and builder implications as they happen. The benchmark data in this article is sourced from Cognition’s official SWE-1.7 announcement and cross-referenced against BenchLM.ai.