Changes confirmed high confidence

Cognition Ships SWE-1.7: RL-Trained Openly on Moonshot's Kimi K2.7, Served at 1,000 Tokens/Second via Cerebras

A US agent company built its flagship coding model on a Chinese open-weight base — disclosing it up front — and matched near-frontier performance at a fraction of per-task cost.

Cognition released SWE-1.7 on July 8, 2026, inside its Devin agent (Web, Desktop and CLI), disclosing up front that the model was RL post-trained on Moonshot AI's open-weight Kimi K2.7 Code — a reversal of the hidden-base controversy around SWE-1.5 — per explainx.ai's launch coverage.

Context

SWE-1.5 drew criticism for shipping on a concealed GLM base. With SWE-1.7, Cognition made the base model a selling point: training ran across four data centers on three continents, with self-compaction enabling roughly six-hour tasks.

What changed

Why it matters

The open-model supply chain has inverted assumptions about the frontier: a US agent company now builds its flagship on a Chinese open-weight base, optimizes cost-per-accepted-task rather than leaderboard position, and uses specialized inference silicon for speed. Near-parity with GPT-5.5 at a fraction of per-task cost pressures the closed labs' coding-agent pricing.

Details

Corroborating coverage appeared in WinBuzzer (July 9) and Radical Data Science (July 10). The 1,000 tok/s figure reflects Cerebras serving; per-task cost derives from Cognition's FrontierCode accounting.

Market context

SWE-1.7 shipped the same day as OpenAI's GPT-Live and ByteDance's Seedream 5.0 Pro, and one day before GPT-5.6 — part of a nine-day stretch in which frontier-adjacent input prices spread roughly 12x across a narrow capability band. Cognition's bet is that procurement increasingly routes on cost-per-accepted-task: at roughly $1.97 per FrontierCode task, SWE-1.7 undercuts hosted frontier agents by a wide margin while conceding only a few points of benchmark accuracy. The three-continent training run also demonstrates that open-weight bases now support serious post-training programs by companies other than the original lab.

Limitations and caveats

FrontierCode is a benchmark family associated with Cognition's own ecosystem, so cross-vendor comparisons carry provenance caveats; Terminal-Bench 2.1 figures are vendor-reported. Real-world Devin task performance may differ from benchmark conditions.

Sources

*Update note: This post was last reviewed on 2026-07-22. The free month for paid Devin users ends in early August; watch for API availability announcements.*

Sources

Drafted with AI assistance from source briefs; reviewed for citation completeness and label accuracy.