Cognition SWE-1.7
Cognition's most capable SWE model — reinforcement learning stacked on RL, served at 1,000 tokens per second.
About
SWE-1.7 launched July 8, 2026 as the most capable model Cognition had trained, continuing the SWE line that began at Windsurf (acquired by Cognition in 2025). The headline claim is methodological: a model already extensively post-trained with RL by one lab can be trained significantly further by another, producing large additional gains rather than hitting the diminishing-returns wall the field widely assumed. It shipped immediately in Devin across Web, Desktop, and CLI at 1,000 tokens per second via Cerebras inference. Earlier in the line, the SWE-1.5 'Fast Agent Model' was reported at up to 950 tokens per second. Public benchmark figures for 1.7 itself are thin in the sources read — the interesting artifact here is the training recipe claim, which the research community will be testing.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
No public benchmark figures in sources read
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (metrics tier, metrics-v1). Evidence: Launch date verified: TechTimes article dated 2026-07-09 says 'launched on Wednesday' = 2026-07-08 (confirmed Wednesday via date check); 1,000 tok/s via Cerebras and RL-on-RL methodology claim from the same report; Winbuzzer corroborates SWE-1.5 speed lineage; No public benchmark figures for 1.7 itself in sources read — thin quantitative evidence.