Siarhei HolubLinkedInGitHub

Nov 4, 2026 · found Nov 8, 2026, 9:10 PM PT

Cognition SWE-2 Flash: cheap, fast coding model with vendor-run benchmarks

Verdict: Vendor-reported · Model releases · Agents and coding

The release is real and in Devin, but every number is Cognition's own, the harness is not named, and the speed claim comes from a simulated replay.

Cognition released SWE-2 Flash on 4 November, a low-cost coding model built on Thinking Machines’ Inkling-Small.

  • Claim: Terminal-Bench 4.0 20.0 and FrontierCode 2.0 24.0, with the larger SWE-2 at 27.3 and 23.8, at one-fifth the price. Cognition
  • Evidence: It is available in Devin Desktop and the Devin CLI, free through 30 November. Weights are not public, and no independent run was found. Cognition
  • Caveats: The post names no harness or runner. Cognition authored FrontierCode and sells the product being scored. The 200 tokens per second figure comes from a simulated replay of one task, and “one-fifth the price” gives no absolute prices. Cognition
  • Why it matters: Treat every number as unproven until someone else runs it. Cognition

Sources: Cognition, Introducing SWE-2 Flash (primary) · Artificial Analysis, Inkling 2 Small (independent)

Verdict history
  1. Nov 8, 2026, 9:10 PM PT: Vendor-reported. Primary post read. No independent run of SWE-2 Flash found. Artificial Analysis data on a different model (Inkling 2 Small) is context only.