GoKawiilCognition's SWE-2 model scores 92.8 on Terminal-Bench 2.1, trails on longer tasks
Cognition released SWE-2, a coding model built by post-training on a 2.8-trillion-parameter Kimi K3 base with 104B active parameters, marking its first RL push into the multi-trillion-parameter range. The model posts a class-leading 92.8 on Terminal-Bench 2.1 and competitive FrontierCode scores at lower cost than rivals like Claude Fable 5.1 and GPT-6 Astra, but falls well behind both on the harder Terminal-Bench 4.0 benchmark (27.3 vs 55.8 and 57.9). SWE-2 is closed-weight and available now through Devin Desktop and CLI, with no local deployment option.
GoKawiil's interpretation of the reporting above, not reported fact.
The results suggest Cognition can match or beat frontier labs on short-horizon coding tasks while running cheaper, thanks to serving optimizations like FP8/NVFP4 quantization and speculative decoding tweaks. But the steep drop-off on long-horizon agentic benchmarks shows that scaling RL on a big MoE base hasn't closed the gap on sustained, multi-step reasoning, which remains the harder problem for autonomous coding agents. Keeping weights proprietary also means enterprises evaluating SWE-2 must rely entirely on Cognition's hosted product rather than independent verification or self-hosting.
- SWE-2 hits 92.8 on Terminal-Bench 2.1, the top score in Cognition's published comparison table
- It trails Claude Fable 5.1 and GPT-6 Astra by a wide margin on Terminal-Bench 4.0, exposing weakness in long-horizon agentic tasks
- Cognition claims major cost savings versus rivals but keeps SWE-2's weights closed, offering it only via Devin Desktop and CLI
Keychron K2 Wireless Mechanical Keyboard — If you're spending your days steering coding agents through terminal benchmarks, the Keychron K2 gives you a compact 75% mechanical layout with dedicated function row for all those terminal shortcuts. It pairs over Bluetooth or USB-C and works with Mac and Windows layouts, so it slots into whatever machine your agent runs on. Hot-swappable switch options mean you can tune the feel for long sessions of prompt-and-review.
See Keychron K2 Wireless Mechanical Keyboard on Amazon → Affiliate link — we may earn a commission on purchases, at no extra cost to you. Product picked by AI based on this article; it is not a tested recommendation.Source: tokenstead.ai, 2026-09-10
Published there as: “Cognition's SWE-2 achieves 92.8 on Terminal-Bench 2.1”
Read the original report → The summary and analysis above are GoKawiil's own, written from reporting by the source above. Facts and quotes belong to the original publisher.