Cognition's SWE-2 scores 92.8 on Terminal-Bench 2.1 with 2.8T-parameter RL-trained model

Cognition has released SWE-2, a coding agent model that the company says scores 92.8 on Terminal-Bench 2.1 — the highest figure in the model's published benchmark table. The details come from a Tokenstead model page that reached the Hacker News front page, and every number in it is flagged as vendor-reported, pending independent replication.

A 2.8-trillion-parameter MoE on a Kimi K3 base

According to Tokenstead, SWE-2 is a mixture-of-experts model with 2.8 trillion total parameters and 104 billion active per token. The foundation is Kimi K3, which Cognition then post-trained on top. Tokenstead notes that the base was already heavily tuned with reinforcement learning for agentic coding, and that Cognition's training pass added roughly five to six points across most benchmarks. This also marks the first time Cognition has pushed RL training into multi-trillion-parameter territory.

Self-reported benchmark results

The scores listed by Tokenstead, all attributed to Cognition's own model card, are:

• FrontierCode 1.1 Main: 50.0
• DeepSWE 1.1: 73.0
• Terminal-Bench 2.1: 92.8
• Terminal-Bench 4.0: 27.3

The FrontierCode result puts SWE-2 one point behind Claude Fable 5.1 (50.9) and 3.3 points behind GPT-6 Astra (53.3) — while Cognition claims a 64% lower cost than Fable 5.1 and about a quarter of Astra's cost. The clear weak spot is Terminal-Bench 4.0, where SWE-2's 27.3 sits far below Fable 5.1 (55.8) and GPT-6 Astra (57.9). Per Tokenstead's read of the numbers, long-horizon agentic work is where the gap to the frontier models still lives.

Fewer steps per task, lower cost per run

Tokenstead lists mean steps per task of 53 at medium effort, 80 at high and 98 at max, compared with 127 for the previous SWE-1.7. At medium effort, SWE-2 reportedly beats SWE-1.7's FrontierCode score using 58% fewer turns and 81% lower average cost. It also reaches its first meaningful code edit at a median of step 18, where SWE-1.7 needed until step 48.

The serving stack is described as MoE inference on NVFP4 and FP8 kernels with quantization-aware training, with FP8 handling the K, Q, V and score computations in the MLA layers. A draft model retrained with SpecForge delivers 15% longer accept lengths for speculative decoding, and a prefill delayer improves throughput per GPU and tokens per second per request by 10–20%, at the expense of slower time-to-first-token.

Proprietary weights, Devin-only availability

Cognition has not released SWE-2's weights, so there is nothing to download and no local deployment path. There is also no per-token API pricing, which means the cost comparisons — such as the 64% saving versus Fable 5.1 at FrontierCode parity — are framed per task rather than as a rate card. According to Tokenstead, the model is available now in Devin Desktop and the Devin CLI, with a rollout underway on Devin Web and Fusion. The model is listed as proprietary, developed in the USA and released in September 2026.

Why it matters

SWE-2 is a coding-agent vendor shipping its own RL-trained model at multi-trillion-parameter scale, rather than building purely on someone else's frontier API — a meaningful shift in how agentic coding products are built. The claimed price-performance, near FrontierCode parity at a fraction of the cost of frontier models, could pressure pricing across the category if it survives third-party testing. The benchmark split is equally telling: 92.8 on Terminal-Bench 2.1 against 27.3 on Terminal-Bench 4.0 suggests shorter-horizon terminal tasks are approaching saturation, while long-horizon agentic work remains genuinely unsolved. And with closed weights, no per-token API and only self-reported numbers, independent verification is currently the main missing piece.