Cognition launches SWE-2, a coding model near frontier quality at 64% lower cost

Cognition has released SWE-2, a new coding model that the company says pushes the cost-performance frontier for software engineering agents. According to Cognition's announcement, SWE-2 scores 50.0% on FrontierCode 1.1 Main — within one point of Fable 5.1 while being 64% cheaper — and lands within a few points of GPT-6 Astra at roughly a quarter of that model's cost. It is available starting now in Devin Desktop and Devin CLI, with a rollout under way on Devin Web and Fusion.

Benchmark position

Cognition's self-reported numbers place SWE-2 at 50.0% on FrontierCode 1.1 Main, ahead of Grok 4.6 (48.0%) and GPT-5.6 Sol (47.5%), marginally behind Fable 5.1 (50.9%) and a few points behind GPT-6 Astra (53.3%). On DeepSWE 1.1 it posts 73.0%, second only to GPT-6 Astra's 74.1% and just above GPT-5.6 Sol's 72.7%. On Terminal-Bench 2.1 it leads the comparison set at 92.8%.

The picture is less uniform on Terminal-Bench 4, a harder suite, where Cognition's table shows SWE-2 at 27.3% against 55.8% for Fable 5.1 and 57.9% for GPT-6 Astra. In other words, the cost advantage does not translate into parity everywhere, and the gap on this particular suite is wide.

Training approach

SWE-2 is post-trained from Kimi K3, a 2.8-trillion-parameter base model that Cognition notes had already undergone extensive RL for agentic coding. The company says its own RL run found substantial headroom, adding five to six points on many benchmarks and shifting the base model's entire cost-performance frontier. Cognition describes this as its first time scaling RL to the multi-trillion-parameter regime, building on the training infrastructure and recipe from SWE-1.7.

The central methodological claim is a cost-penalized reward that trains all reasoning-effort levels in a single RL run. The reward subtracts a linear cost penalty from the success signal, with the penalty coefficient tuned per effort level to match the local slope of the base model's Pareto frontier. Cognition argues that a linear penalty is the principled choice here because it makes the expected reward depend only on average cost and solve rate. The post also cites length-weighted reward baselines for training stability, an online draft model plus NVFP4/FP8 kernels and quantization-aware training to raise decoding throughput and cut memory use, a tripling of the number of RL environments, and a flywheel in which earlier SWE-2 checkpoints are used to iteratively harden the verifiers.

Efficiency and behavior

Efficiency is where the gains are most concrete. Cognition reports that SWE-2 medium scores higher than SWE-1.7 on FrontierCode 1.1 Main while taking 58% fewer turns and costing 81% less on average. It also makes its first real edit after a median of 18 steps, versus 48 for SWE-1.7 — a direct response to user feedback that SWE-1.7 tended to over-explore and overthink simple tasks before touching code.

Cognition lists several behavioral improvements observed in internal testing: better end-to-end test coverage that catches regressions and edge cases; more willingness to find alternate routes within user-set boundaries (in one case, reconstructing needed data from Slack channel history when an MCP integration was unavailable); and verification discipline, meaning the model re-derives conclusions when challenged and runs artifacts to gather evidence rather than accepting surface-level claims. The effort levels are reported to behave differently, with medium stepping into action quickly on simple and intermediate tasks, while high and max spend more on planning, exploration and verification for complex work.

Why it matters

Coding agents are unusually cost-sensitive because they burn many turns, file reads and test runs per task, so a model that sits near the capability frontier at a fraction of the price changes the economics of running them at scale. If Cognition's numbers hold up under independent evaluation, SWE-2 becomes a credible budget option rather than a compromise. The technique of training every effort level in one RL run is also worth watching for anyone building multi-tier models. At the same time, the Terminal-Bench 4 results are a useful corrective: near-frontier averages can hide large gaps on the hardest suites, and every figure published so far is the vendor's own.