Cognition has launched SWE-2, which the company calls its most advanced coding model to date, and claims it lands just one point behind Fable 5.1 on a key benchmark while costing 64% less. The announcement also says the model holds its own against GPT-6 Astra at a fraction of the price.
Starting today, SWE-2 can be found in Devin Desktop and CLI. Cognition is currently introducing it to Devin Web and Fusion.
The Benchmark Numbers
Cognition reports that SWE-2 scores 50.0% on FrontierCode 1.1 Main1, placing it just one point from Fable 5.1, a major rival. The firm claims SWE-2 outperforms its earlier model SWE-1.7 and Grok 4.6 across both score and cost on FrontierCode 1.1 Main and DeepSWE 1.1.
Cognition said the model matches GPT-5.6 Sol and Fable 5/5.1 for a smaller share of the price, and it falls short by just a few points of GPT-6 Astra while costing a quarter as much.
| Model | FrontierCode 1.1 Main1 Score | Relative Cost |
|---|---|---|
| SWE-2 | 50.0% | Baseline |
| Fable 5.1 | Within one point of SWE-2 | 64% more expensive |
| GPT-6 Astra | A few points higher | 4x SWE-2’s cost |
| SWE-1.7 | Lower | Higher than SWE-2 |
Training at a New Scale
According to Cognition, SWE-2 is the first model to scale reinforcement learning into the multi-trillion-parameter range. Its training process draws upon the infrastructure and recipe from SWE-1.72.
The core addition is an algorithm that applies reinforcement learning across every reasoning-effort level during a single training session. The company said this moves the entire cost-performance boundary forward, not just a single spot on it.
Cognition reports that the RL process behind SWE-2 found considerable room to grow after Kimi K33, a 2.8T-parameter model with extensive agentic coding training, was used as the starting point. The RL added 5-6 points on many benchmarks, according to cognition.
Faster, More Focused Behavior
Cognition explained that the efficiency gains stem from targeted exploration: the model determines which portions of a codebase are truly relevant to a given task and begins implementing sooner as a result.
SWE-2 medium first appears on FrontierCode 1.1 Main with a median of 18 steps since the last edit. SWE-1.7 required 48 steps. The new version also ranks above SWE-1.7, completing its task in 58% fewer turns and costing 81% less on average.
SWE-2 answers feedback from cognition that SWE-1.7 tended to over-explore and overthink on simple tasks.
Three Behavioral Improvements
Cognition observed several behavioral patterns in internal testing.
SWE-2 writes tests that check implementations end-to-end, which catches regressions and edge cases more reliably, as SWE-2 does for SWE-1’s test coverage. When the obvious path is blocked, SWE-2 finds another route, as shown by one case where an MCP integration was unavailable and the model reconstructed the data from Slack channel history it already had access to. Verification discipline means SWE-2 re-derives conclusions when challenged rather than re-asserting them; it verifies user hypotheses instead of simply agreeing.
The SWE-2 data reveals actual distinctions across various effort levels. A medium effort starts working on a simple task sooner. Higher and maximum efforts take more time to plan, do more exploration, and add more verification when dealing with complex work.
The Cost Penalty Method
Cognition’s training approach applies a linear cost penalty per effort level in a single RL run. Each penalty is tuned to the local slope of the base model’s Pareto frontier.
According to the company, the technique comes from first principles. Its goal is to move the model’s whole Pareto frontier forward while keeping its shape intact, and to mirror real user costs during training as closely as possible.
The method cognition adopted for its length-weighted reward baseline came from SWE-1.6, which it has relied on ever since, and it demonstrates how that approach steadies training.
Infrastructure and Rollout
Better scheduling and an online draft model were added to cognition’s RL rollout serving to increase decoding throughput. With NVFP4/FP8 kernels and quantization-aware training, memory usage was cut and train-inference mismatch stayed lower than SWE-1.7 at roughly the same throughput, even though the base model had nearly 3x more parameters.
The firm more than doubled the count of RL environments, put instruction-following overlays into place, and constructed a flywheel driven by earlier checkpoints of SWE-2 that work to steadily strengthen its verifiers.
According to the announcement, the cost penalty combines both the price of running inferences in dollars and the duration of the rollout.
What This Means
Cognition’s latest model, SWE-2, is its closest approach yet to the frontier. The company describes the release as a push toward the Pareto frontier, balancing capability against cost.
The model’s availability across Devin Desktop, CLI, Web, and Fusion means users can test the claims directly. Whether SWE-2’s benchmark scores translate to real-world coding performance remains to be seen, but the cost advantages are hard to ignore.
Source: cognition.com
Get the Notebook.
The day's best stories and every fresh verdict, in plain English, in your inbox by seven. One email a day, no more.