Cognition's SWE-2 Beats GPT-5.6 Sol at 64% Lower Cost

Cognition just released SWE-2, its most capable software engineering model to date, and the pitch is aggressive: frontier-level agentic coding at a fraction of the price. On FrontierCode 1.1 Main, SWE-2 scores 50.0%, within one point of Fable 5.1 while costing 64% less. The company also claims it edges out Grok 4.6 and matches GPT-5.6 Sol at meaningfully lower cost, coming within a few points of GPT-6 Astra at roughly a quarter of the price.

The more interesting part sits under the hood. SWE-2 is Cognition's first model with configurable reasoning effort levels, and the entire cost-performance curve was optimized inside a single reinforcement learning run using a mathematically motivated cost penalty. That is a departure from how most labs handle effort tiers today.

The headline numbers

SWE-2 is post-trained from Kimi K3, a 2.8T-parameter model that had already undergone extensive RL for agentic coding. Even on top of that strong starting point, Cognition's RL recipe reportedly adds 5-6 points across many benchmarks. Here is where it lands:

Benchmark	SWE-2	Kimi K3	Grok 4.6	Fable 5.1	GPT-5.6 Sol	GPT-6 Astra	SWE-1.7
FrontierCode 1.1 Main	50.0%	44.2%	48.0%	50.9%	47.5%	53.3%	42.0%
DeepSWE 1.1	73.0%	68.5%	67.5%	67.4%	72.7%	74.1%	37.7%
Terminal-Bench 2.1	92.8%	88.3%	88.4%	91.4%	88.8%	89.9%	81.5%
Terminal-Bench 4	27.3%	21.5%	20.3%	55.8%	37.3%	57.9%	7.6%

Terminal-Bench 4 is the notable weak spot, with SWE-2 well behind Fable 5.1 and GPT-6 Astra on the harder terminal tasks, so this is not a universal frontier claim. A fairer summary: SWE-2 dominates the price-performance chart while conceding raw ceiling on the hardest agentic problems.

Where the efficiency actually comes from

The most tangible engineering win is that SWE-2 stops over-thinking. On FrontierCode 1.1 Main, SWE-2 medium scores higher than SWE-1.7 while taking 58% fewer turns and costing 81% less on average. Cognition attributes this to what they call focused exploration: the model has developed better judgment about which parts of a codebase actually matter for a given task.

The behavioral difference shows up in a concrete metric. On FrontierCode 1.1 Main, SWE-2 medium makes its first real edit after a median of 18 steps, compared with 48 for the previous generation. Internal testing surfaced a few other patterns worth flagging:

• Better test writing: SWE-2 is more reliable at writing tests that check an implementation end-to-end, catching regressions and edge cases.
• Resourcefulness with guardrails: When the obvious path is blocked, SWE-2 is more willing to look for another route to the same answer, in one case reconstructing data from Slack channel history when an MCP integration was unavailable.
• Verification over agreement: When challenged, the model re-derives conclusions rather than re-asserting, checks user hypotheses instead of simply nodding along, and runs artifacts to gather evidence.

Training all three effort levels in one RL run

This is the real methodological contribution. Most labs train separate models or separate reasoning modes and stitch them together. Kimi K3, for reference, trains a separate expert per domain-effort combination and then distills them. Cognition took a different route.

They use a single reward function of the form R = S - λ_e * C, where S is a binary success signal, C is the actual dollar-and-time cost of the rollout, and λ_e is a per-effort-level penalty coefficient. The clever bit is how they set λ_e: it is tuned to match the local slope of the base model's Pareto frontier at that effort level.

The intuition is geometric. If your iso-reward line is tangent to the current Pareto frontier at a given effort point, then any reward improvement necessarily pushes the frontier outward rather than sliding along it. Set λ too high at the high-effort tier and you get a degenerate outcome where the high-effort model collapses into the medium-effort model to save cost. Cognition also proves in an appendix that a linear penalty is the only choice that preserves the property of depending only on average cost and average solve rate, which is a nice first-principles result.

The upshot: one RL run advances the whole cost-performance frontier at once. Medium becomes cheaper and smarter, while max learns to spend more tokens and turns in exchange for higher scores.

Variance reduction and inference plumbing

A second piece of the recipe is a length-weighted reward baseline that Cognition has quietly been using since SWE-1.6. The optimal variance-minimizing baseline requires an extra backward pass per rollout to compute gradient norms, which is expensive. They observed that the squared gradient norm correlates strongly with rollout length, so they approximate the optimal baseline as a length-weighted average of rewards. It costs nothing extra and keeps the inference-training KL divergence low during RL, which matters more as models get larger and drift becomes harder to control.

On the serving side, they trained an online draft model for speculative decoding using SpecForge so the draft keeps tracking the policy as it updates during RL, recovering roughly 15% longer accept lengths. They also moved to NVFP4 and FP8 kernels with quantization-aware training. The net effect: throughput similar to SWE-1.7 despite a base model nearly 3x larger.

Availability and pricing

SWE-2 ships today inside Devin Desktop and Devin CLI, with rollout to Devin Web and Fusion in progress. Cognition is making it free for all Pro, Max, and Teams subscribers for the next month, an aggressive attempt to pull developers off Claude Code and Codex during evaluation windows.

The three effort levels have distinct personalities worth knowing before you pick one:

1. Medium: Steps into action quickly. Best for simple to intermediate tasks where cost matters. This is where the 81% cost reduction versus SWE-1.7 shows up.
2. High: More planning, more codebase exploration, better handling of ambiguous requirements.
3. Max: Maximum turns and tokens, aimed at the hardest tasks with complex verification chains.

What this changes about the frontier conversation

Two things stand out. First, the gap between an open-weight base model plus a strong RL recipe and closed frontier labs keeps narrowing. Cognition is post-training on Kimi K3 and landing within striking distance of GPT-6 Astra on agentic coding benchmarks. That is a data point about how much of the frontier is base-model quality versus post-training craft.

Second, the multi-effort single-run recipe is likely to get copied. If you can push an entire Pareto curve with one RL run instead of training and distilling separate experts, that changes the economics of shipping tiered products. The first-principles derivation Cognition published makes it hard to argue you should be doing something else.