Cognition put out SWE-2 a few hours ago, the newest model behind its Devin coding agent, and the numbers it's publishing put the model within a few points of OpenAI's GPT-6 Astra on Cognition's own coding benchmark while running at roughly a quarter of the price. The company says it beats its previous model, SWE-1.7, and xAI's Grok 4.6 on both score and cost, and ties Fable 5/5.1 and GPT-5.6 Sol at a fraction of what those models charge per task.

That's a big claim from a company that also grades its own homework. Cognition built the FrontierCode benchmark it's citing, so the numbers deserve a skeptical read before anyone rewrites their agent stack around them. But the underlying method is worth understanding regardless of whose scoreboard it's measured on, because it's a genuinely different way to train a coding model than "throw more RL at it and hope."

RelatedClaude Sonnet 5: Near-Opus Coding at Half the Price

  • SWE-2 is post-trained from Moonshot's 2.8-trillion-parameter Kimi K3, not built from scratch.
  • It ships in three effort tiers, medium, high, and max, so a task can trade speed for thoroughness.
  • On Cognition's FrontierCode 1.1 benchmark it scores 50.0%, half a point behind Fable 5.1's 50.9% and 3.3 points behind GPT-6 Astra's 53.3%, at a fraction of the inference cost.
  • It's live today in Devin Desktop, CLI, Web, and Fusion, no waitlist mentioned.

What did Cognition actually announce?

SWE-2 replaces SWE-1.7, which shipped in July after Cognition raised prices and reputation on the strength of its coding-agent benchmarks. The new model's headline numbers: 50.0% on FrontierCode 1.1 Main (SWE-1.7 scored 42.0%), 73.0% on DeepSWE 1.1 (up from 37.7%), and 92.8% on Terminal-Bench 2.1 (up from 81.5%). Terminal-Bench 4, a harder and newer variant, jumped from 7.6% to 27.3%. Those are large single-generation gains, and Cognition attributes most of them to how the model was trained rather than to a bigger base model.

The efficiency story is arguably the bigger deal than the raw scores. On FrontierCode tasks, SWE-2's medium tier averages 53 steps to solve a problem where SWE-1.7 needed 127, a 58% cut. Cognition says that translates into an 81% average cost reduction against SWE-1.7, and a 64% discount against Fable 5.1 for a score within one point of it.

How does the training actually work?

This is the part most coverage of AI model launches skips, because "novel RL techniques" is a phrase you can write without understanding anything. Cognition's version is specific enough to check: they train with a reward function R = S − λₑC, where S is task success, C is the dollar-and-time cost of a rollout, and λₑ is a penalty weight tuned separately for each effort level to match the slope of that level's own cost-performance curve.

The team says they proved a linear penalty is the only form that gives the same result whether you apply it before or after averaging across a batch of rollouts, which is a real mathematical property, not marketing language, and it's what lets them train medium, high, and max as one coherent family instead of three separately tuned models. In plain terms: instead of rewarding the model only for getting the right answer, they're rewarding it for getting the right answer cheaply, with "cheaply" defined differently depending on which effort tier is generating the rollout.

How SWE-2 was trained Diagram showing Kimi K3 post-trained with a cost-penalized reinforcement learning reward to produce SWE-2 at three effort levels. KIMI K3 2.8T params base model RL POST-TRAINING cost-penalized reward R = S − λₑ·C λₑ fit to each tier's cost-performance slope SWE-2 medium / high / max 3 effort tiers genztech.blog
Fig 1 Cognition post-trains Kimi K3 with a cost-penalized reward, then splits the result into three effort levels rather than training three separate models.

How does SWE-2 actually stack up?

Here's the comparison Cognition is publishing, all on FrontierCode 1.1 Main:

SWE-1.7Grok 4.6GPT-5.6 SolSWE-2Fable 5.1GPT-6 Astra
FrontierCode 1.1 Main42.0%48.0%47.5%50.0%50.9%53.3%
DeepSWE 1.137.7%67.5%72.7%73.0%67.4%74.1%
Terminal-Bench 2.181.5%——92.8%——
Steps to solve (medium)127——53——
FrontierCode 1.1 Main scores by model Bar chart of FrontierCode 1.1 Main scores: SWE-1.7 42.0%, Grok 4.6 48.0%, GPT-5.6 Sol 47.5%, SWE-2 50.0%, Fable 5.1 50.9%, GPT-6 Astra 53.3%. 42.0% SWE-1.7 48.0% Grok 4.6 47.5% GPT-5.6 Sol 50.0% SWE-2 50.9% Fable 5.1 53.3% GPT-6 Astra genztech.blog
Fig 2 · benchmark SWE-2 lands between Fable 5.1 and GPT-5.6 Sol on Cognition's own FrontierCode 1.1 Main score, at a fraction of either model's per-task cost.

Why does the cost number matter more than the score?

Benchmark scores move a point or two every product cycle across the whole industry right now. What doesn't move as often is a model beating its predecessor on accuracy while cutting the number of steps it takes nearly in half. Devin bills by usage, not by a flat subscription, so every step an agent takes is a real, metered cost passed to the customer. A model that solves the same class of bug in 53 steps instead of 127 changes the unit economics of running an agent on a large codebase all day, which is the actual pitch Cognition is selling to engineering teams, not the leaderboard placement.

RelatedKimi K3 Is Third Best at Coding. It Is Also Not Open.

Who does this actually affect?

Teams already paying for Devin get a quiet upgrade with no plan change needed. Teams evaluating coding agents now have one more entrant to test against Cursor, GitHub Copilot's agent mode, Windsurf, and the frontier labs' own coding modes, and SWE-2's pitch is specifically to the buyer who wants near-frontier output without paying frontier per-token rates. It also puts pressure on rivals: if a mid-tier lab can close most of the gap to GPT-6 Astra through training method alone, that's a cheaper path to competitiveness than the compute-scaling race everyone else is running.

What's the skeptical read here?

Every number in this piece comes from Cognition, measured on a benchmark Cognition built, run on Cognition's own harness. That's not disqualifying, but it means SWE-2 hasn't cleared the bar that actually matters for buyers: an independent, third-party evaluation on a neutral harness. Our own AI coding leaderboard tracks exactly this gap, vendor-claimed scores routinely run several points above what independent evaluators like vals.ai measure once a model leaves the vendor's own test environment. SWE-2 goes into that leaderboard's verification queue today, unranked, until an independent run confirms or corrects these numbers. There's also no published token pricing or context window size in the announcement, both of which matter more to a buyer doing math than a percentage on a proprietary benchmark.

What it means for the market

Cognition isn't public, but it's a useful signal for where its valuation conversation stands. We covered on August 13 that Cognition was in talks putting the company near a $40 billion valuation on the strength of Devin's revenue multiple. A model that Cognition itself says cuts serving cost by 81% against its predecessor, if that holds up under independent scrutiny, directly supports the margin story behind that number: cheaper inference per solved task means better unit economics on every Devin seat sold. Watch whether Cognition discloses actual token pricing in the coming days, and whether any funding update references SWE-2 by name.

  1. Oct 29, 2025SWE-1.5 launchesfirst model behind Devin's public coding-agent push
  2. Apr 2026SWE-1.6 shipsfocused on agent UX rather than raw score gains
  3. Jul 8, 2026SWE-1.7 launches"near-frontier intelligence at a fraction of the cost"
  4. Sep 10, 2026SWE-2 launchespost-trained from Kimi K3, three effort tiers, live today

Our take

The training method here is the actual news, not the leaderboard position. A linear cost-penalty term that provably behaves the same before and after batch averaging is the kind of detail that either means Cognition's research team found something real, or means they're very good at making an incremental RL tweak sound rigorous. Given SWE-1.7 to SWE-2 is a four-model iteration in under a year with real efficiency gains each time, we lean toward the former. The number to watch isn't 50.0%, it's whether vals.ai or another independent evaluator reproduces anything close to it once SWE-2 is running outside Cognition's own harness. Until then, treat this as a strong internal result from a company that has been directionally honest about its numbers before, not a verified frontier-model result.

What to watch · next few weeks
  • Independent verification. Whether vals.ai or another neutral evaluator posts a SWE-bench Verified score for SWE-2 that holds within a few points of Cognition's internal numbers.
  • Token pricing. Cognition hasn't published per-token cost for SWE-2; the 81% efficiency claim is only checkable once that number exists.
  • Rival responses. Whether xAI, OpenAI, or the Fable team adjust pricing or ship an update that specifically targets SWE-2's cost-per-task pitch.
  • The valuation talks. Whether Cognition's reported $40B valuation conversation references SWE-2's efficiency numbers as supporting evidence.
Primary sources

Original analysis by GenZTech, based on Cognition's official announcement. Source.