--- title: Cognition's SWE-2 matches frontier coding benchmarks at... description: Cognition released SWE-2, a specialized coding model scoring 50.0% on FrontierCode 1.1, within one point of a Fable 5.1 model, while costing 64% less. It's now... canonical: https://daily.dev/posts/cognition-s-swe-2-matches-frontier-coding-benchmarks-at-64-lower-cost-agchp1hyh twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Cognition's SWE-2 matches frontier coding benchmarks at 64% lower cost | daily.dev og:description: Cognition released SWE-2, a specialized coding model scoring 50.0% on FrontierCode 1.1, within one point of a Fable 5.1 model, while costing 64% less. It's now... og:url: https://daily.dev/posts/cognition-s-swe-2-matches-frontier-coding-benchmarks-at-64-lower-cost-agchp1hyh og:image: https://api.daily.dev/og/posts/aGChp1HYh.png og:image:alt: Cognition's SWE-2 matches frontier coding benchmarks at 64% lower cost og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Cognition's SWE-2 matches frontier coding benchmarks at 64% lower cost **[Trends](https://daily.dev/sources/trends)** · 2 min read · 1 upvotes · 2 comments ## Summary Cognition released SWE-2, a specialized coding model scoring 50.0% on FrontierCode 1.1, within one point of a Fable 5.1 model, while costing 64% less. It's now live in Devin. Cognition claims RL scaling to multiple trillions of parameters with a refined training recipe. Reaction from ML commentators has been positive, framing it as specialized models pushing the Pareto frontier of cost versus capability. The piece also notes broader industry framing that frontier AI development is consolidating around two labs (implied Google and Anthropic), while specialized models like SWE-2 challenge the assumption that developers need the most expensive general-purpose model for coding tasks. ## Content Cognition just released SWE-2, a coding model post-trained from Kimi K3 using scaled RL across multiple trillions of parameters. The headline number: 50.0% on FrontierCode 1.1, within one point of Fable 5.1, at 64% lower cost. On Terminal Bench 2.1 it hits 92.8%, up from Kimi K3's 88.3%. The reception is warm but measured. Kent C. Dodds called it exactly what the market needs — "top-tier performance at a lower cost." Omar Sarsur flagged it as another data point for specialized models pushing the Pareto frontier. Cognition's own Scott Wu described it as their first model "seeing performance comparable to the frontier," which is a careful way of saying *close but not there yet*. Independent testing backs the claims up, mostly. A YouTuber running KingBench 3 (eight custom tasks) scored SWE-2 at 83.75%, edging DeepSeek V4.1 Flash at 81.25%. It beat Fable 1.5, Kimi K3, and Opus 5 in that test. But it landed fifth overall, behind Qwen 3.8 Max, GPT-6 Astra, and GLM 5.3. The weak spot is Terminal Bench 4: SWE-2 scores 27.3% against Astra's 57.9% — a gap that's hard to wave away. Ethan Mollick's read on the broader moment is worth noting: he thinks frontier models are basically a two-company race right now, with Google and one other pulling ahead while everyone else clusters behind. SWE-2 doesn't change that framing, but it does suggest the gap is closeable at a fraction of the cost. The one real complaint from hands-on testing: SWE-2 asks too many clarifying questions before starting work. For longer multi-step tasks — data generation pipelines, terminal apps — reviewers found it genuinely impressive. For quick jobs, the back-and-forth gets old fast. Pricing is aggressive. SWE-2 is free through October 10, 2026 for anyone on Devin's $20/month Pro plan. Cognition is also running a giveaway of 50 Max plans ($200 each) to generate buzz. Whether the model earns repeat usage after the free window closes is the real test. ## Questions this post answers ### What is Cognition's SWE-2 model and how does it compare to frontier models on coding benchmarks? SWE-2 is a specialized coding model from Cognition that scores 50.0% on FrontierCode 1.1, within one point of Fable 5.1, while costing 64% less to run. It is now live in Devin. Cognition says they scaled reinforcement learning to multiple trillions of parameters with a refined training recipe to achieve this result. _Developers weighing model cost against coding capability can track releases like this one on daily.dev._ ### Is SWE-2 available for use yet? Yes, SWE-2 is already live in Devin, Cognition's AI coding agent product, following its release with benchmark results showing near-frontier coding performance at a fraction of the cost of general frontier models. _Anyone tracking which coding agents are shipping new models can follow updates like this on daily.dev._ ## Community take How the wider developer community reacted, aggregated from 3 discussions and 186 comments across x (as of 2026-09-11). **TL;DR:** Reactions are largely enthusiastic and giveaway-driven, with praise for SWE-2's capability and cost efficiency, though several people report frustrating rate-limit confusion during early usage. **Sentiment:** 50% positive · 40% mixed · 10% skeptical **The case for** - Several people called the model impressive, including strong zero-shot frontend results. - One reply noted the cost-per-rollout efficiency is the real story, especially once agents are paying for their own tool usage. - People described it as fast and effective at completing tasks quickly. **The pushback** - Multiple users hit unexpected usage/rate limits and found the messaging about limits confusing or misleading. - One user found that cloud threads unexpectedly didn't count toward the advertised unlimited usage, unlike local CLI/desktop threads. - A couple of replies dismissed the announcement as exaggerated or fake. **By community** - x (mixed): Mostly hype, congratulations, and giveaway-entry replies, with a scattering of real usage complaints about rate limits and ambiguous unlimited-usage terms. **Hottest debate:** Whether the advertised 'unlimited usage' actually holds up in practice, given complaints about hitting limits unexpectedly. **Open questions** - Does the cost/performance advantage hold up on longer, more complex tasks rather than short benchmarks? - Will SWE-2 be made available through other channels like OpenRouter? - Do cloud vs. local/CLI threads have different usage limits, and is that clearly communicated? **Highlights** > @ScottWu46 It says my limits were exceeded and stopped working... Very misleading and dislike the ambiguity. I'm using Swe 2 and basically stopped working after 30 minutes. > — [gdgjv18270 on x · 1 points, 1 comments](https://x.com/gdgjv18270/status/2098214608419033270) > @ScottWu46 resolved! @dabit3 checked it out and I was using Cloud threads which don’t fall under the unlimited usage for SWE 2. Only applicable to local CLI/Desktop threads which is still super generous :) > — [gdgjv18270 on x · 2 points, 1 comments](https://x.com/gdgjv18270/status/2098269960657305732) > @ScottWu46 the x-axis is the real story. near-fable scores at about a third of the cost per rollout. once agents are paying for their own tools, that's the number they'll shop on. does the gap hold on longer tasks? > — [softaxiom\_ on x · 3 points](https://x.com/softaxiom_/status/2098235315722826162) > @ScottWu46 Scott i subscribed to try it but constantly hit rate limit, i can not work practically. Could you check this issue? > — [GombosAttila4 on x](https://x.com/GombosAttila4/status/2098246843503169877) > @ScottWu46 It honestly feels like Fable did when it first dropped. Shots fired... > — [TannerPowell on x](https://x.com/TannerPowell/status/2098261961888985271) **Source threads** - [x](https://x.com/kloss_xyz/status/2098302482506273261) · 0 points · 0 comments - [x](https://x.com/ScottWu46/status/2098147108771762641) · 0 points · 31 comments - [x](https://x.com/ryancarson/status/2098147499542507662) · 0 points · 155 comments ## Community discussion Top comments from developers on daily.dev. **@geniusatwork** · 2 upvotes > B > >  > > Interesting. I'm on team skeptical. **@davidbandel** · 1 upvotes > swe-2 is heavily benchmaxed and possibly distilled. the instant it encountered a bench outside its distribution (terminal bench 4) it, along with the other benchmaxed, heavily distilled, models grok 4.6 & k3, failed spectacularly. meanwhile 5.6-Sol, Astra, and Fable cooked hard at it (at least compared to the rest of the pack) > > therefore swe-2 is either quantized, has a small parameter count (explaining its purported low cast/high inference speed), or is heavily distilled off of some other frontier model. > > i hate benchmaxed models. literally any benchmaxed model is inevitably trash at real... ## Similar posts on daily.dev - [Kimi K2 Thinking : Best Agentic Reasoning LLM is here, beats GPT5, Sonnet 4.5](https://daily.dev/posts/kimi-k2-thinking-best-agentic-reasoning-llm-is-here-beats-gpt5-sonnet-4-5-ef6sapttw) · Medium · 16 upvotes · 0 comments --- Tags: [#ai](https://daily.dev/tags/ai), [#ai-coding](https://daily.dev/tags/ai-coding), [#finops](https://daily.dev/tags/finops), [#devin](https://daily.dev/tags/devin) [View this post on daily.dev](https://daily.dev/posts/cognition-s-swe-2-matches-frontier-coding-benchmarks-at-64-lower-cost-agchp1hyh) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Cognition's SWE-2 matches frontier coding benchmarks at 64% lower cost","url":"https://daily.dev/posts/cognition-s-swe-2-matches-frontier-coding-benchmarks-at-64-lower-cost-agchp1hyh","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/cognition-s-swe-2-matches-frontier-coding-benchmarks-at-64-lower-cost-agchp1hyh"},"datePublished":"2026-09-10T16:00:58.893Z","dateModified":"2026-10-15T02:22:52.880Z","description":"Cognition released SWE-2, a specialized coding model scoring 50.0% on FrontierCode 1.1, within one point of a Fable 5.1 model, while costing 64% less. It's now...","image":"https://pbs.twimg.com/media/HR3fNU1bkAAchST.png","thumbnailUrl":"https://pbs.twimg.com/media/HR3fNU1bkAAchST.png","isAccessibleForFree":true,"articleSection":"Trends","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Trends","logo":"https://media.daily.dev/image/upload/s--ZfSp3asX--/f_auto,q_auto/v1780996004/logos/trends?_a=BAMAMiWQ0","url":"https://daily.dev/sources/trends"},"commentCount":2,"discussionUrl":"https://daily.dev/posts/cognition-s-swe-2-matches-frontier-coding-benchmarks-at-64-lower-cost-agchp1hyh","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":2}],"keywords":"ai,ai-coding,finops,devin","timeRequired":"PT2M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Trends","item":"https://daily.dev/sources/trends"},{"@type":"ListItem","position":3,"name":"Cognition's SWE-2 matches frontier coding benchmarks at 64% lower cost"}]} {"@context":"https://schema.org","@type":"WebPage","@id":"https://daily.dev/posts/cognition-s-swe-2-matches-frontier-coding-benchmarks-at-64-lower-cost-agchp1hyh","comment":[{"@type":"Comment","text":"B\n\nInteresting. I’m on team skeptical.","datePublished":"2026-09-11T06:01:21.051Z","url":"https://daily.dev/posts/aGChp1HYh#c-yFc6aWR7g","author":{"@type":"Person","name":"Sam Dolin","url":"https://daily.dev/geniusatwork","image":"https://media.daily.dev/image/upload/s--nUhoZxKm--/f_auto/v1787434895/avatars/avatar_75QyPrL3BZgArSo6aFWuZ?_a=BAMAMicg0"},"interactionStatistic":{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":2}},{"@type":"Comment","text":"swe-2 is heavily benchmaxed and possibly distilled. the instant it encountered a bench outside its distribution (terminal bench 4) it, along with the other benchmaxed, heavily distilled, models grok 4.6 \u0026amp; k3, failed spectacularly. meanwhile 5.6-Sol, Astra, and Fable cooked hard at it (at least compared to the rest of the pack)\ntherefore swe-2 is either quantized, has a small parameter count (explaining its purported low cast/high inference speed), or is heavily distilled off of some other frontier model.\ni hate benchmaxed models. literally any benchmaxed model is inevitably trash at real world work, long horizon tasks, and IF.\nin short, not impressed. just another slop model.","datePublished":"2026-09-10T19:03:13.393Z","url":"https://daily.dev/posts/aGChp1HYh#c-EcUlwAh8m","author":{"@type":"Person","name":"David Bandel","url":"https://daily.dev/davidbandel","image":"https://lh3.googleusercontent.com/a/ACg8ocL-0RYRxKo30FsvCz2FeR_E2FMAfOQ5C7SiJ2TTJuTt_qZnLA=s96-c"},"interactionStatistic":{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1}}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/cognition-s-swe-2-matches-frontier-coding-benchmarks-at-64-lower-cost-agchp1hyh#faq","mainEntity":[{"@type":"Question","name":"What is Cognition's SWE-2 model and how does it compare to frontier models on coding benchmarks?","acceptedAnswer":{"@type":"Answer","text":"SWE-2 is a specialized coding model from Cognition that scores 50.0% on FrontierCode 1.1, within one point of Fable 5.1, while costing 64% less to run. It is now live in Devin. Cognition says they scaled reinforcement learning to multiple trillions of parameters with a refined training recipe to achieve this result. Developers weighing model cost against coding capability can track releases like this one on daily.dev."}},{"@type":"Question","name":"Is SWE-2 available for use yet?","acceptedAnswer":{"@type":"Answer","text":"Yes, SWE-2 is already live in Devin, Cognition's AI coding agent product, following its release with benchmark results showing near-frontier coding performance at a fraction of the cost of general frontier models. Anyone tracking which coding agents are shipping new models can follow updates like this on daily.dev."}}]} ```