Pular para o conteúdo
Menu
Mix Vale

SWE-2 from Cognition matches GPT-5.6 Sol at 64% lower cost

Beatriz
Por 6 min de leitura
Compartilhe
Cognition logo - Divulgação
Foto: Cognition logo - Divulgação

Cognition launched SWE-2, an engineering model that matches frontier systems on software benchmarks while decreasing operating costs by up to 70 percent. On FrontierCode 1.1 Main, SWE-2 recorded a score of 50.0%, trailing Fable 5.1 by less than one percentage point while operating at 64% lower expense, alongside reported performance matching GPT-5.6 Sol and coming within range of GPT-6 Astra at roughly one-fourth of the price. Testing also placed the system ahead of Grok 4.6 across identical coding assessments.

Publicidade

SWE-2 represents the first release from Cognition featuring adjustable reasoning levels, configured through an optimization curve generated entirely within a solitary reinforcement learning run. Developers structured this process using linear cost penalties rather than building separate specialized checkpoints for each tier.

Engineers built SWE-2 on top of Kimi K3, a 2.8-trillion-parameter base system that received reinforcement learning, with the added training routine contributing 5 to 6 points across multiple industry benchmarks.

Publicidade
  • FrontierCode 1.1 Main: SWE-2 reached 50.0%, Kimi K3 hit 44.2%, Grok 4.6 scored 48.0%, Fable 5.1 reached 50.9%, GPT-5.6 Sol marked 47.5%, GPT-6 Astra posted 53.3%, and SWE-1.7 finished at 42.0%.
  • DeepSWE 1.1: SWE-2 attained 73.0%, compared to 68.5% for Kimi K3, 67.5% for Grok 4.6, 67.4% for Fable 5.1, 72.7% for GPT-5.6 Sol, 74.1% for GPT-6 Astra, and 37.7% for SWE-1.7.
  • Terminal-Bench 2.1: SWE-2 posted 92.8%, surpassing Kimi K3 at 88.3%, Grok 4.6 at 88.4%, Fable 5.1 at 91.4%, GPT-5.6 Sol at 88.8%, GPT-6 Astra at 89.9%, and SWE-1.7 at 81.5%.
  • Terminal-Bench 4: SWE-2 registered 27.3%, ahead of Kimi K3 at 21.5%, Grok 4.6 at 20.3%, and SWE-1.7 at 7.6%, but behind Fable 5.1 at 55.8%, GPT-5.6 Sol at 37.3%, and GPT-6 Astra at 57.9%.

Performance dipped on Terminal-Bench 4, where the model trailed Fable 5.1 and GPT-6 Astra on intricate terminal workflows. The overall metrics place SWE-2 at the top of price-to-performance rankings while yielding peak capability on the most demanding tasks.

How the system reduces redundant reasoning steps

Developers eliminated extraneous computing loops to prevent the model from over-analyzing routine code segments. Under medium settings on FrontierCode 1.1 Main, SWE-2 exceeded SWE-1.7 benchmark scores while requiring 58% fewer turns and lowering average expenses by 81%. Cognition credited the change to focused exploration, which isolates relevant code sections faster during execution.

Metric evaluations revealed that SWE-2 medium initiates code modifications after a median of 18 steps, down from 48 steps in SWE-1.7. Internal diagnostics documented distinct operational behaviors across three core development areas.

  • Automated testing: SWE-2 constructs end-to-end verification tests that track regressions and unusual failure modes across an entire implementation.
  • Problem solving under system boundaries: The model navigates blocked software paths by seeking alternative routes, including rebuilding dataset information from Slack history when MCP connections failed.
  • Independent verification: Rather than accepting user assertions, SWE-2 recalculates technical findings, tests user proposals against data, and generates test artifacts to validate claims.

Unified optimization across multiple computation tiers

The training procedure marks a technical shift from conventional development practices. Many labs generate independent models for distinct reasoning tiers before merging weights together. Kimi K3 trained individual specialists for each workload combination prior to distillation. Cognition took a different route.

The pipeline relies on a unified mathematical reward function where success signals interact directly with rollout execution costs and level-specific penalty coefficients. Technicians calibrated the penalty parameters to match the slope of the Pareto frontier belonging to the foundation system.

Publicidade

Tangency between the reward line and the efficiency curve ensures that training gains expand overall system capability instead of shifting trade-offs along existing baselines. Excessive penalties on intensive reasoning tiers risk causing advanced tiers to regress into medium behavior to trim expenses. Cognition demonstrated mathematically that a linear formulation remains the only penalty structure relying exclusively on mean expense and overall resolution rates. The resulting architecture prevents performance degradation across varying operational profiles.

A single training run elevated performance across all tiers, enabling medium settings to reduce expenditures while directing maximum-effort configurations to allocate additional tokens toward solving complex programming challenges.

Technical adjustments in reward stabilization and serving

Cognition incorporated a length-weighted baseline deployed internally since SWE-1.6 to minimize variance without incurring the computational cost of secondary backward passes. Because squared gradient measurements correlate with generation length, researchers approximated baseline rewards through token-length weighting. This method limits divergence between training and inference phases as network sizes grow.

Deployment architecture integrated an online draft engine created via SpecForge to maintain synchronization with active policies, expanding acceptance lengths by approximately 15 percent. By adopting NVFP4 and FP8 computing kernels alongside quantization training, Cognition preserved processing speeds comparable to SWE-1.7 despite deploying a base parameter footprint nearly three times larger.

The model rolled out through Devin Desktop and Devin CLI, with impending expansions planned for Devin Web and Fusion environments. Cognition granted complimentary access for one month to Pro, Max, and Teams subscribers to encourage platform adoption during user evaluation cycles. The move targets market competition against Claude Code and Codex.

The software engineering system provides three distinct processing modes tailored to differing project complexities:

  • Medium: Prioritizes rapid turnaround on uncomplicated programming assignments, yielding the documented 81% reduction in operating costs compared to SWE-1.7.
  • High: Allocates greater resources toward codebase analysis, architectural planning, and resolving incomplete project specifications.
  • Max: Uses extensive token counts and sequential turns to resolve deep algorithmic problems requiring multi-stage verification steps.

Implications for competitive software engineering models

Industry metrics reflect a closing performance divide between proprietary models and open-weight architectures enhanced through reinforcement learning pipelines. SWE-2 applied post-training techniques to Kimi K3 and achieved results comparable to GPT-6 Astra across complex coding benchmarks. This development highlights the growing influence of specialized refinement compared to base parameter scale alone.

Software laboratories face economic incentives to adopt single-run training regimens capable of optimizing full efficiency curves rather than maintaining discrete specialist models. The mathematical framework from Cognition provides an empirical foundation for deploying cost-tiered programming systems across enterprise development workflows.

Publicidade

Beatriz Lima é colaboradora da redação do Mix Vale, portal de notícias com cobertura de Brasil e mundo. Produz conteúdo nas editorias de economia e benefícios, com foco em informação de serviço para o leitor brasileiro. Escreve para o portal desde 2024.

Notícias relacionadas

Mais lidas

  1. Ciência

    Meteoro no Nordeste: bólido cruza o céu de cinco estados às 21h34 e mobiliza buscas por fragmentos na Paraíba

  2. Caixa Econômica Federal

    Pé-de-Meia paga parcela de R$ 1.000,00 a aprovados no ensino médio de 2026 entre fevereiro e março de 2027

  3. Política

    Com obstrução parcial no intestino, Jair Bolsonaro é internado no hospital DF Star em Brasília

  4. INSS

    Recuperação da senha Gov.br pelo celular exige reconhecimento facial ou código por SMS para o Meu INSS

  5. EUA

    Terremoto em Los Angeles atinge magnitude 4,2 e causa tremores leves em Commerce na madrugada

  6. Matéria Publicitária

    MineMV é o novo jogo grátis de construir com blocos; veja como jogar e liberado para todas as idades

  7. Política

    Regulamentação da escala 5×2 pelo governo encarece cálculo da hora extra a partir de 16 de dezembro

  8. Santa Catarina

    Acidente de trânsito na SC-108 envolve carro e moto e deixa motociclista morto em São Ludgero

  9. Mundo

    Jornalista italiano Corrado Augias morre aos 91 anos; presidente da Itália lamenta a perda

  10. Loterias

    Rendimento de R$ 40 milhões da Mega-Sena 3088 chega a R$ 337 mil por mês no Tesouro Direto