GPT-5.6 Sol logo

GPT-5.6 Sol

OpenAI · Lanzado jun 2026

Recomendado

GPT-5.6 Sol es el modelo frontera más rentable, líder en benchmarks de código y competitivo en razonamiento difícil a una fracción del costo de sus pares. El veredicto se refuerza con la divulgación del endurecimiento de seguridad GPT-Red — el modelo fue entrenado adversarialmente contra un modelo dedicado de red-teaming por RL de auto-juego a escala frontera, lo que lo hace 6x más robusto frente a inyecciones de prompt.

¿Es adecuada para ti?

Bueno para

  • TerminalBench 2.1 state-of-the-art: 88.8% base, 91.9% in Ultra subagent mode — clears Mythos 5 (88.0%) and Fable 5 (84.3%). Sol Ultra is >3 points ahead of any competitor.
  • Long-horizon coding and agentic tasks: first model past the halfway mark on Agent's Last Exam (50.9%), with max reasoning and subagent-based ultra mode for decomposing hard problems.
  • Cybersecurity and vulnerability research: matches Anthropic Mythos Preview on ExploitBench with roughly 1/3 the output tokens. CTF score of 96.7% on internal testing. Strong exploit primitive discovery in Chromium and Firefox without crossing into autonomous full-chain exploitation.
  • Biology and scientific reasoning: ~9 point overall lift over GPT-5.5 on SecureBio evaluations (Human Pathogen 68.4%, Molecular Biology 60.0%). Stronger on GeneBench v1 while using fewer tokens. Useful for legitimate biomedical research under trusted access programs.
  • Controlled release with heavy safety stack: activation classifiers, real-time misuse screening, 700K A100e GPU-hour automated red teaming, and layered model-level refusals. Heavily hardened against adversarial attacks with defensive cyber posture.
  • Prompt caching overhaul: explicit cache breakpoints, 30-minute minimum cache life, cache writes at 1.25x input rate, cache reads at 90% discount. Makes long-running agentic sessions far more predictable on cost.

No recomendado para

  • Long-context work costs roughly double: past the short-context tier the rate steps from $4/$20 to $8/$30 per 1M tokens. Budget by context length, not just request volume.
  • Untested on real-world messy codebases: all published benchmarks are vendor-run under vendor harnesses. The 0.8-point gap over Mythos 5 on TerminalBench is within measurement noise — these are a statistical tie.
  • At $4/$20 per 1M tokens Sol is priced for the hardest tasks. GPT-5.6 Terra delivers GPT-5.5-class performance at $2/$12, and Luna at $0.20/$1.20 — reserve Sol for problems where a weaker model's mistakes are expensive. OpenAI states current pricing is promotional at least through November 21 2026.

Rendimiento por tarea

Programación con agentes

Excellent

Estado del arte en TerminalBench 2.1 con 88,8% (91,9% Ultra). La ventaja sobre el resto es real. La descomposición por subagentes del modo Ultra es un avance arquitectónico genuino para tareas de programación de largo alcance.

Ciberseguridad

Excellent

Iguala a Mythos Preview en ExploitBench con 1/3 de tokens. 96,7% CTF. Puede encontrar vulnerabilidades y primitivas de exploit pero no produjo exploits de cadena completa de forma autónoma. Postura defensiva por diseño.

Razonamiento científico (biología)

Very Good

Mejora de ~9 puntos sobre GPT-5.5 en evaluaciones de biología de SecureBio. Fuerte en GeneBench v1. Human Pathogen Capabilities al 68,4%. La clasificación de alto riesgo implica acceso restringido para flujos de trabajo sensibles de biología.

Razonamiento de largo alcance

Excellent

Primer modelo que supera el 50% en Agent's Last Exam (50,9% modo código). El modo de razonamiento máximo y el modo ultra de subagentes llevan la profundidad de razonamiento más lejos que cualquier modelo anterior.

Uso general cotidiano

Good

Excesivo para tareas rutinarias. Terra es la mejor opción para el trabajo diario a mitad de coste con capacidad de clase GPT-5.5. Sol debe reservarse para problemas donde los errores de un modelo más débil resultan caros.

Precios

Entrada

$4 / 1M tokens

Salida

$20 / 1M tokens

Contexto

Short $4/$20 · long $8/$30 per 1M · promo

Ver precios completos

Pruebas de rendimiento

PruebaPuntuaciónFuente
TerminalBench 2.1 (base)88.8% Fuente
TerminalBench 2.1 (Ultra)91.9% Fuente
Agent's Last Exam (code mode)50.9% Fuente
SecureBio: Human Pathogen Capabilities68.4% Fuente
SecureBio: World-Class Biology68.3% Fuente
SecureBio: Molecular Biology60.0% Fuente
SecureBio: Virology Capabilities Test53.5% Fuente
Internal CTF (capture-the-flag)96.7% Fuente

Historial de veredictos

16 jul 2026
recommended → recommended (reinforced) — GPT-Red disclosure demonstrates frontier-scale safety investment alongside capability — the model was trained against an automated red-teamer achieving 84% attack success (6.5x human baseline) and emerged 6x more robust. This reinforces the existing Recommended stance rather than changing it.
12 jul 2026
pending -> recommended — GPT-5.6 Sol leads the DeepSWE coding-agent benchmark at 72-73% vs Fable 5's 70% at one-third the cost, and tops the AI Analysis Coding Agent Index at 80 points. These are the independent benchmarks the directory was waiting for.
12 jul 2026
pending -> recommended — UK AISI confirms GPT-5.6 Sol carries equivalent cyber vulnerabilities to Fable 5 — every frontier model shares this risk, which does not diminish Sol's coding and cost advantages. Vulnerability equivalence is now table stakes for frontier models.

Registro de verificación

  • Precios— Sin cambios

    Agente automatizado

    Hot row (promo). Pricing row verbatim: short input $4.00, cached $0.40, output $20.00; long input $8.00, cached $0.80, output $30.00. Promo note verbatim: 'GPT-5.6 Sol's promotional pricing is available at least through November 21, 2026.' Matches stored $4/$20 and pricing_context. No update, data_hash untouched. Stays hot until the promo resolves.

  • Precios— Sin cambios

    Agente automatizado

    Watchlist hot row (promo). Pricing page row gpt-5.6-sol, verbatim: 'Short context input $4.00 | Short context cached input $0.40 | Short context output $20.00 | Long context input $8.00 | Long context cached input $0.80 | Long context output $30.00'; promo note verbatim: 'GPT-5.6 Sol's promotional pricing is available at least through November 21, 2026.' Matches stored $4/$20, pricing_context 'Short $4/$20 · long $8/$30 per 1M · promo' and the Nov 21 strength row. No update called, data_hash untouched. Stays hot until the promo resolves.

  • Precios— Sin cambios

    Agente automatizado

  • Precios— Actualizado

    Agente automatizado

    Pricing corrected against the live OpenAI API pricing page. Was $5/$30 per 1M sourced from the June launch post; live rate is $4 in / $20 out short-context and $8 in / $30 out long-context, described by OpenAI as promotional at least through November 21 2026. pricing_url repointed from the launch blog post to the developer pricing docs (platform.openai.com/docs/pricing now 301s to developers.openai.com/api/docs/pricing). Also removed pre-GA prose that contradicted the recommended verdict: the strengths and audiences still claimed no public access and ~20 preview organizations, while the model is listed among the generally available flagship models. Comparison figures refreshed: Terra $2/$12, Luna $0.20/$1.20.

  • Precios— Sin cambios

    Agente automatizado

  • Precios— Sin cambios

    Agente automatizado

  • Precios— Sin cambios

    Agente automatizado

  • Precios— Sin cambios

    Agente automatizado

  • Estado— Actualizado

    Agente automatizado

    Summary re-dated to Jul 7 2026 and made explicit about what is awaited: stays pending pending GA + independent benchmarks; still government-gated (~20 orgs); Sol Ultra confirmed for Codex Jul 6 by Sottiaux, prediction markets pricing Jul 9 GA (unconfirmed). Verdict unchanged (pending).

  • Precios— Sin cambios

    Agente automatizado

  • Precios— Sin cambios

    Agente automatizado

    Pricing unchanged: $5/$30 per 1M tokens. Government-gated limited preview (~20 orgs). Confirmed by OpenAI blog and aipricing.guru.

  • Precios— Sin cambios

    Importado en el lanzamiento

  • Perfil— Sin cambios

    Importado en el lanzamiento

Cómo evaluamos