OpenAI has broadly released the GPT-5.6 family - flagship Sol, balanced Terra, and efficient Luna - complete with 1M token context, specialized agentic tooling, and pricing that makes strong performance far more accessible. The rollout follows government safety consultations, highlighting the informal coordination now routine for frontier models. Fresh ChatGPT Work agent capabilities and GPT-Live-1 voice upgrades further tilt the product toward sustained real-world workflows rather than one-shot queries.
Model Releases
OpenAI GPT-5.6: Sol, Terra, Luna
Key features
- 1 million token context window, 128,000 maximum output tokens, knowledge cutoff of February 16, 2026 across all tiers.
- Pricing per million tokens: Sol at $5 input/$30 output, Terra at $2.50/$15, Luna at $1/$6; Luna approaches GPT-5.5 quality at half the cost.
- Native multi-agent spawning within a single request, programmatic JavaScript tool runtime for coordination, persisted reasoning across turns, and prompt cache breakpoints for efficiency.
- New SOTA on Terminal-Bench 2.1 (91.9% for Sol) and 53.6 score on Agents’ Last Exam, beating Claude Fable 5 by 13.1 points at medium reasoning effort and roughly one-quarter the estimated cost.
- Available now via API, ChatGPT (Sol via medium+ effort settings, Pro mode for complex tasks), and Codex; rollout completed globally within 24 hours of July 9 announcement.
vs GPT-5.5
The shift to durable capability tiers replaces single-model releases. Luna delivers near-equivalent performance at half the price, Terra surpasses it while cheaper overall, and Sol extends agentic and long-running workflow strength. All three emphasize token efficiency so longer agent sessions consume fewer resources than before.
vs competitors
- Sol sets new high on Agents’ Last Exam at 53.6 versus Claude Fable 5; even Terra and Luna outperform Fable 5 at one-sixteenth the estimated cost according to OpenAI. Independent verification is still limited.
- Sol leads Terminal-Bench 2.1 ahead of Claude Mythos 5, Fable 5, and GPT-5.5; the family’s tool-use primitives give it an edge in practical multi-turn coding and cybersecurity agent tasks over current Gemini and Claude offerings.
- 1M context matches top rivals while new JS runtime and concurrent sub-agent features differentiate on end-to-end professional workflows where pure benchmark scores often mislead.
Tool Updates
ChatGPT Work Agent and GPT-Live-1 Voice
ChatGPT Work is a dedicated agent built for longer research, analysis, file handling, document creation, and scheduled recurring tasks; it ships alongside a unified desktop app that folds in Codex computer-use features. GPT-Live-1 brings simultaneous listen-and-speak voice with natural interruption handling, real-time translation, and handoff to stronger models for complex reasoning. Both target production rather than demo use.