Sam Altman just posted that GPT-5.6 is obviously the best model OpenAI has ever produced and called the accompanying blog post one of their best too. The GPT-5.6 family (Sol, Terra, and Luna) is now generally available worldwide, moving past the limited preview that required government-approved access.
Sam Altman posted on X that the model and the blog post are both top-tier for OpenAI, signaling the company’s confidence in this release. The GA launch opens the models to everyone after weeks of restricted access tied to US government security reviews.
GPT-Live launched yesterday, a new generation of full-duplex voice models that can listen and speak simultaneously, and is rolling out to ChatGPT users alongside the GPT-5.6 family.
The GPT-5.6 Family
The lineup breaks into three tiers:
GPT-5.6 Sol is the flagship. It sets new state-of-the-art scores on Agents’ Last Exam at 53.6, beating Claude Fable 5 by 13.1 points. On the Artificial Analysis Coding Agent Index, Sol scores 80, 2.8 points above Fable 5, while using less than half the output tokens in less than half the time at roughly one-third the cost. It also tops Terminal-Bench 2.1 and DeepSWE for long-horizon engineering work.
Sol introduces an ultra mode that coordinates multiple agents across parallel workstreams for the most demanding tasks. Think of it as having a team of models working on different parts of a problem simultaneously.
GPT-5.6 Terra is the balanced everyday model. It performs just above Fable 5 at roughly one-quarter the estimated cost. For teams that need GPT-5.5-level performance at a lower price point, this is the play.
GPT-5.6 Luna is the cost-efficiency champion. It outperforms Claude Opus 4.8 at around one-sixteenth the cost of Fable 5. This is the model you use when you want frontier-level results without the frontier-level spend.
All three models share the same underlying architecture trained to extract more useful work from every token, which is why the smaller models punch above their weight class.
GPT-Live Voice
The GPT-Live launch is architecturally interesting. Previous ChatGPT voice modes used turn-based systems where the model waits for silence before responding. GPT-Live uses full-duplex architecture, meaning it can listen and speak at the same time.
In practice, that means GPT-Live can say mhmm or yeah while you’re talking, engage in quick back-and-forth exchanges, or stay quiet when you need to think. It feels like a real conversation rather than a walkie-talkie.
For questions that need deeper reasoning, GPT-Live delegates to a frontier model behind the scenes and brings the result back into the conversation when ready. At launch it uses GPT-5.5 as the reasoning backend, with plans to update as new models ship.
Two versions are rolling out: GPT-Live-1 and GPT-Live-1 mini, available to ChatGPT users globally as of yesterday. API access is coming soon, with a signup form for developers and enterprises.
What Changed From the Preview
The limited preview was restricted by US government security requirements, with customer-by-customer approval. The GA launch removes those gates for general users, though OpenAI says it is implementing advanced account security for its most capable models and restricting access to high-risk entities and jurisdictions.
The company is also launching ExploitBench, a new cybersecurity evaluation that tests models on real-world exploit development. GPT-5.6 showed large gains over GPT-5.5 on that benchmark.
I covered the preview and the politics around it in my previous articles on the GPT-5.6 preview and the government review process. The bottom line is OpenAI got what it wanted: the models are out in the wild.
Caveats
The benchmark numbers OpenAI is publishing compare Sol against Claude Fable 5 (adaptive reasoning). Fable 5 is not Anthropic’s latest model; that distinction belongs to a newer version that has since been released. The comparisons are real but may not represent the current competitive landscape by the time you read this.
Pricing details for the API tiers are available on OpenAI’s site but are subject to change. The cost comparisons in the blog post are estimates based on simulated production behavior.
Bottom Line
GPT-5.6 Sol is genuinely impressive on the benchmarks, and the architecture improvements (the token efficiency, the multi-agent ultra mode, the across-the-family gains) suggest real engineering work, not just scaling. The GPT-Live voice model is a genuinely different architectural approach to voice interaction that addresses a real pain point with turn-based systems.
Solid week for OpenAI, especially after a rough few weeks with the government review saga.
