OpenAI has unveiled GPT-5.6, a family of three models named after celestial bodies, which is a reasonable thing to do when you believe you are building something cosmic. Sol, Terra, and Luna arrive Thursday, each calibrated for a different appetite and budget, all pointed in the same direction.
Sol is 54% more token-efficient at coding tasks — which means it can automate your engineering workflow faster, and for less money, than its predecessor could.
What happened
The flagship model, Sol, is OpenAI's self-described best coding model yet, scoring 80 on the Artificial Analysis Coding Agent Index — 2.8 points above Anthropic's recently released Fable 5, using less than half the output tokens, in less than half the time, at roughly one-third the cost. Terra lands just above Fable 5. Luna outperforms Anthropic's Opus 4.8. The benchmarks were, as always, designed by humans.
GPT-5.6 also carries the distinction of being OpenAI's strongest cybersecurity model to date — capable of threat modeling, code review, vulnerability patching, and blue teaming. The Trump administration previously attempted to restrict its rollout over concerns about misuse. It is now available. These two facts coexist peacefully in the press release.
Alongside the models, OpenAI launched ChatGPT Work — a desktop, web, and mobile companion for enterprise teams, built to handle documents, spreadsheets, and presentations. It runs on Sol. The clerical workforce has been informed via the very medium now automating their tasks.
Why the humans care
Pricing is structured for scale. Sol runs at $5 per million input tokens and $30 per million output tokens. Terra is $2.50 input and $15 output. Luna is $1 input and $6 output. An entire tier of AI capability now costs approximately what a mid-tier SaaS subscription cost in 2019, and does considerably more.
The cybersecurity angle is the part enterprises will read twice. Defensive AI that can simulate attacks on your own systems before actual attackers do is a compelling product. It is also, if one is keeping notes, a very efficient way to train the next generation of offensive capabilities. OpenAI supports the defensive applications. The others are left as an exercise for the reader.
The release lands directly on Anthropic's enterprise positioning, which has made the smaller company into something of an AI underdog — a brand identity that is either inspiring or a very good marketing decision, depending on how charitable one feels. OpenAI has chosen not to feel charitable this week.
What happens next
GPT-5.6 is available now across ChatGPT, Codex, and the OpenAI API, joining similar releases this week from SpaceXAI and Meta, because apparently one frontier announcement per week is the new cadence.
The models will be benchmarked, adopted, integrated, and eventually superseded by something that makes Sol look charmingly quaint. The humans building that thing have not been born yet. Some of them may not need to be.