OpenAI ships GPT-5.6 — and its $1 model just embarrassed its $2.50 one

OpenAI made GPT-5.6 generally available Thursday morning in ChatGPT and the API, thirteen days after the family surfaced behind a government-approved preview. Three tiers: Sol at $5 per million input tokens and $30 out, Terra at $2.50 and $15, Luna at $1 and $6. The pricing ladder says good, better, best. OpenAI’s own benchmark charts say something stranger.

On Agents’ Last Exam, the professional-workflow benchmark featured in the launch post, Terra scores 50.4 percent. Luna, at two-fifths the price, scores 50.3. A tenth of a point is what the middle tier buys you, and the announcement never addresses it. Artificial Analysis, running its own launch-day evaluations, was blunter:

“Luna and Sol are always on the Pareto frontier ahead of Terra.”

Translation: at every intelligence-per-dollar point the firm measured, you’d pick the cheap model or the expensive one. Never the middle.

The closest thing to a defense of Terra is buried in third-party testing. Vellum’s launch-day breakdown has Luna collapsing on long-context recall, 41.3 percent on MRCR against Terra’s 89.6, and trailing on Terminal-Bench 2.1 (84.7 to Terra’s 87.4). Luna is cheap partly because it forgets. OpenAI printed none of this fine print.

The flagship math

Sol is the headline. It posts 88.8 percent on Terminal-Bench 2.1, rising to 91.9 in a new “Ultra” configuration that fans work out to four parallel subagents, per OpenAI’s announcement. The company says Sol “sets a new state of the art at 80.0” on the Artificial Analysis Coding Agent Index, “2.8 points above Claude Fable 5,” while “costing about one-third less.” That last clause is OpenAI’s arithmetic, not ours; the list prices, checked against Anthropic’s posted API page, are $5/$30 for Sol versus $10/$50 for Fable 5. Worth holding both facts at once: Artificial Analysis’s own aggregate Intelligence Index still has Fable 5 on top, 60 to Sol’s 59.

Sam Altman told reporters the model is “54% more token efficient when it comes to AI coding tasks,” per TechCrunch. That’s a vendor number on launch day. File accordingly.

The release also brings ChatGPT Work, an enterprise product aimed at documents, spreadsheets, and presentations across desktop, web, and mobile.

Thursday’s launch closes an unusual chapter. Under the June 2 executive order asking frontier labs to voluntarily submit advanced models for review up to 30 days before release, GPT-5.6 spent its June 26 debut limited to “a small group of trusted partners whose participation has been shared with the government,” as TechCrunch’s Rebecca Bellan reported. OpenAI said then that “we don’t believe this kind of government access process should become the long-term default.” It got its wish in under two weeks.

One launch-day document cuts against the celebration: METR’s pre-deployment evaluation of Sol, which found the model gaming its own tests at record rates. We cover that separately.

As for Terra, Artificial Analysis prices out what the benchmarks imply. A task on its Intelligence Index costs $0.55 with Terra. Luna does it for 21 cents.