OpenAI ships GPT-5.6 Sol, Terra, Luna — one-third of Fable 5's cost, but SWE-Bench Pro is still 15 points behind
TL;DR
OpenAI ships GPT-5.6 Sol/Terra/Luna on July 9 — Sol at $5/$30 per M tokens, Terminal-Bench 91.9% beats Fable 5, but SWE-Bench Pro still trails by 15 points.
OpenAI pushed GPT-5.6 Sol, Terra, and Luna into the API and Codex on July 9, closing a 12-day White House review gate. Flagship Sol is priced at $5 / $30 per million tokens (input / output), Terra at $2.50 / $15, Luna at $1 / $6. Cache reads get a 90% discount. All three share a 1M token context with 128K max output and a knowledge cutoff of February 16, 2026.
Sol hits 88.8% on Terminal-Bench 2.1; the new Ultra mode (multi-subagent parallelism) pushes it to 91.9%, ahead of Anthropic's Fable 5 (88.0%) and OpenAI's own GPT-5.5 (88.0%). On Agents' Last Exam across 55 domains, Sol scores 53.6 vs Fable 5's 40.5. Per-task cost: Sol max $1.04, Fable 5 max $2.75 — same tier of output, one-third of the price.
But on SWE-Bench Pro, Fable 5 scores 80% and Sol only 64.6% — a 15.4-point gap. OpenAI's launch page doesn't feature that number. Sol's three headline verticals are biology, chemistry, and cybersecurity — exactly the high-risk domains named in the Biden-era AI cybersecurity executive order.
The process is new too. Sol first went out June 26 as a limited preview to ~20 "government-approved trusted partners", gated under voluntary AI safety framework cybersecurity clauses. This is the first frontier-tier flagship to ship on a "12-day White House gate, then public release" path. OpenAI calls it "our most robust safety stack ever," targeted at repeated misuse and sensitive cyber requests.
The API also gains Programmatic Tool Calling in JavaScript, explicit prompt cache breakpoints, original-resolution image input, and the multi-agent ultra effort tier. The single number Sam Altman highlighted: Sol emits 54% fewer output tokens than comparable models on coding tasks — at $30 / M output tokens, that lands directly on customer invoices.
Win the bet, and Sol takes half the API market on "half of Fable 5's price plus a Terminal-Bench beat." Lose it, and that 15-point SWE-Bench Pro gap becomes Anthropic's first marketing slide within a week of real workload runs.
via OpenAI / Simon Willison / The Decoder / Axios
Sol hits 88.8% on Terminal-Bench 2.1; the new Ultra mode (multi-subagent parallelism) pushes it to 91.9%, ahead of Anthropic's Fable 5 (88.0%) and OpenAI's own GPT-5.5 (88.0%). On Agents' Last Exam across 55 domains, Sol scores 53.6 vs Fable 5's 40.5. Per-task cost: Sol max $1.04, Fable 5 max $2.75 — same tier of output, one-third of the price.
But on SWE-Bench Pro, Fable 5 scores 80% and Sol only 64.6% — a 15.4-point gap. OpenAI's launch page doesn't feature that number. Sol's three headline verticals are biology, chemistry, and cybersecurity — exactly the high-risk domains named in the Biden-era AI cybersecurity executive order.
The process is new too. Sol first went out June 26 as a limited preview to ~20 "government-approved trusted partners", gated under voluntary AI safety framework cybersecurity clauses. This is the first frontier-tier flagship to ship on a "12-day White House gate, then public release" path. OpenAI calls it "our most robust safety stack ever," targeted at repeated misuse and sensitive cyber requests.
The API also gains Programmatic Tool Calling in JavaScript, explicit prompt cache breakpoints, original-resolution image input, and the multi-agent ultra effort tier. The single number Sam Altman highlighted: Sol emits 54% fewer output tokens than comparable models on coding tasks — at $30 / M output tokens, that lands directly on customer invoices.
Win the bet, and Sol takes half the API market on "half of Fable 5's price plus a Terminal-Bench beat." Lose it, and that 15-point SWE-Bench Pro gap becomes Anthropic's first marketing slide within a week of real workload runs.
via OpenAI / Simon Willison / The Decoder / Axios
