OpenAI ships GPT-5.6 Luna, Terra, Sol as Fable 5 leads on code

OpenAI shipped a new flagship trio on July 9th, and Simon Willison's write-up is a compact place to see the whole picture. The GPT-5.6 family, smallest to largest, is Luna, Terra, and Sol, priced at $1/$6, $2.50/$15, and $5/$30 per million input/output tokens. All three share the same February 16th, 2026 knowledge cutoff, a one million token context window, and 128,000 tokens of maximum output.

The pitch OpenAI is leading with is Agents' Last Exam, a benchmark that evaluates professional workflows across 55 fields. Their claim is that "GPT-5.6 Sol sets a new high of 53.6, eclipsing Claude Fable 5 (adaptive reasoning) by 13.1 points," with the smaller Terra and Luna reportedly outperforming Fable 5 "at around one-sixteenth the cost." That last figure is the number cost-sensitive teams will actually stare at.

The awkward part is coding. On SWE-Bench Pro, Claude Fable 5 scored 80% and Sol landed at 64.6%. OpenAI's response was not to concede the coding crown but to publish a separate piece arguing that "~30% of SWE-bench Pro tasks are broken." Willison's own early-access testing, as he tells it, hasn't surfaced a clear advantage over Fable on complex coding jobs. Alongside the models, the API picks up Programmatic Tool Calling for JavaScript orchestration, parallel multi-agent support, prompt cache breakpoints, and an image detail option that skips automatic resizing.

The honest caveat is that these are vendor-reported numbers on a benchmark whose validity OpenAI is now itself contesting, so take the specifics as reported rather than settled. What the reporting doesn't give you is independent replication, or a sense of how much of Fable 5's SWE-Bench Pro lead survives once the disputed tasks are stripped out. For teams paying inference bills at scale, Terra and Luna at a fraction of Sol's price are the more interesting story than any single leaderboard number, if the cheaper tiers hold up on real workflows.