OpenAI's GPT-5.6 family goes public Thursday, capping a nine-day stretch that brought SpaceXAI's Grok 4.5 and Anthropic's returning Claude Fable 5, with prices spanning $1 to $50 per million tokens.
Key Points:
- GPT-5.6 Sol, Terra and Luna reach general availability with tiered pricing from $1 to $5 per million input tokens.
- Grok 4.5 launched at $2 input and $6 output, ranking fourth on an independent intelligence index.
- Claude Fable 5 holds the top coding scores, including 80.4% on SWE-Bench Pro, at the steepest price.
GPT-5.6 Launch Week
OpenAI announced the GPT-5.6 series on Jun. 26 and held it in a limited preview coordinated with the US government before this week's public rollout. Sol, the flagship, costs $5 per million input tokens and $30 for output, while Terra runs $2.50 and $15. Luna, the fastest tier, sits at $1 and $6.
SpaceXAI beat that debut by a day. The company shipped Grok 4.5 to developers Wednesday and to the public Thursday at $2 per million input tokens and $6 for output. Elon Musk described it as "an Opus-class model, but faster, more token-efficient and lower cost."
Anthropic restored Claude Fable 5 worldwide on Jul. 1 after Washington lifted export controls imposed on Jun. 12, completing the three-way field.
The model costs $10 per million input tokens and $50 for output, the steepest rate of the three.
Also Read: Solana's ETF Moment Gets Harder To Ignore After New Bitwise Filing
Fable 5 Benchmark Lead
Vendor figures show Fable 5 ahead on the hardest coding tests, scoring 80.4% on SWE-Bench Pro against 64.7% for Grok 4.5 and 58.6% for GPT-5.5, the model GPT-5.6 replaces. The gap narrows on Terminal-Bench 2.1, where all three earlier flagships landed within roughly one point of one another. GPT-5.6 changes that chart, with OpenAI reporting Sol at 91.9% on the same terminal test through a new ultra mode that splits work across parallel subagents.
A fuller benchmark suite is promised at general availability.
Independent testing frames the contest around cost, with Artificial Analysis ranking Grok 4.5 fourth of 168 models on its Intelligence Index, behind Fable 5, GPT-5.5 and Claude Opus 4.8. A completed agentic coding task, however, costs $2.49 on Grok against $11.80 for Fable 5 running inside Claude Code, a gap wide enough to reshape purchasing decisions. The same testers flagged a rising hallucination rate, which jumped from 25% to 54% on one knowledge measure as the model grew more confident when wrong.
Grok 4.5 Expert Verdicts
Cursor chief executive Michael Truell, whose company helped train Grok 4.5, posted that the model has "become the daily driver for many on our team." Musk pegged it internally as roughly comparable to Opus 4.7, only faster, though reviewers noted both endorsements come from parties with a direct stake in the launch.
Analysts increasingly describe a routing market rather than a single winner, with cheap tiers such as Luna and Grok 4.5 absorbing bulk work while Fable 5 takes the problems nothing else solves. Sol's subagent mode targets the long terminal sessions in between.
Washington, meanwhile, now shapes the release calendar, because both GPT-5.6 and Fable 5 passed through government review before reaching users.
Fable 5's own June path shows what is at stake when regulators step in. Anthropic launched the model on Jun. 9, and Amazon researchers reported a jailbreak that surfaced software vulnerabilities, prompting export controls three days later. The model then spent 19 days offline before returning behind a classifier that blocks the reported technique in more than 99% of cases.
Read Next: Grok 4.5 Challenges OpenAI And Anthropic With Cheaper Agentic AI