GPT-5.6 vs Claude Fable 5: pricing, safety and performance

OpenAI released GPT-5.6 as three distinct large language models: Sol, Terra and Luna. Sol is available to paid ChatGPT subscribers and is priced at $5 per million input tokens and $30 per million output tokens. Anthropic’s Claude Fable 5 lists $10 per million input tokens and $50 per million output tokens and has been operating with temporary access limits ahead of a planned July 19 shift to usage credits.

OpenAI’s split model strategy assigns different cost and capability targets to each model. Luna is priced at $1 per million input tokens and $6 per million output tokens. Benchmark reports show Luna outperforming Anthropic’s Opus 4.8 on coding tasks. Sol is positioned as the mid-tier model and is included in ChatGPT paid plans without an announced expiration.

Anthropic removed Fable 5 from global availability after researchers reported a jailbreak that could turn the model into an unintended vulnerability scanner. U.S. officials restricted access on June 12. Anthropic took the model offline for 19 days, added a safety classifier and reintroduced Fable 5 on July 1 with compressed access limits. The company posted on its account on July 12 that access on paid plans would be extended through July 19.

Public benchmark comparisons show mixed results. On the Artificial Analysis Coding Agent Index, Sol scored 80 and Fable 5 scored 77.2; tests run on Sol used roughly half the tokens, ran in under half the time and cost about a third as much. On Agents’ Last Exam, which tests professional workflows across 55 fields, Sol scored 53.6% versus Fable 5 at 40.5%. Terminal-Bench 2.1 in a four-agent Sol configuration reached 91.9% against Fable 5’s 83.1%. On a broader Intelligence Index that aggregates nine benchmarks, Fable 5 held a one-point lead over GPT-5.6.

Independent practical tests returned nuanced results. In a creative-writing prompt, both models produced long-form narratives; reviewers judged Fable 5 stronger on cultural specificity and implied causality, and Sol clearer in exposition. In an associative-thinking prompt using a twig metaphor to discuss worker exploitation, reviewers found Sol more explicit and Fable 5 more embedded in metaphor. On a bridge-crossing logic puzzle both models returned the same incorrect numeric answer, 17 minutes, while the prompt allowed for a 10-minute solution if all four crossed together at the slowest pace. In a one-shot coding task to build a typing-based shooter game, Fable 5 produced music, animated enemies, words-per-minute tracking and power-ups; Sol produced a simpler user interface and fewer atmospheric details.

Pricing and access differ between the vendors. OpenAI’s Sol, Terra and Luna are available inside ChatGPT paid subscriptions with no expiration announced. Anthropic has scheduled a transition of Fable 5 to a usage-credits system on July 19; if implemented, the $10/$50 per-million-token rates would apply. Anthropic and OpenAI have both applied model-level changes tied to safety and product design: Anthropic added a safety classifier before restoring Fable 5, and OpenAI delivers three separate models with distinct pricing and capability tiers.