OpenAI has released GPT-5.6 in three separate models, Sol, Terra and Luna, shifting the competitive landscape with Anthropic’s flagship public model, Fable 5. In coding, Sol, the top GPT-5.6 model, held an edge. In writing and general conversation, Fable 5 maintained its strengths. With performance differences not large, analysts say users’ choices are increasingly likely to be driven by price and accessibility.

Decrypt, a blockchain media outlet, reported on July 18 that the core of the comparison is a head-to-head matchup between Sol, the top GPT-5.6 model, and Fable 5.

On price competitiveness, GPT-5.6 edged ahead. Sol costs $5 per 1 million input tokens and $30 per 1 million output tokens, while Fable 5 costs $10 and $50, respectively, about double. Luna, a lower-tier model, costs $1 per 1 million input tokens and $6 per 1 million output tokens and still outperformed Anthropic’s Opus 4.8 in coding performance. Fable 5 is also set to switch back to a usage-based credit billing system from July 19, raising expectations of a heavier cost burden.

Fable 5 has also changed its access policy several times over the past month. The U.S. government banned use of the model after confirming in June that Amazon researchers could use a jailbreak technique to exploit it as an unintended vulnerability scanner. Anthropic then took Fable 5 down worldwide for 19 days, applied a new safety classifier and resumed service on July 1. The end date for availability under paid subscription plans was also pushed back several times, to the 7th, 12th and 19th. The company said it would keep Claude Code’s weekly usage limit raised by 50 percent.

In benchmarks, GPT-5.6 Sol led in coding. In an AI coding agent evaluation, Sol scored 80, ahead of Fable 5’s 77.2, while using about half the tokens and taking less than half the processing time. In “Agents Last Exam,” which performs 55 professional tasks, Sol scored 53.6 percent, well above Fable 5’s 40.5 percent. In TerminalBench 2.1, in top performance mode, Sol scored 91.9 percent and Fable 5 scored 83.1 percent. In an intelligence index aggregating 9 benchmarks, however, Fable 5 scored 1 point higher than GPT-5.6, indicating the overall performance gap was not large.

Results were mixed in practical-use tests. In creative writing based on a time-travel paradox, Fable 5 received higher marks for more naturally linking cultural background and causality. GPT-5.6 Sol, by contrast, was closer to a more direct style of explaining story structure. In prompts requiring metaphor, Fable 5 stood out for symbolic expressions using objects, while Sol focused on clearly explaining meaning.

On a logic problem, both models made the same mistake. In a modified version of the classic problem of 4 people crossing a bridge with 1 torch, both models answered 17 minutes, but the correct answer was 10 minutes, the crossing time of the slowest person. The assessment was that they applied a representative solution learned previously as-is.

In a coding test to build a browser-based typing shooter game, Fable 5 produced a more complete result. Sol offered original ideas, but its graphics and visual effects were relatively simple. Fable 5 implemented music and sound effects, animation, a display of words per minute, and power-up elements, delivering a more polished output.

Decrypt assessed that Sol is advantageous if users prioritize coding benchmarks and cost efficiency, while Fable 5 remained competitive for general conversation, writing and creative work. It also analyzed that with performance differences not large, and with Fable 5 returning to usage-based billing while GPT-5.6’s Sol, Terra and Luna are included in paid ChatGPT plans, users’ choices are likely to be determined more by price and accessibility than performance.