GPT-5.6 vs. Grok 4.5 vs. Fable 5: A chaotic week reshaping the AI landscape

OpenAI\'s GPT-5.6 series was publicly released on Thursday, ending an intensive nine-day release period. Previously, SpaceXAI\'s Grok 4.5 and Anthropic\'s Claude Fable 5 have been unveiled, with a pricing range of US$1 to US$50 per million tokens.

Core Points

Three versions of GPT-5.6, Sol, Terra and Luna, are fully launched, and input token pricing ranges from US$1 to US$5 per million.
Grok 4.5 was released at a price of $2 per million input tokens and $6 per million output tokens, ranking fourth in the Independent Intelligence Index.
Claude Fable 5 maintains the highest score in programming tests, with SWE-Bench Pro scoring 80.4%, but is also the most expensive priced.

GPT-5.6 Release Week

OpenAI announced the GPT-5.6 series on June 26, with a limited preview after coordination with the U.S. government, and was subsequently launched to the public this week. The flagship version of Sol is priced at $5 per million in input tokens and $30 per million in output tokens; the Terra version is priced at $2.5 and $15; and the fastest Luna version is priced at $1 and $6.

SpaceXAI released its own product a day ahead of schedule. The company made Grok 4.5 available to developers on Wednesday and made it available to the public on Thursday, priced at $2 per million input tokens and $6 per million output tokens. Elon Musk described it as an \"Opus level model, but faster, more token efficient, and lower cost.\"

Anthropic resumed providing Claude Fable 5 globally on July 1. Export controls previously implemented by Washington on June 12 were lifted, completing the three-party issuance pattern. The model is priced at $10 per million of input tokens and $50 per million of output tokens, the highest price of the three models.

Fable 5 benchmark leads

Vendor data shows that Fable 5 leads the most difficult programming tests, scoring 80.4% on SWE-Bench Pro, compared to 64.7% for Grok 4.5 and 58.6% for GPT-5.5 (the model replaced by GPT-5.6). The gap narrowed in Terminal-Bench 2.1, with the scores of the three previous early flagship models within a percentage point of each other. GPT-5.6 changed this situation. OpenAI reports that Sol achieved a score of 91.9% in the same terminal test through a new super model of assigning tasks to multiple parallel sub-agents. A more complete benchmark suite will be available when it is fully available.

Independent testing looks at the competition from a cost perspective. Artificial Analysis ranks Grok 4.5 fourth in its Intelligence Index (out of 168 models), behind Fable 5, GPT-5.5 and Claude Opus 4.8. However, completing an agent-style programming task costs $2.49 on Grok and $11.80 to run Fable 5 in Claude Code, which is enough to change purchasing decisions. The same tester also pointed out that hallucination rates are rising: on one knowledge dimension, as models become more confident when making mistakes, hallucination rates jump from 25% to 54%.

Grok 4.5 Expert Review

Michael Truell, CEO of Cursor, whose company helped train Grok 4.5, wrote that the model \"has become the daily choice of many in our team.\" Musk internally rated it as roughly equivalent to Opus 4.7, only faster. However, commentators pointed out that both endorsements came from parties with direct interests in the release.

Analysts are increasingly describing the market as a routing market rather than a single winner-with cheaper versions like Luna and Grok 4.5 taking on a lot of routine work, while Fable 5 handles problems that other models cannot solve. Sol\'s sub-agent model targets long terminal sessions in between.

Meanwhile, Washington is influencing the release schedule. Both GPT-5.6 and Fable 5 reached users after government censorship. Fable 5 \'s experience in June illustrates the risks it can face when regulatory intervention comes in. Anthropic released the model on June 9, and Amazon researchers reported discovering a jailbreak that exposed software vulnerabilities, triggering export controls just three days later. The model was therefore rolled offline for 19 days, and then re-launched with a classifier that blocked reported attack techniques in more than 99% of cases.