GPT-5.6 'Crushes' Claude Fable 5: OpenAI is Confident and Claims the AI Throne
OpenAI has made the GPT-5.6 family available to everyone after a period of limited preview. It includes three distinct models: Sol, the new flagship model, Terra, designed for everyday work with a good balance between cost and capability, and Luna, the most economical version of the range.
The thread of this generation is efficiency. On the Agents' Last Exam (55 professional fields with extended workflows), Sol scores 53.6, crushing – or rather, 'asphalted', as stated in the press release – Claude Fable 5 by 13.1 points. Even set to a medium reasoning level, Sol remains ahead of Anthropic's model by 11.4 points at about a quarter of the estimated cost. The gap widens as one moves down the spectrum: Terra and Luna outperform Fable 5 spending only about a sixteenth as much. On the Artificial Analysis Intelligence Index, Sol, at maximum reasoning, stops just one point short of Fable 5, but completes tasks in 61% less time and at about half the estimated cost.
Fewer tokens for more useful work: OpenAI's recipe with GPT-5.6. The most operational novelty is ultra, the mode that coordinates four agents in parallel (up to sixteen in the most extreme tests) to accelerate complex tasks, exchanging higher token consumption for better results and reduced response times. Those working within the Responses API can replicate the same scheme via the multi-agent beta function. This is complemented by Programmatic Tool Calling, which allows the model to write and run small programs capable of orchestrating external tools and filtering intermediate results without sending each individual response back to the model itself.
On the programming front, Sol scores 80 points on the Artificial Analysis Coding Agent Index, 2.8 points above Fable 5, using less than half the output tokens and about a third of the time and cost. New records are also set on Terminal-Bench 2.1 and DeepSWE. Computer use capabilities grow alongside: 92.2% on BrowseComp and 62.6% on OSWorld 2.0, where Sol surpasses Opus 4.8 using 85% less output tokens. The model inspects and corrects the graphical result produced by itself; it does not merely generate code, resulting in presentations, documents, and spreadsheets that more faithfully adhere to the reference models provided by the user.
On the cybersecurity front, the numbers are also clear: 73.5% on ExploitBench2 (it was 47.9% with GPT-5.5), almost double the success rate on ExploitGym3 (from 15.1% to 24.9% within two hours, up to 33.7% with six hours available) and 71.2% on SEC-Bench Pro compared to the previous 45.8%. OpenAI maintains, however, that the model does not exceed the Critical threshold either in biology or cybersecurity: it is more effective at finding and correcting vulnerabilities than in conducting end-to-end autonomous attacks against well-protected targets. The most sensitive functions remain reserved for those participating in the Trusted Access for Cyber program, which as of September 1 will require the activation of hardware passkeys to avoid losing access to the most capable models in this area. New control systems, OpenAI explains, block about ten times more potentially harmful activity compared to the previous generation, following approximately 700,000 hours of dedicated GPU A100e for automated red teaming.
On the commercial front, GPT-5.6 is already active on ChatGPT, Codex, and APIs, with a global rollout currently underway. The prices per million tokens remain tiered: Sol costs $5 in input and $30 in output, Terra $2.50 and $15, Luna just $1 and $6.