GPT‑5.6: A Critical Look Beyond the Hype

OpenAI officially released GPT‑5.6, rolling out three variants—Sol, Terra, and Luna—simultaneously, and reset subscription quotas for all users. The author notes this is the most significant version of the year and likely the last before the upcoming GPT‑6.x series.

Sol introduces two new reasoning tiers, Max and Ultra. Max allocates more time for deep thinking, while Ultra runs multiple sub‑agents in parallel, consuming substantially more tokens and requiring careful use. Official evaluations suggest Sol performs well while saving tokens.

Alongside the model update, OpenAI launched ChatGPT Codex (renamed from Codex) and ChatGPT Work. The latter hides technical details, targets non‑developers, and can integrate with tools such as Google Drive, Slack, and email to autonomously break down complex projects, execute steps, and only query the user when necessary.

The browser component also received a major upgrade: the Computer Use feature now works more smoothly, supports authenticated sites, multiple tabs, file downloads, and retains user tags, making the experience faster and more token‑efficient.

Official benchmark data compare the three GPT‑5.6 variants with Claude Fable 5. The Intelligence Index (average of nine tests) scores Fable 5 at 60, GPT‑5.6 Sol at 59, and Terra at 55. The Coding Agent Index (sum of DeepSWE, Terminal‑Bench, SWE‑Atlas‑QnA) places Sol first at 80, with Terra and Fable 5 tied at 77.

The author emphasizes that these numbers do not represent a pure model‑to‑model duel; different toolchains, reasoning tiers, and test sets can shift rankings. Max mode offers the highest reasoning strength, while a safety fallback configuration runs for Fable 5.

Simon Willison’s blog notes that GPT‑5.6 may not outperform Fable 5 overall. In his “intelligent agent ultimate exam” covering 55 long‑running professional workflows, Sol achieved a score of 53.6, 13.1 points higher than Fable 5’s adaptive‑reasoning mode, and in medium mode it cost roughly a quarter of Fable 5’s estimated expense.

Cost analysis shows the lowest‑priced configuration is gpt-5.6-luna (inference strength None) at $0.0071 per request, while the most expensive is gpt-5.6-sol (Max) at $0.4855.

On the SWE‑Bench suite, Fable 5 achieved 80% while GPT‑5.6 Sol reached 64.6%. OpenAI later clarified that about 30% of SWE‑Bench problems may be flawed—e.g., ambiguous descriptions or overly strict criteria—so the results should not be taken as definitive.

In the author’s personal experiments, Terra Ultra felt slower than GPT‑5.5, with the first conversation ending in an empty final message and the second mixing English and Chinese output. These issues led the author to question whether the model’s advantages in token efficiency and multi‑agent coordination outweigh its stability problems.

Considering the mixed evidence, the author suggests that while Sol excels in coding‑agent tasks and Luna/Terra offer cost benefits, Fable 5 remains stronger in comprehensive intelligence and complex engineering workflows. The author also mentions testing Grok 4.5, which performed well, and indicates a willingness to switch if GPT‑5.6 Sol does not meet expectations.