GPT-5.6 Launches Worldwide as Claude 4.5 Benchmarks Leak

OpenAI opened GPT-5.6 to every developer and enterprise account worldwide this week, ending a months-long "US government preview" restriction. In the same stretch of weeks, a benchmark leak began circulating that shows Anthropic's newest model ahead of GPT-5.6 on several standard tests. Both events raise the same question for engineering teams choosing a frontier model: what changed, and does it matter enough to re-benchmark now?

What Actually Shipped

OpenAI officially announced that GPT-5.6 is now fully available to developers and enterprise users worldwide, ending the restricted rollout that had limited access to government-approved organizations. That restriction followed weeks of regulatory drama: GPT-5.6 spent its "limited preview" period cleared only for government-approved organizations before the Trump administration greenlit a public rollout. OpenAI's announcement quotes CEO Sam Altman calling GPT-5.6 "the best model we have ever produced."

OpenAI marked the wider release by unveiling a new agent, ChatGPT Work, the same day. It's billed as a combination of ChatGPT and Codex, letting non-technical users apply Codex's coding capabilities to everyday, non-coding tasks, and it runs on the GPT-5.6 model suite of Sol, Terra, and Luna. OpenAI is positioning the agent as a direct competitor to Anthropic's Claude Cowork, which pairs Claude with Claude Code.

Plus and Business subscribers will get access "over the next few days," OpenAI wrote, adding that the rollout "is starting globally and will continue gradually toward full availability over the next 24 hours." The three GPT-5.6 tiers carry distinct list prices per million tokens:

Altman has said the new models are orders of magnitude more efficient and cost-effective than earlier versions. He told CNBC that Sol is 54% more token-efficient for coding tasks.

What the Leaked Benchmarks Say

A benchmark leak tied to Anthropic's newest model, circulating alongside the GPT-5.6 launch, claims wins on three widely cited tests. According to that report, the Anthropic model leads GPT-5.6 by roughly 3.2 percentage points on MMLU, by 4.7% on HumanEval code-generation accuracy, and by 1.5% on GSM8K math reasoning — its narrowest margin. The same report claims Anthropic's stock surged in pre-market trading and says market analysts widely read the leak as an intentional "accidental" move to counter GPT-5.6's launch buzz.

A separate leaderboard, citing Vellum and MarkTechPost data, lists different figures for the same rivalry. It puts Anthropic's model — named there as Fable 5, the same Anthropic model discussed above — at 80.3%, ahead of Anthropic's own Opus 4.8 at 69.2%, while GPT-5.6 Sol trails at 64.6%, a 15.7-point gap in Fable 5's favor. OpenAI, for its part, says Sol "sets a new state of the art at 80, 2.8 points above Fable 5, while using less than half the output tokens, taking less than half the time, and costing about one-third less."

Why METR Isn't Buying It

Independent evaluator METR added a note of caution about Sol's numbers on June 26, writing: "we do not consider any of these numbers to represent a robust measurement of GPT-5.6 Sol's capabilities." That skepticism targets the exact kind of headline figure OpenAI is promoting, and it predates any independent scrutiny of the leaked Anthropic numbers.

Commentator Brian Wang, writing at NextBigFuture, offered a more colorful framing of the same rivalry: Fable 5 behaves like "a wise owl that thinks wider and asks better questions," while Sol acts like "a rottweiler that grabs the problem and doesn't let go."

What This Means for Your Stack

None of this week's numbers, vendor-claimed or leaked, should move a production model decision on their own. OpenAI's "state of the art at 80" figure and the leaked MMLU/HumanEval/GSM8K numbers describe different tests run by different parties, so a side-by-side ranking from headlines alone is not reliable. METR's caution about Sol is a useful reminder that the same standard should apply to any leaked numbers favoring Anthropic until a named source verifies them.

Teams with an existing evaluation harness are better positioned than teams chasing headline percentages. Rerunning your own task suite against both model families, on your own workloads, will tell you more than any leaked spreadsheet or vendor blog post this week.

Pricing is the one figure in this story that isn't in dispute. Sol's $5/$30 per-million-token rate sits well above Terra and Luna, so teams building high-volume agent workflows should model cost against measured token usage, not against Altman's efficiency claim, before moving tiers.

What to Watch

• Whether OpenAI or an independent evaluator publishes a reproducible methodology behind Sol's "54% more token-efficient" and "state of the art at 80" claims.
• Whether the leaked MMLU/HumanEval/GSM8K numbers get confirmed by a named source, since they are currently circulating without one — the separate Vellum/MarkTechPost leaderboard figures already cite named sources.
• How ChatGPT Work's early reviews compare with Claude Cowork once both agents reach general availability, since that comparison will matter more to most teams than the benchmark leak.