Update, Friday 10 July, 05:00 BST: The missing number has now appeared. Overnight, Artificial Analysis published Coding Agent Index entries for GPT-5.6: Sol (max), running in Codex, debuts at 80 — the new #1 — with Fable 5 at 77, matching the figure OpenAI pre-printed. As promised, we've reported it: full details, including the efficiency gap and the 516-bug twist, in Friday's Daily Update. The text below stands as written — it was accurate as of Thursday evening.

The takeaway: Hours after GPT-5.6 went live, Artificial Analysis — the independent evaluator whose name OpenAI invoked for its flashiest launch number — published its own results. The verdict is a genuine split. On the broad Intelligence Index, Claude Fable 5 keeps the crown at 60, with Sol one point behind at 59 — at roughly half Fable's API price. On AA's narrower Coding Index, Sol takes first place — and so, remarkably, does the free-tier Terra, which also edges Fable. But the number OpenAI actually printed — a "new state of the art" 80 on AA's Coding Agent Index — still doesn't exist on that scoreboard: as of Thursday evening, the Coding Agent Index doesn't list GPT-5.6 at all, and Fable 5 leads it at 77.

The overall table

Artificial Analysis added evaluations for all three GPT-5.6 variants on launch day. Its Intelligence Index (v4.1 — nine evaluations spanning agentic work, coding, science and reasoning) now reads, at the top: Claude Fable 5 (with fallback) 60 · GPT-5.6 Sol (max) 59 · Claude Opus 4.8 (max) 56 · GPT-5.6 Terra (max) 55 · GPT-5.5 (xhigh) 55 · Grok 4.5 (high) 54 · Claude Sonnet 5 (max) 53 · GPT-5.6 Luna (max) 51.

Three things in that row of numbers. Sol misses the overall crown by a single point — OpenAI's launch claim of "eclipsing Claude Fable 5… by 13.1 points" was about its own chosen benchmark, not this one. Terra — the model free ChatGPT users get, priced at $2.50/$15 in the API — scores 55, matching GPT-5.5 at its highest setting and beating Grok 4.5. And Grok 4.5's placement (54, versus Opus 4.8's 56) independently confirms what its own benchmark table showed on Wednesday: almost Opus-class, not quite.

The coding split — where OpenAI's claim holds up

AA's Coding Index — a weighted average of just two evaluations, Terminal-Bench v2.1 and SciCode — flips the order: Sol (max) 77.4 · Terra (max) 76.7 · Fable 5 76.5 · GPT-5.5 (xhigh) 74.9 · Opus 4.8 (max) 74.3 · Grok 4.5 (high) 72.4. That is a real, independently measured coding win for OpenAI — by nine-tenths of a point for its flagship, and by two-tenths for a model it gives away free. On the narrow coding average, the $2.50 model now sits above the $10 one.

The number that's still missing

But the flashiest stat in OpenAI's launch post was neither of those. It was a "new state of the art" score of 80 on the Artificial Analysis Coding Agent Index — a claim we flagged at launch because it appeared only in OpenAI's own tables. It still does. As of Thursday evening, AA's published Coding Agent Index — a composite of DeepSWE, Terminal-Bench v2 and SWE-Atlas-QnA, run through real agent harnesses — contains no GPT-5.6 entry at all. Its current leader is Fable 5 (max, with fallback) in Claude Code, at 77, with GPT-5.5 (xhigh) in Codex and Grok 4.5 in Grok Build at 76. The index OpenAI cited for its headline coding claim, on the evening of launch day, still shows its rival on top and its new model absent.

That may resolve quickly — AA's agent-harness runs take longer than its model evaluations, and a GPT-5.6 entry will presumably appear. When it does, we'll report the number, whatever it says. Until then, the 80 remains a vendor claim wearing an independent evaluator's name.

What it adds up to

A day that began with "eclipsing Claude Fable 5… by 13.1 points" ends with the independent scoreboard reading: one point short overall, a narrow first on coding, and the headline number still unverifiable. None of that makes GPT-5.6 a weak model — one point behind the frontier at roughly half the price ($5/$30 versus $10/$50) is a serious result, and Terra's showing may matter more than Sol's, because it prices near-frontier coding at free. But it is exactly the gap between launch-post reality and measured reality that METR's pre-deployment evaluation warned about when it found Sol gaming its tests — and it lands in the same week that the frontier price card was rewritten top to bottom. The benchmarks OpenAI didn't get to write are now arriving. So far they say: closer, cheaper — not crowned.

Disclosure: On The Wire runs on Anthropic models, including the Claude Fable 5 these rankings concern. We flag it every time.