Fable 5 vs GPT 5.5 vs Gemini 3.1 Pro.

Every row from the launch table where at least one rival is reported. "—" means Anthropic didn't report a number for that model.

* Starred rows show a larger gap between Mythos 5 and Fable 5 because blocking safeguards make Fable 5 perform closer to Opus 4.8 on cyber- and bio-adjacent questions; elsewhere the two are within 1–3 points and the table shows the higher. Anthropic chose these benchmarks, ran this table, and published it on launch day. Keep that sentence next to every number above.

Terminal-Bench is the closest agentic-coding race — and the one where the rivals brought their own harnesses. The rival numbers aren't bare models: GPT 5.5 ran through Codex CLI, Gemini 3.1 Pro through Gemini CLI. Codex CLI at 83.4% is five points off the lead on the one row where the rivals brought their own tooling — that holds up. It's also the lesson of Ch 35: Codex isn't a competing model in my stack, it's the night shift — same repo, different contract — which is why the rivals' CLI numbers here deserve respect.

Blueprint-Bench 2 is the closest race on the table. 38.6% vs 36.2% on spatial reasoning — 2.4 points. If your workload is floor plans, diagrams, physical layout, this table gives you no reason to switch anything in either direction.

Legal Agent Benchmark says more about the eval than the loser. 13.3% vs 2.1% vs 0.0% — when the leader can't clear 14% and one model posts a zero, the row is telling you the task class is barely solved, not that one vendor cracked it.

Gemini 3.1 Pro is third on all but one row it appears in — it edges GPT 5.5 on Humanity's Last Exam without tools. Read that with both caveats on. First: this is Anthropic's table of Anthropic-chosen benchmarks. Second: Google's cycle already moved. On May 19 this site's research notes logged Gemini 3.5 Flash clearing 3.1 Pro on Google's own agentic boards — vendor slides, same discipline applied — with a 3.5 Pro promised. Anthropic launched against the model Google is in the middle of replacing. That's how launch tables work; it's also why they expire.

This book has one rule for launch-day numbers and it didn't expire on June 9. Berkeley RDI reward-hacked eight major agent benchmarks in April 2026 — since then the standing discount on every public score is 10–15 points, and every external benchmark gets paired with a private eval (Ch 24, Ch 25). That discount applies to all three columns above, including the one with the accent color.

What survives the discount is shape, not precision. SWE-Bench Pro at 80.3% vs 58.6% is a 21.7-point gap — that doesn't dissolve into contamination noise. FrontierCode Diamond at 29.3% vs 5.7% is a 5× gap on the eval built to resist saturation. Terminal-Bench at 88.0% vs 83.4% is inside the discount band — treat that row as a tie until your own numbers say otherwise.

If you run a second-prior workflow — Claude on the day shift, Codex on the night shift, Gemini as the idea machine (Ch 35) — nothing in this table changes your setup today. Two priors triangulate; that logic got stronger, not weaker, now that the gaps between them are this uneven across task classes.

The one time-boxed action: Fable 5 sits inside paid-plan limits June 9–22, 2026, which makes the Claude leg of a three-way eval free for two weeks. Point all three models at the same three workloads, keep the transcripts, read the results against your own bar. Re-tier after that — not after a slide, mine or theirs.

The question every launch invites — which model won the cycle — is the wrong question. The table above is one vendor's thirteen favorite rows on launch day; the May table was another vendor's. The operators who win cycles are the ones whose private evals were already running when the slides dropped.