AINEWS
Anthropic's newest model jumped from below 10% to 88% in six months on the hardest tier of FrontierMath, an expert authored math benchmark run by Epoch AI — and the pace of gain, not the leaderboard, is now the story.
Six months ago, Anthropic's best model cracked fewer than one in ten of the hardest problems on FrontierMath, an expert-authored math benchmark widely cited as one of the toughest tests of AI reasoning. On Friday, the company reported that its newest model, Claude Fable 5, hit 88% on the same tier. The acceleration, not the leaderboard, is the story.
FrontierMath, operated by Epoch AI, is a curated set of problems written by working mathematicians and designed to resist the pattern-matching that carried earlier large language models. Its "tier 4" problems are the hardest subset; the current variant is labeled v2, and scores are not directly comparable across versions. In early 2026, Anthropic's Opus 4.5 scored below 10% on tier 4. By June 13, 2026, Claude Fable 5 reportedly reached 87% on tiers 1–3 and 88% on tier 4 on the same test.
That puts Fable 5 about 13 points ahead of OpenAI's GPT-5.5, which the-decoder reports reaches roughly 75% on tier 4. GPT-5.6 is reportedly in development at OpenAI, though the company has not confirmed timing. All comparisons used Epoch AI's standard scaffold with maximum reasoning effort, which is load-bearing context: this is benchmark performance with heavy scaffolding support, not autonomous mathematical reasoning.
A few caveats belong near the top of any honest reading. FrontierMath is curated, and tier 4 is its most recent variant, so gains there can reflect scaffold and reasoning-effort changes as well as model capability. Epoch AI is the primary source for the figures cited; the upstream reporting is a single outlet (the-decoder) summarizing Epoch AI's leaderboard. Anthropic has not, as of this writing, published its own confirmation of the "Fable 5" naming or the 88% figure, and the model name is unusual; the more familiar Anthropic family in this period includes Claude Opus 4.5 and its successors.
The harder question is what tier 4 still measures once multiple frontier models clear it. A benchmark designed to be unsolvable becomes, by construction, less useful the moment it is solved. If the curve from 10% to 88% in roughly six months holds across the next cycle, the next move belongs to the benchmark operator: harder problems, new problem families, or some acknowledgment that the leaderboard has stopped ranking capability in the way it once did.
There is a separate, real-world beat worth watching. Recent reporting has suggested that both an OpenAI model and an Anthropic model referred to as "Claude Mythos" have tackled a longstanding Erdős problem, a class of open conjectures in discrete mathematics named for the prolific twentieth-century mathematician Paul Erdős. The "solved" framing in the source material needs primary-source verification, including which problems, what kind of solution, and on whose authority, before any claim of mathematical breakthrough can stand. Until then, it is a depth beat, not a fact.
What to watch next: Epoch AI's next leaderboard update, any Anthropic confirmation of the Fable 5 release, and OpenAI's first public numbers for GPT-5.6. The slope of the curve is the story; the names on the leaderboard will keep changing.