GPT-6 Astra solves the last FrontierMath Tier 4 problem
A research-grade math benchmark built to stay ahead of models just ran out of road, Apple's dedicated neural chip gets its autopsy, and Oracle puts a price on the AI jobs it is cutting.
Epoch AI's FrontierMath Tier 4 is saturated. GPT-6 Astra solved the final unsolved problem — a research-level question authored by combinatorialist Jay Pantone — and Epoch now considers the tier closed out at 97.6%. When Tier 4 launched in mid-2025, the best model cleared just 5% of it; the runners-up are GPT-5.6 Sol at 83.0% and GPT-5.6 Terra at 68.3%. Mathematicians did not report the model exploiting an unintended shortcut this time, which is what makes the result land. The honest read: a benchmark that was supposed to measure the reasoning frontier for a decade flatlined in fourteen months, so the signal moves to open-ended problems and long-horizon evals — and to cautionary notes like the one we filed in September, OpenAI quietly rewrote Astra's benchmark numbers after launch.
A reverse-engineering deep dive explains why Apple folded the Neural Engine into the GPU. Eileen Yoon mapped the M1 ANE's internals — 16 fixed-function cores, no instruction set, a 1 MiB kernel memory that cannot be fed from L2 — and showed the chip is pinned by memory bandwidth, not compute: peak throughput needs 162 operations per byte fetched from DRAM, and the kernel path tops out at 38 GB/s. The ANE was architected around CNNs' predictable weight reuse, the assumption transformers broke; the M5's "LLM performance" headline, with ANE cores absorbed into the GPU, reads as Apple conceding the standalone NPU.
Oracle added $700 million to its restructuring bill — to pay people it is cutting for AI data centers. The company disclosed the charge as it ramps capex, taking the total restructuring cost to about $2.8 billion, according to Reuters. The same report had shares rising on an AI cloud backlog that beat estimates — the same ledger we read last night in Oracle just proved the AI cloud trade is real — the bill for it is coming due anyway. Labor severance is now a line item inside infrastructure economics.
What to watch: Epoch's next hard eval — FrontierMath's open-problems track and MirrorCode — is where the differentiation signal now lives.
If the hardest closed-form math benchmark is solved, what should we ask models to prove next?
