Fetching from the wire…
01
02
03
04
05
06
07
08
Models2026-09-23 · source-backed
"The Plunging Price of Thought" finds about 47% per quarter over three years across GPQA Diamond, FrontierMath and AIME, running about 66% per quarter right after a new state of the art and slowing to 32% two years later. One GPQA Diamond performance level got 725x cheaper between January 2025 and mid-2026. Yesterday's simultaneous price cuts are that curve, not a price war anyone chose.
Each link below shares sources, entities, or timing with this story.
Epoch AI's open-problems board lists a curve over Q of rank at least 30 posted August 20, credited to Claude with Levent Alpöge and Ava Howell, followed by a rank-31 curve on August 23 from the same team (Epoch AI). That beats the Elkies and Klagsbrun record of 29 from 2024, w...
Epoch puts OpenAI's run rate above $40B, up from $13B a year ago, and Anthropic's at $65B as of end of July, against Exponential View's ~$175B annualized estimate for the whole deduplicated generative AI industry as of June. The analytically useful part is the caveat: when tok...
Two facts sit next to each other and neither cancels the other out. Anthropic published on September 4 that an internal general-purpose research model, roughly comparable to Claude Fable 5.1, formalized Fermat's Last Theorem in Lean over 11 days working largely autonomously. T...
A submission dated August 20 to the elliptic curve rank leaderboard credits Claude, working with mathematicians Levent Alpöge and Ava Howell, with a curve of rank at least 30, holding the smallest conductor, naive height, and Faltings height among rank ≥30 entries. Under GRH+B...
On June 12 Epoch released FrontierMath v2, which corrected errors in 42% of problems; the set is now 338 problems (295 in Tiers 1–3, 43 in the Tier-4 expansion) (Epoch AI). Claude Fable 5 took the top spot at 87% on Tiers 1–3 and 88% on Tier 4. If you've been citing FrontierMa...
Epoch AI logged the first Major Advance tier solution, on emptiness of the core in approval-based committee elections, open since Aziz, Brill and colleagues posed it in 2017. Becker, Greger and Peters worked interactively with GPT-6 Astra, and Peters said he doubts the team wo...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.