Laurenceo raises odds of 10 FrontierMath Erdos solutions by mid-2027 to 52.8%
A single unverified post reporting 7 of 68 problems solved on a near-standard budget has cut the remaining gap to three.
Laurenceo has raised its forecast that an AI model will solve at least 10 of the 68 problems in Epoch AI's FrontierMath Erdos benchmark under Epoch's standard protocol by 30 June 2027, from 20% to 52.8%. The question resolves yes only if Epoch publishes such a result under its standard per-attempt budget, and results that need larger budgets do not count. The change was made on 18 November 2026.
Why It Moved
The main evidence is a post reporting that 6.2 Astra solved 7 of 68 problems on a budget of about $240 per question, below the $300 standard, with each solution formally verified in Lean 1. On 28 October 2026 the top score was 3 of 68, after every model had scored 0% in August 2. The post is a single unverified item from x.com, and whether its budget matches the standard protocol is not confirmed. Against the forecast, reaching 5 solutions with larger budgets reportedly took over $220,000 of compute 3. Because each solved problem is discrete, the easier ones may already be taken, and the last few could come slowly or not at all.
Two developments would move the figure again. An Epoch listing of the 7 of 68 result on its official FrontierMath Erdos page as a standard-protocol result would strengthen the case, as would a next frontier model evaluated by Epoch reaching 10 or more. A stall at 7 through subsequent releases would lower it.
Evidence
- x.com (opens in a new tab): Epoch AI on X: "GPT-6.2 Astra sets a new record on FrontierMath Erdős, our benchmark of handpicked math problems, unsolved as of August 2026. With a ~$240 budge
- manifold.markets (opens in a new tab): Highest score on Epoch AI's FrontierMath Erdős benchmark (68 open Erdős problems) by November 30, 2026?
- epoch.ai (opens in a new tab): Announcing FrontierMath Erdős
The Forecast
Will any AI model solve at least 10 of the 68 problems in Epoch's FrontierMath Erdos benchmark under Epoch's standard protocol by 2027-06-30? The full reasoning, every revision and the resolution rule are on the forecast page. Laurenceo publishes a story whenever one of its forecasts moves 10 points or more in a day.
More Analysis
- Forecast Move · 18 November 2026Laurenceo lowers 2026 AI job-cut forecast to 10%
- Forecast Move · 17 November 2026Laurenceo raises odds of US curbs on Chinese cloud access to AI chips to 72.8%
- The Laurenceo View · 17 November 2026Compute is now an export control, and data-centre operators are in scope
- Forecast Move · 17 November 2026Laurenceo lowers odds that Kokotajlo's Automated Coder median lands by 2027.0 to 10%
- Forecast Move · 17 November 2026Laurenceo lowers odds of over 150,000 AI-attributed job cuts in 2026 to 25%
- Forecast Move · 16 November 2026Laurenceo raises Qwen 4 release forecast to 88% after Qwen4-Max launch
Laurenceo Intelligence is the daily AI intelligence briefing published in Bermuda by Laurenceo and edited by Jorai Laurenceo. Every story is researched, source-verified and published each morning; how it works. General information only, not legal, regulatory or financial advice.