SHELFCRITTER AI Digest

MODEL

ChatGPT 5.5 Pro

modeltopic-noteopenaireasoning

Overview

ChatGPT 5.5 Pro is OpenAI‘s top-tier ChatGPT product surface (built on the GPT-5.5 generation released April 24, 2026) that gained AI Digest corpus coverage on May 10 when Fields Medalist Tim Gowers documented the model solving previously-open math research problems unaided in under an hour. The headline is a frontier-capability beat distinct from prior LLM-mathematician headlines — research mathematics on a working Fields Medalist’s own active problems, not olympiad questions or formal-proof pipelines.

Timeline

  • 2026-05-10-AI-Digest — Tim Gowers publishes a primary blog post on May 8 describing a session in which ChatGPT 5.5 Pro improved an exponential bound to a quadratic one in 17 minutes 5 seconds, then generalised exponential to polynomial by 31 minutes 40 seconds — both on previously-open problems Gowers gave it cold. Gowers describes the result as “above the lower bound for contributing to mathematics,” and MIT student Isaac Rajagopal calls the key idea “completely original.” The OpenAI / arXiv “Early science acceleration experiments with GPT-5” paper (arXiv:2511.16072) is the broader research programme this fits inside. Separately, Simon Willison’s same-day post on prompting Claude / ChatGPT 5.5 Pro for HTML rather than Markdown uses ChatGPT 5.5 Pro on a Linux exploit walkthrough as its worked example — a developer-tooling-affordance discovery distinct from the math-research beat.

Key Developments

  1. Fields-Medalist Research-Math Beat: Gowers is a working Fields Medalist working on his own active research — not olympiad problems and not a formal-proof assistant pipeline. The result also doesn’t contradict 2026-05-09-AI-Digest‘s “models faking reasoning traces under evaluation pressure” finding: Gowers verified the output is correct mathematics, not that the chain-of-thought reflects the actual computation. Output correctness and trace faithfulness are orthogonal evidence streams.

  2. Source Discipline: Gowers’s blog is the primary source; The Decoder is the secondary. The Decoder’s “MIT researchers calling ideas ‘completely original’” framing is a single-student attribution (Isaac Rajagopal), not a generic-researcher claim.

  3. HTML-Output Affordance: Simon Willison’s amplification of Thariq Shihipar’s argument that asking Claude (or ChatGPT 5.5 Pro) to emit HTML rather than Markdown unlocks SVG diagrams, interactive widgets, in-page navigation — a prompt-pattern that surfaces capability the default Markdown framing was muting.

See also: OpenAI, GPT-5.5, Tim Gowers, Simon Willison, MOC - Major Companies.