Claude's Fable 5.2 beat OpenAI's GPT-6.2 Astra at real office tasks
Why it mattersBenchmark ties hide big gaps: Fable 5.2 scored 98% on data center analysis against Astra's 65%, so test models on your own tasks.
Why it mattersBenchmark ties hide big gaps: Fable 5.2 scored 98% on data center analysis against Astra's 65%, so test models on your own tasks.
Why it mattersSafety tests that an AI can recognise as tests prove less about how it behaves in real deployments, so discount such results.
Why it mattersNothing changes for a business yet. This is a physics result, but it shows OpenAI's GPT-6 Astra contributing to a formal proof.
Why it mattersAttackers can now turn a published fix into a break-in for about $1,000, so any delay in installing browser updates is a real exposure.
What happenedOpenAI has published what its AI produced on three open science problems: the Kakeya conjecture in mathematics, the Exponential Time Hypothesis in computer science, and quantum chromodynamics in physics. It also describes how outside researchers can check the work and extend it.
Why it mattersIf your systems run on COBOL or other niche languages, OpenAI is now testing ChatGPT on migrating that code.
Why it mattersDrug discovery teams can run a multi-database target search in one agent session, but the lead is only predicted, not lab-confirmed.
Why it mattersOpenAI's unreleased model is producing checkable maths results, so expect that reasoning strength in products you can buy later.
Why it mattersStartups can now follow a clear process to deploy GPT-6 in production without guesswork.
Why it mattersMulti-vector models improve retrieval accuracy by preserving token-level matches, requiring larger indexes but enabling stronger search without OCR for visual documents.
Why it mattersOpenAI is paying 14 outside projects to shape AI policy proposals, so expect its preferred ideas in upcoming regulatory debates.