Anthropic Launches Claude Opus 5, Plus GPT-5.6 Solves Erdős Problems & Imagination Models!

Anthropic has released Claude Opus 5, a model it describes as thoughtful and proactive, coming close to Fable 5's frontier intelligence at half the cost. It's priced the same as its predecessor, Opus 4.8, at five dollars per million input tokens and twenty-five per million output, versus Fable 5's double that rate. On coding and knowledge-work benchmarks it sets new highs for Anthropic, scoring over forty percent on Frontier-Bench compared with Opus 4.8's under twenty, though Anthropic acknowledges it still trails Mythos 5 on cybersecurity tasks. The company is positioning it for daily professional use rather than the longest autonomous runs, where Fable 5 remains the recommended choice. It's now the default on Claude Max and top of the list on Pro. A researcher going by Shouqiao Wang says he solved six open Erdős problems in five days using OpenAI's GPT-5.6 Sol, out of thirteen he attempted. He didn't need deep mathematical expertise to do it, just a carefully written prompt that functioned more like a contract than a question, specifying exactly what a finished proof had to establish and which near-misses wouldn't count. From there it became a loop of attempt, failure, diagnosis, and a fresh proof draft, with the model auditing its own work. Not everyone in the mathematics community is convinced these count as genuine solutions, but it's a striking demonstration of what a long, patient reasoning loop can achieve. OpenAI's ChatGPT Work agent can now use websites that require sign-in, rather than being limited to public pages as before. Users take over the cloud browser themselves to log in, then hand control back to the agent to continue the task, and that login persists across sessions so the process only needs to happen once. It addresses a real limitation, since agents restricted to logged-out pages could only accomplish so much on a user's behalf. It's a small update, but the kind of unglamorous plumbing that determines whether an agent is genuinely useful day to day. Induction Labs has introduced what it calls imagination models, a new foundation model architecture designed to learn from internet-scale video rather than labelled actions. Its first model, Photon-1, learned to use a computer by watching the equivalent of eighteen years of screen recordings, with no one labelling what was clicked or why. That's a markedly different training approach to the reinforcement-learning setups most computer-use agents rely on, betting that raw observation at enormous scale can outperform curated demonstrations. It's an early, unproven architecture from a little-known lab, but the underlying bet is the notable part.