The frontier model race has turned into an open price war. Anthropic and OpenAI shipped major new models within about 90 minutes of each other, both cheaper than what came before. Meanwhile AI moved from talking about science to doing it, with Claude surfacing a genuine biology discovery on its own.

Anthropic released Claude Opus 5.5 at $4 per million input tokens and $20 per million output, roughly 40% cheaper than the previous Opus and 30% faster. About 90 minutes later OpenAI countered with GPT-6 Sol at $2/$10 and GPT-6 Luna at $0.10/$0.50, both around 50% cheaper on the API. Opus 5.5 topped one leading intelligence benchmark, while OpenAI leaned on cost. Both firms are now pushing a measure of cost per successful task rather than raw price.

Why it matters: Each release delivers more capability for less money, good for buyers but a squeeze on the providers' margins.

Anthropic's biology lab ran roughly 950 Claude agents for 21 hours, burning 210 million tokens to sift more than 200,000 reverse transcriptases. The system narrowed 3,500 candidates down to 20 and flagged one standout, dubbed ART, that pairs a reverse transcriptase gene with a partner gene and repeated DNA sequences. Its structure resembles CRISPR-like systems and may cut, copy and paste DNA. Human scientists then confirmed ART produces small RNAs.

Why it matters: This is AI producing a real scientific finding at scale, not just generating hypotheses, though physical lab work still sets the pace.

OpenAI says a single internal model has solved over 100 open maths problems since late August, including a claimed breakthrough on the Navier-Stokes problem. To handle the flood of results, it has enlisted nine mathematicians at Princeton's Institute for Advanced Study to review them. The panel includes Timothy Gowers, Martin Hairer and Melanie Matchett Wood. Their job is to check correctness, originality and proper credit before anything is released.

Why it matters: It shows human oversight becoming a formal step in verifying AI-driven research rather than an afterthought.

At its Connect event Meta showed Muse Charm, a pocket-sized, Tamagotchi-style device powered by a personal Muse AI agent. It lets users chat with an AI companion in a playful, portable form. Meta also demonstrated VR glasses weighing about a fifth of the Quest 3 headset, able to turn any flat surface into a keyboard and carrying an always-on AI agent.

Why it's interesting: It signals Meta pushing consumer AI hardware into playful, wearable formats ahead of rivals.

A user set an AI agent loose on an airline's customer service over a delayed flight. The agent kept pushing where a person would give up and secured a $250 credit and a rebooking within minutes. It is a small example of automated escalation changing how support works.

Why it's interesting: Agents that never tire and never back down could become a major cost centre for companies handling refunds and complaints.

Source: Yahoo Finance

A fund called Bracket22 is reportedly operated fully by AI agents. It surfaced alongside other novel setups, including a claim that IPOs for SpaceX, OpenAI and Anthropic could together exceed the value of every tech IPO since 1980.

Why it's interesting: It hints at agents moving from assistants to autonomous operators in high-stakes finance.

Context: A two-tier model stack keeps costs down by using an expensive model only for planning and review, and cheaper models for the bulk of the work. Use it whenever a task can be split into clear, independent parts.

Prompt: You are the lead coordinator. First, break this task into a clear plan with independent subtasks I can hand to cheaper models running in parallel. For each subtask, write a self-contained instruction, the exact output format required, and success criteria. After the subtasks return, review all outputs, remove redundancy, flag weak or contradictory results, and produce a single synthesised final answer. Task: [describe your task].

Example use case: Migrating a large codebase: the strong model designs the architecture and splits the files, cheaper models handle individual modules, then the strong model reviews and merges.

A study of more than 69,000 medical preprints found em-dash use jumped from about 4% to nearly 20% after ChatGPT launched. Researchers treat it as a population-level fingerprint of AI-assisted writing rather than proof any single paper was machine-written.

Why it's important: It shows how AI style choices are quietly seeping into scholarly communication and shaping how people judge authenticity.

Source: arXiv

A survey of over 4,000 people aged 11 to 24 in England found nearly half trust AI more than a human to verify information. While 87% had used AI, fewer than 10% felt schools were preparing them for an AI-driven future. The report urges pilots in 100 schools and clearer government guidance.

Why it's important: Education risks falling behind as students lean on AI for learning and truth-checking faster than institutions can adapt.

Source: Public First

Andreessen Horowitz unveiled a one-year programme for recent high school graduates that replaces degrees with startup projects, co-ops at major tech firms, and instruction from industry figures. A founding class of about 50 is planned, with each participant getting $50K in compute credits and a $5K travel allowance.

Why it's important: It reflects a wider push to rethink higher education as AI reshapes what skills matter.

That's it for today's Daily Pulse. Forward this to one person who wants to stay ahead of AI. See you in the next one. - Nicolas