Models & Research

Codenamed Watermelon, trained at the Prometheus datacenter in Ohio. On MemoryBench, Meta claims it beats Fable, Opus and Astra.

The commitment includes training and onboarding, down to agencies that still have to get a first Genesis project running.

Mathematicians at Cambridge and King's College London found two places where the prose proof and the Lean code diverge.

The model was handed roughly 4,000 problems. Two results fell outside the standard procedure: a zero-free region for the Riemann zeta function, and the Hodge Conjecture for CM abelian varieties.

Haiku 4.5 scored a flat 0.0% on Terminal-Bench 4.0. The 5.5 gets 39.2 — at a tenth of the price.

Behind the generated interfaces sits a compiler that renders the answer while the model is still writing it. Codex and Work are left out.

Strip out the financial profile and let the models read only an email inbox, and much of the gap survives. Gemini read the two finance emails first in 97% of trials.

The reinforcement learning run produced over 100 million rollouts, on 10,500 GB300 GPUs, across four weeks.

AI Plus subscribers paying $4.99 lose Pro as well. In exchange, the Pro tier picks up Deep Think, which used to start at $99.99 a month.

Since October 1, Google has stopped taking product vulnerability reports. The announcement puts the restart in the first quarter of 2027.

To build the index, Claude scored about 19,000 O*NET task descriptions and hunted for a real robot behind each one. A side finding from 50 years of backtesting: robots gain roughly 2% of previously impossible physical work every year.

Google's new frontier model writes up to a million output tokens, where earlier Gemini models stopped at 64,000. Google's own agents used it to rewrite the libgav1 video decoder in safe Rust: 2.7 times faster, identical output.

Artificial Analysis clocked roughly 193,000 output tokens per task at max effort. That's the heaviest token use they have ever measured, and about seven times what GPT-6 Astra needs.

You'll summon a skill with a forward slash inside a task thread. An interface engineers tend to like and everyone else tends not to.

950 agents spent 21 hours combing through 200,000 reverse transcriptases. One of them wrote, mid-analysis: 'This is spectacular.'

One early tester pushed through a 680,000-line migration in under a day. Cache reads, the biggest line item in any agent bill, drop from 50 to 20 cents per million tokens.

On AutomationBench, Sol at xhigh effort beats Claude Opus 5 at max effort for nine percent of the cost per task. Luna comes in at 10 cents per million input tokens.

TypeSafe calls its yes/no question a Noul. The CEO confirmed on Hacker News that it is short for Bernoulli.

Xiaomi puts the entire reinforcement-learning run at $2.62 million for Pro and $850,000 for Flash, finished in under six days. More than half that money never touched the weights.

One line in the Fable 5.1 system card has gone almost unquoted these past two weeks: Anthropic sees no clear sign of dramatic acceleration.