OpenAI published AI-generated full or partial solutions to more than 370 outstanding math problems, sparking debate among mathematicians over whether the approach helps or harms the discipline. OpenAI said its unreleased internal model used about three hours of computing time per solution on average. The field is shifting toward “Math 2.0,” with calls for changed publication and research norms and more human-led scrutiny and community building rather than solution-only progress.
Simon Willison’s Weblog·1 week ago·
4
● 16 sources
Anthropic released Claude Haiku 5.5 as a faster, lower-cost model and adjusted its pricing and tokenizer compared with Haiku 4.5. It is priced at $0.10 per million input tokens and $0.50 per million output tokens up to 100,000 tokens, then rises 5x beyond that. Pricing becomes similar to GPT-6 Luna only within 100,000 tokens, while long prompts cost more due to a less generous tokenizer, and Anthropic also halved cache-read prices and added monthly API credits for subscribers (with non-rollover credits).
Meta announced it took action against 33.2 million pieces of child sexual exploitation content on Facebook and Instagram in the first half of 2026 and is rolling out new AI tools aimed at detecting ads and accounts that may secretly direct users to illegal child abuse material elsewhere online. It said more than 97% of the content it acted on was found by its systems before users reported it. As a result, Meta will use new AI to detect ad “signposting,” block destinations and accounts, expand scanning for missed content, run red-teaming on its defenses, and improve detection of users returning with new accounts.
OpenAI launched GPT-6.1 Sol and it was tested against GPT-6 Astra on three identical agentic coding and log/decoder tasks. GPT-6.1 Sol completed 15 runs in $2.66 total versus $14.77 for Astra, which is 18% of the cost. The tests found the models tied on accuracy across all runs, while GPT-6.1 Sol was faster on the longer incident logs and resolver tasks.
Mistral released Le Chonk, a new freely available AI model it says can compete with top US and China models. It has 1 trillion parameters and is in preview now, with a final version due by the end of the month. The model’s broad public availability and niche optimization for coding, cyberdefense, and specific industry tasks change how developers can access and customize advanced AI.
OpenAI released mathematical results generated by an unreleased frontier model covering work across number theory, complexity theory, and mathematical physics. It produced 722 manuscripts in 372 families, with an average result taking the equivalent of three hours of ChatGPT Pro thinking. The article argues this points toward a split between “machine mathematics” that AI systems can verify and a smaller “human mathematics” subset humans can understand and use.
Simo, a “System 1.5” judgment model from Temprl Labs, was listed as an AI infrastructure tooling launch for asking typed questions using text, screenshots, or video. The model is described as System 1.5 and returns calibrated probabilities needed to act in milliseconds. The listing positions Simo as a new developer tool, but it does not report any measurable performance change beyond the stated response-time claim.
Falcon-ASR was released as a 1.6B-parameter speech recognition model built by TII, focused on transcribing Arabic—especially the Emirati dialect—and also supporting English plus French, Spanish, and Portuguese. It reported an average Arabic word error rate of 20.92% across six test sets and achieved 22.73% WER with 10.19% CER on an internal Emirati evaluation. It adds word-level timestamps and is offered via a Hugging Face demo with planned API and native applications.
Aleph Alpha and Mistral each released new AI models, fueling expectations that Europe has more credible sovereign AI contenders. One figure mentioned in the piece is €794k. The story frames these releases as helping shift Europe toward a more competitive AI landscape with multiple serious players.
Anthropic launched Claude Haiku 5.5, its fastest and most capable small model for high-volume, cost-sensitive tasks. The release is effective today. It shifts Anthropic’s small-model lineup toward faster deployment for summarization, classification, coding subagents, customer support, and browser use, but provides no specific benchmarks in the text.
OpenAI released hundreds of new math findings across areas like algebra and topology using an internal model that was not made public. The release covered 377 math problems. The debate shifts to whether the system is genuinely creative or mainly completes proofs by reusing human-derived ideas.
OpenAI published 722 math manuscripts from an unreleased internal Navier-Stokes-based model, claiming solutions to 90 of the top 500 open math problems. The reports cite about 3 hours of ChatGPT Pro thinking compute per result on average. As a result, the release adds new paper files and proof artifacts publicly, with ongoing debate about what is verifiable and what may contain errors.
Every AI story that matters,
in your inbox by 8am.
TLDRocket reads all relevant sources, removes duplicate coverage, and summarises the
day in two minutes. Follow companies and topics for alerts, or get the
briefing in Slack. Free, no spam, unsubscribe anytime.
Reading TLDRocket needs no cookies, and the readership counts we rely on come from
our own cookieless analytics. Google Analytics is the exception: it sets cookies and
reports to Google, so it stays switched off until you allow it. You can change your
mind any time from “Cookie settings” in the footer.
Strictly necessary
Session security and form protection (tldrocket-session,
XSRF-TOKEN, 2 hours). The site cannot work without them,
so they need no consent.
Always on
Google Analytics 4 (_ga,
_ga_<id>, up to 2 years). Measures which
stories and sections readers use. Google acts as a third-party processor and may
store the data outside the EU. No advertising, no profiling, no data sold.
