The week ending
16 August 2026.

Claude and the Riemann zeta function

• Anthropic: Learning more about Claude's mathematical capabilities — published 10 August and updated 13 August; the company's account of the result and its validation.
• Mathematical paper — dated 11 August; raises a lower bound concerning simple zeros on the critical line, but does not prove the Riemann hypothesis.
• Anthropic: zeta-23 Lean files — the formal artefact available for independent inspection. Examination by named mathematicians is not the same as formal endorsement or wider peer review.

Persistence, latency and price

• xAI: Grok 4.6 — announced 12 August through coding and developer channels, with vendor-reported capabilities and pricing.
• Google: Introducing Gemini 3.7 Flash — launched 13 August with temporary introductory pricing.
• OpenAI: Previewing GPT-5.6 Sol Ultrafast — a limited Cerebras-backed preview announced 13 August; its maximum speed figures are OpenAI's claims.

These products, prices and benchmark claims were not measured under one shared test. They support no clean cross-company ranking.

Muse Glimmer and local open weights

• Meta: Muse Glimmer 30B model card — the primary description of the model, Apache 2.0 artefacts, context length and memory targets.

The verified description is open weights under a permissive licence, not a claim that every part of training is open source. The stated memory requirements place uncompressed local use well above an ordinary laptop, while compressed-quality claims still come from Meta's own evaluations.

Gemini at mass-market scale

• Google: Gemini reaches one billion monthly users — published 11 August; usage and image-generation totals are unaudited company figures.
• Google: new connected applications and services — announced 12 August, with rollout planned over subsequent weeks rather than complete availability on announcement day.

Advertising inside ChatGPT

• OpenAI: Testing ads in ChatGPT — updated 11 August to add the United Kingdom, Mexico, Brazil, Japan and South Korea to an experiment that began in the United States on 9 February.

Eligibility, advertisement matching, aggregate reporting, answer independence and privacy are described here as OpenAI's published design and policy commitments, not as independent findings.

AMIE's simulated video consultations

• Google Research: advancing AMIE toward audio-visual clinical consultations — published 11 August; Google's account of the study and results.
• AMIE study paper — first posted 10 August; reports a randomized simulated examination involving trained patient actors, consulting physicians and physician evaluators.

The study used actors rather than real patients, covered only conditions the setup could portray and reported occasional perception, reasoning and technical errors. AMIE remains a research system.

GLM-5.3 and the dual-use release decision

• Z.ai: GLM-5.3 launch — hosted access launched 14 August; the company said public weights would follow after two weeks of further safety evaluation and hardening.
• Z.ai: coordinated vulnerability-disclosure ledger — the company's record of reported vulnerabilities, most of which remained under embargo when checked.

The vulnerability totals, capability claims and reason for delaying the weights are company-reported and largely cannot be inspected independently while disclosures remain embargoed. The planned release date is an intention, not a guarantee.