OpenAI and Anthropic public-post watch

Checked 17 August 2026 at approximately 3:05 am AEST (5:05 pm UTC on 16 August). Priority window: the preceding 24 hours. Items covered in the 16 August bulletin were excluded.

Headlines

• OpenAI’s Tibo Sottiaux argues that cost per token is a misleading cross-model comparison — Sottiaux noted that tokenizers can split identical text differently, so a cheaper token need not mean a cheaper completed task. In one small sample spanning English, technical, multilingual, and numerical text, he said GPT-5.6 Sol’s tokenizer used 766 tokens versus an estimated 1,170 for Claude Opus 5, or about 34.5% fewer. His more useful conclusion is to compare price per successful outcome on your own workloads. Treat the sample as an OpenAI product lead’s illustrative comparison, not a general benchmark proving that Sol is always cheaper. Original post

• Dario Amodei rejects a simple regulation-versus-decentralization trade-off — In a rare two-part thread, Anthropic’s CEO argued that carefully designed rules can slow frontier labs while exempting or advantaging smaller challengers. He cited California’s SB 53, differentiated frontier-model testing, and Anthropic’s preferred interpretation of the “Pacing the Frontier” proposal. Amodei said he supports the US administration’s reported direction of pre-deployment testing for frontier models, extending testing to open-weight models as they approach the frontier, but explicitly reserved judgment until he sees the details. His stated target is rules that address cyber, biological, and alignment risks, constrain frontier-company power, and still leave room for open weights. Part one

• Anthropic is increasing its biology and medicine effort, but Amodei says proof must come before promotion — In the second half of the thread, Amodei repeated his personal forecast that AI could help cure most human disease in roughly five to ten years and said Anthropic is ramping up work in biology and medicine, hoping for “early glimmers” in the coming months and stronger results over the coming years. He also called the industry’s failure to deliver its promised public benefits a valid criticism and argued that trust will come from concrete achievements rather than marketing. These are ambitions and forecasts, not announced medical breakthroughs or a product roadmap. Part two

Other activity

• No additional high-signal OpenAI research or product launch surfaced in the window — Greg Brockman briefly endorsed data showing a large gap between the highest-spending AI adopters and the median company, while Sottiaux continued promoting /goal and Sol-managed fleets of Luna agents. The fleet capability was covered in yesterday’s bulletin and is not repeated as a new launch.

• Anthropic’s Thariq Shihipar shared a watermarking explainer — Shihipar posted a Claude-built interactive artifact intended to explain how watermarking can work without reducing output quality. Because Anthropic’s content-marking rollout and its EU AI Act context were covered earlier this week, this is best read as educational follow-up rather than a new policy or product change. Original post

Steve Yegge watch

• No fresh visible main-feed post in the priority window — Steve Yegge’s live public profile still showed 11 August as its newest main-feed post, so no older item was recycled.

Coverage note

Live public timelines were reviewed across representative OpenAI leadership, product, and research staff including Sam Altman, Greg Brockman, Tibo Sottiaux, Noam Brown, and Andrew Ambrosino, plus the recent multi-agent thread from Eric Provencher. Anthropic checks included Dario Amodei, Boris Cherny, and Thariq Shihipar, with recent policy, product, and research posts cross-checked against the prior bulletin. Steve Yegge’s public profile was checked separately. Previously covered multi-model delegation, ChatGPT feature drops, Ultrafast mode, Claude content marking, EU AI Act guidance, GPT-5.6-Cyber and Daybreak access, usage resets, Claude Code diagnostics, prompt-injection benchmarks, and auto mode were not repeated.