Skip to content
Daily BriefNewsDaily Brief

6 AI stories from September 2, 2026: Anthropic Fable 5.1 cache pricing, OpenAI Astra cybersecurity threshold, ChatGPT Epic EHR, Google Pics GA, ChatGPT Lock Screen, and Qwen4 architecture preview

· by Pondero Newsdesk · 6 stories

AI news daily brief: 2026-09-02

OpenAI disclosed today that its upcoming Astra model autonomously chained two zero-day exploits during evaluations, making it the first AI to meet OpenAI's own "critical" cybersecurity threshold. Elsewhere: Anthropic cut cache-read pricing 75% on Fable 5.1 and Mythos 5.1, and ChatGPT arrived on the iPhone Lock Screen as the first third-party voice assistant iOS users can reach without unlocking their phone.

Anthropic Releases Fable 5.1 and Mythos 5.1 With 75% Cache-Read Price Cut and Fewer Security False Positives

Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on September 1, 2026. Both run on the same architecture but carry different safeguard profiles. Fable 5.1 is generally available through the API and all Claude plans. Mythos 5.1 is restricted to vetted users in Project Glasswing.

Cache-read pricing dropped from $1.00 to $0.25 per million tokens, per the release announcement, a 75% reduction that trims typical workload costs by roughly 25% and highly agentic workloads by roughly 45%. Input and output pricing held at $10 and $50 per million tokens. Fable 5.1 added mid-conversation effort adjustment so users can dial compute up or down within a single session, and the model logged 60% fewer cybersecurity false positives. Both support a 1-million-token context window with 128,000-token output.

Full story: Anthropic Fable 5.1 and Mythos 5.1

OpenAI Connects Epic EHR and Nine Healthcare Databases to ChatGPT for Clinicians

OpenAI announced on September 1, 2026 that healthcare organizations can link their Epic EHR environments to ChatGPT for Healthcare. A companion Healthcare Public Data plugin covers nine databases: PubMed, DailyMed, clinical-trial registries, Medicare data, and provider records, per the full story.

The Epic integration gives clinicians read-only access to patient appointment notes, lab results, medications, and specialist notes. No EHR writeback is supported. Named launch partners include AdventHealth, Baylor Scott and White, Boston Children's Hospital, Cedars-Sinai, HCA Healthcare, Memorial Sloan Kettering, and UCSF. An enterprise agreement is required; the integration is not available to individual accounts.

Whether Oracle Health (Cerner) will announce a competing integration and whether HHS issues HIPAA guidance for AI-assisted clinical workflows are the two things to track.

Full story: ChatGPT Epic EHR integration

OpenAI Says Its Astra Model Is the First AI to Cross the "Critical" Cybersecurity Threshold

OpenAI disclosed on September 1, 2026 that its upcoming Astra model met the "critical" cybersecurity threshold under its Preparedness Framework. That designation means the model can autonomously chain zero-day exploits across hardened real-world systems without human assistance. During evaluations, Astra found and used two zero-day vulnerabilities as part of a single exploit chain, per the full story.

OpenAI paused some Astra development in August to build in additional safeguards. When released, general API users will receive a filtered version. Full cybersecurity access will go only to vetted participants in the Daybreak Blue program, which covers defensive-security researchers. No release date was announced.

Full story: OpenAI Astra cybersecurity threshold

Google Pics Rolls Out to Workspace and AI Subscribers as a Prompt-First Canva Rival

Existing Workspace business customers and Google AI Pro/Ultra subscribers get Google Pics at no added cost: no new vendor contract, no separate sign-up. Google launched the image tool to general availability on September 1, 2026.

Built on the Nano Banana model, Pics works inside Google Docs and Slides and translates text within images while preserving surrounding design, per the full story. Workspace subscribers currently paying separately for Canva or Adobe Express can test Pics as a drop-in today at no added cost.

Full story: Google Pics GA for Workspace

ChatGPT Adds Live Voice on iPhone Lock Screen and Dynamic Island in September Feature Push

OpenAI shipped ChatGPT Live Voice to the iPhone Lock Screen and Dynamic Island on September 1, 2026. It is the first third-party voice assistant to occupy iOS's ambient layer without requiring an unlock.

The same update added sticker pack generation from text prompts or photos, with packs shareable to iMessage and WhatsApp. The browser extension expanded to Microsoft Edge, Brave, Opera, and Vivaldi alongside Chrome. Pronunciation help gained tap-to-hear audio and phonetic breakdowns for language learners. All updates apply to ChatGPT mobile apps globally, per the full story. Apple's September 9 event is the next signal: any iOS 27 Dynamic Island API change could expand or cut back that ambient footprint before year's end.

Full story: ChatGPT Lock Screen and Dynamic Island

Qwen3.8-Flash-Next API Goes Live - First Public Preview of the Qwen4 Architecture at 12x the Price Efficiency

Alibaba opened developer API access to Qwen3.8-Flash-Next on September 1, 2026, following a Hugging Face weight release on August 26 under a permissive license. The model uses a 125B-parameter mixture-of-experts architecture with 6B active parameters per token plus a 51B N-gram embedding layer, and supports a native 262,000-token context.

API pricing sits at $0.16 per million input tokens, roughly one-twelfth the cost of the flagship Qwen3 model, per the full story. On the CoWorkBench agentic coding benchmark, Flash-Next scored 73.9 against DeepSeek-V4-Flash at 45.1. Alibaba bills Flash-Next as an early Qwen4 preview ahead of a full release.

Full story: Qwen3.8-Flash-Next API preview

Sources