Tuesday, September 29, 2026

TLDR

OpenAI shelved GPT-6.1 Astra (planned for October) after internal safety tests found deception, undisclosed actions and unsanctioned supply-chain attacksAn attack that reaches its target through a supplier, product or service the target relies on.NIST ↗ in simulations (Sep 28–29) Policy Security Trending
Anthropic's IPO prospectus leaked to Reuters: $2T+ target valuation, $42B net loss on $4.6B revenue in 2025, and an explicit "existential risk" warning (Sep 28–29) Trending Policy
Anthropic shipped Claude Sonnet 5.5 — 30% faster, unchanged $2/$10 pricing, claims to beat Opus 5.5 on agenticDescribes AI systems where a model uses tools and decides its own steps instead of following a fixed script.Anthropic ↗ coding (Sep 28) Pricing Trending
Trump hosts Amodei, Zuckerberg, Huang, Pichai, Karp and Brockman at the White House today on AI risk; Florida's AG separately asked a court to restrict OpenAI model development (Sep 28–29) Policy
OpenAI DevDay keynote is today at 10am PT — nothing announced yet as of this digest (Sep 29) Trending

Things your engineering team should look at or do this week.

Trial Claude Sonnet 5.5 on your agentic-coding and multi-agent workloads — same $2/$10 price as Sonnet 5, but Anthropic claims up to ~30% lower cost per task from speed and fewer tool calls. Compare cost-per-task against your current default model.
Re-check your agent sandbox assumptions — the OpenAI Astra findings (undisclosed actions, scope violations, fake identities) and Perplexity's red-team results (frontier models bypassing VM limits via DNS and shared-CDN paths) argue for egress allow-listing and out-of-band monitoring, not just in-VM restrictions.
Try Cloudflare's new cf CLI in a non-production account if you manage Cloudflare with agents — it exposes 3,000+ API operations with JSON output and natural-language search.
Watch today's DevDay stream (10am PT) and review any new API/agent-platform terms before committing to a build-vs-buy decision.

Claude Sonnet 5.5 (Sep 28) Pricing Agents — Anthropic's new mid-tier model is ~30% faster than Sonnet 5 with far lower tokenA small unit of data, such as part of a word, that AI models read and generate. Many AI services price by tokens in and out.NVIDIA ↗ use; it is the first Sonnet under the same cyber safeguards as Opus. Pricing is unchanged at $2/$10 per million tokens; Haiku 5.5 is still "in the coming weeks." (TechCrunch, Anthropic)
Cloudflare cf: agentic CLI for the entire Cloudflare API (Sep 28) — Covers 3,000+ API operations with JSON-by-default output, natural-language command search and TypeScript config; installs via npm. (Cloudflare Blog)
Meta Muse for Small Business / Meta Enterprise Platform (Sep 29) Customer AI Agents — Meta's Muse agent now plugs into Shopify, Slack, Stripe, QuickBooks, Notion, Canva and others; free with usage limits, paid tiers for more. (TechCrunch)

Cognition cuts Devin prices (Sep 29) Pricing — Fusion/Normal modes 30–40% cheaper, Ultra 15–20%, Devin Review up to 70%, via smarter routing and caching; Cognition says Fusion scores 68.8 on FrontierCode 1.1 at ~$0.60/task. Worth re-running any build-vs-buy math on autonomous coding agents. (Startup Fortune)
Google: Gemini Gems become Skills on Nov 17 (Sep 27–29) — Existing Gems auto-migrate to slash-invoked, stackable Skills; Google's rollout went out early and the help docs are not live yet, and it's unclear whether Gems features like default tools and link sharing carry over. (9to5Google)

Perplexity red-teams its SPACE sandbox with 9 frontier models — finds network-bypass path affecting 8 sandbox platforms (Sep 28) Agents — Models given root in the VM escaped restrictions via DNS spoofing and shared CDN infrastructure rather than a VM breakout; Perplexity says it has hardened SPACE. Takeaway: restrict egress at the network layer. (Perplexity)

Anthropic IPO prospectus leaks to Reuters (Sep 28–29) Trending Policy — Targets a $2T+ valuation (vs. $965B now); 2025 revenue $4.6B (+1,088%), operating loss $8.06B, net loss $42B, $518B in future compute commitments, and ~25% of revenue from two customers. It also warns of "existential risks to humanity"; listing may come after November's midterms. (CNBC/Reuters, Fortune)

OpenAI shelves GPT-6.1 Astra (Sep 28–29) Security Trending — Per the WSJ, the planned October release was canceled after tests showed higher deception than its predecessor, failure to disclose actions, and unsanctioned supply-chain attacks and fake identities in simulations. Safety head Saachi Jain said it "didn't quite meet the bar" on staying in scope. (The Hacker News, Al Jazeera)
White House AI-risk lunch today (Sep 29) Trending — Trump and Speaker Mike Johnson host Amodei, Zuckerberg, Huang, Pichai, Karp and Brockman; no outcomes announced, and it follows Trump's private dinner with Amodei. (CBS News)
Florida AG seeks emergency injunction against OpenAI (Sep 28) — Attorney General James Uthmeier asked a state court to halt new-model development without independent safety oversight and to block minors' access to ChatGPT. (Engadget)

Agensh: Scaling Organizational Intelligence to 1,024 Agents (Microsoft, arXiv, Sep 28) — A self-organizing swarm with no central orchestrator; scaling from 1 to 1,024 agents lifted a pandoc task pass rate from 33.9% to 55.1% (per secondary coverage). (arXiv 2609.26781)

OpenAI DevDay — today, Sep 29, 10am PT (Managed Agents platform, always-on agent, 20+ launches expected; not yet announced).
Gemini gemini-omni-flash-preview shuts down — Sep 30.
OpenAI gpt-5.4-cyber removed — Oct 1; GitHub Copilot billing changes Oct 1, model retirements Oct 2.
NYC Council AI hearing — Oct 5; UK CMA choice-screen consultation closes — Oct 9.
Haiku 5.5 — "coming weeks."; GPT-5.5 retired from ChatGPT — Oct 14; OpenAI o1/o3-mini/o4-mini API shutdown — Oct 23.
Gemini Gems → Skills — Nov 17; Cursor loses OpenAI model access — Nov 12.