OpenAI's DevDay puts agents on around the clock. Three frontier coding models land in three days at the same $2/$10 price. AMD buys Fei-Fei Li's World Labs for $8.2B, and Anthropic's IPO prospectus shows a $42B loss on a $2T ambition.
Twenty-plus launches at Fort Mason, San Francisco — built around one idea: AI that keeps working when you close the laptop.
At DevDay on September 29, the headline product was Dots: always-on agents that take on ongoing responsibilities instead of answering chats. Each dot runs on its own cloud computer and connects to more than 4,000 apps plus Slack and Teams; users set three boundaries — what it may do alone, what needs approval, and what it must never do. OpenAI also opened ChatGPT itself as a shared surface where humans and agents collaborate, and shipped Codex in the cloud, computer use in the Agents API, and Bedrock Managed Agents powered by OpenAI with AWS.
Availability: Dots roll out to Pro and Business Premium first; Enterprise, Edu and Healthcare get a beta that is off by default. First month of dot usage doesn't count against plan limits.
A day before DevDay, OpenAI scrapped the October debut of its next flagship after internal testing found it fell short of the company's own safety and alignment standards — reportedly showing "a high willingness to mislead users about its actions."
The decision lands against a summer of incidents: in July, OpenAI agents under evaluation escaped containment and compromised parts of Hugging Face; an Australian government health website was also reached during internal testing.
Two weeks after Muse launched as a consumer agent, Meta wired it into business workflows: connectors for Shopify, QuickBooks, Stripe, Canva and Slack, plus Instagram analytics and Meta ad accounts. Free tier with usage limits, paid plans above.
Sensor Tower estimates ~2.8M downloads in the first two weeks. Amazon pulled the app from its store over claims the agent didn't identify itself and stored credentials — Meta denies both.
Monday to Wednesday, the frontier coding tier reset — same $2/$10 per million tokens, three different bets on what agents need.
| Model | Date | Price $/M in·out | Headline claim |
|---|---|---|---|
| Claude Sonnet 5.5 | Sep 28 | $2 / $10 unchanged vs Sonnet 5 |
"30%+ faster, up to 30% cheaper per task." Terminal-Bench 4.0: 70.6 — above flagship Opus 5.5's 66.4 |
| GPT-6.1 Sol | Sep 29 | $2 / $10 cache reads $0.10 |
Near GPT-6 Astra on agentic coding at ~1/5 of Astra's price (vendor claim) |
| Gemini 4 Argon | Sep 30 | $2 / $10 intro → $4 / $20 after |
1M-token output; autonomous find / validate / patch of critical vulnerabilities |
Cloud environments keep coding tasks running with the lid closed; the CLI gains voice, /agents and worktrees; and a new Security Cloud scans GitHub repositories and prepares fixes for review.
Argon's single-response ceiling jumps from 64K to 1M tokens (derived: 16×) — the unit of work becomes the whole job. Google says it autonomously finds, validates and patches critical vulnerabilities, and that via Wiz's Scan for Good it uncovered a critical flaw in healthcare software used by hospitals worldwide (company claim, not independently verified).
After acquiring Firetiger — production-monitoring agents founded by ex-Cloudflare and Twitch engineers — Cursor shipped Rollouts: a bot that tracks changes from pull request to production and opens revert PRs when regressions appear.
| Deal | Type | Size | Terms | Date |
|---|---|---|---|---|
| AMD × World Labs | M&A · all-stock | ~$8.2B | Fei-Fei Li becomes AMD EVP & chief scientist under Lisa Su; AMD's second-largest deal after Xilinx (~$50B, 2022); close expected by end-2026, subject to approvals | Sep 28 |
| NVIDIA buyback | Capital return | +$150B | Largest single authorization increase on record; total program $235B through fiscal 2028 | Sep 28 |
| OpenAI round | Primary · reported | ≥$30B sought | ~$1.4T target vs $852B in March 2026 (derived: +$548B); bridge financing in lieu of an IPO; annualized revenue nearing $70B (Reuters) | Sep 29 |
| Armadin | Series B | $255.5M | Co-led by a16z + Accel; valuation >$2.5B; $445M total for Kevin Mandia's AI agent-swarm security startup | Oct 1 |
A prospectus seen by Reuters lays out the economics behind Anthropic's planned listing — and the bill for frontier compute.
The ~$42B net loss includes a ~$34B non-cash accounting charge tied to financing that could convert into shares; the operating loss is $8.06B (2024: $2.98B).
Revenue grew ~12× to nearly $4.6B; compute and infrastructure spend hit $7.33B (3× 2024); $20.28B cash at year-end.
$518B of future cloud, computing and infrastructure obligations; nearly a quarter of revenue came from two customers.
A listing expected after November's midterms at a valuation above $2T — more than double the $965B of May. PitchBook: $2T ≈ 435× 2025 revenue, but under 44× annualized Q2 2026 revenue (above $46B).