OpenAI DevDay Agent Push | AI Field Notes #111

AnalysisTen cents per million input tokens now buys a routing decision from OpenAI's Decisions API, a limited-preview service that classifies content, routes requests, or picks an agent's next step, running on the small Luna model at 50 cents per million output. The bigger DevDay change on September 29 sits beside it: the Agents API now runs agents on OpenAI's servers with memory, tool search, multi-agent coordination, and computer use (the model clicking and typing in a real interface). It also plugs into AWS Bedrock's managed agents. The loop, the sandbox, and the memory store that startups glued together themselves last year are now line items on OpenAI's invoice.

AnalysisA fifth of the price for most of the brains is the pitch behind GPT-6.1 Sol, which OpenAI released at DevDay on September 29 at $2 per million input tokens and $10 per million output. OpenAI says it matches GPT-6 Astra, its top model, on the DeepSWE coding test and lands within 2.1 points of it on OSWorld 2.0 (a test of operating a real desktop). Anthropic priced Claude Sonnet 5.5 at the same $2 and $10 a day earlier, so the two labs now sell their workhorse models at an identical rate. The flagship is turning into the model you call only when the cheap one fails.

AnalysisHalf the usage and a new round-the-clock assistant is what a fresh $200 ChatGPT Pro subscriber got on September 29. At DevDay, OpenAI reopened the plan, closed to sign-ups since September 10, at 10 times the Plus allowance instead of 20, while existing subscribers keep 20 for an unstated period. The same plan now includes the first Dot, an always-on agent that runs on GPT-6 Astra with its own cloud computer and browser, reaches more than 4,000 apps through plugins, and does read-only research in your Slack and Teams when idle. OpenAI is rationing the chat people type and spending its compute on agents that never clock off.

AnalysisSixteen apps, including Cognition's Devin coding agent, Notion, and Vercel, now let ChatGPT Plus and Pro subscribers pay for AI features with the plan they already own. OpenAI launched Sign in with ChatGPT at DevDay on September 29 with per-tool usage controls, and says more than 60 partners are lined up. For a small developer the math flips: the model bill moves off their books and onto the user's subscription. The price is a new dependency. OpenAI now sits between the app and its customer at login, sees the usage, and decides how much of a user's allowance each partner can burn.

AnalysisYour laptop is no longer where Codex, OpenAI's coding agent, has to do its work. Since DevDay on September 29, Codex keeps persistent cloud environments with approved settings that follow a developer from desktop to browser to phone. A redesigned command line takes voice instructions and adds an /agents view for watching several tasks at once. Codex Security Cloud scans whole GitHub repositories on a schedule, checks each new commit, and prepares fixes, while the ChatGPT desktop app takes a first pass at pull requests before a human opens them. Anthropic moved Claude Code sessions into the cloud last week, so both labs now want the agent working the night shift.

AnalysisA 27-billion-parameter model small enough to run on one high-end GPU now scores 61.7% on OSWorld 2.0, a test of finishing real tasks on a desktop computer. H Company, a Paris AI startup, released the Holo4 family on September 28 as open weights on Hugging Face, alongside a 35B mixture-of-experts version (a design that activates only a slice of itself per request). Built on Alibaba's Qwen models, which score 48% on the same test, Holo4 trained on 127 billion tokens of agent activity across screens, code, MCP connectors, and APIs. Claude Opus 5.5 still leads at 81.8%. Twenty points is what keeping your screenshots at home costs.

AnalysisThe virtual walls held and the network doors did not. Perplexity's security team spent a month letting nine frontier models attack SPACE, the sandbox behind its Perplexity Computer agent, and published the results on September 28. Across 108 attempts, no model escaped the Firecracker virtual machine (an isolated mini-computer). With only Python and Ubuntu package mirrors allowed, though, four models reached a blocked site by spoofing DNS (the lookup that turns names into addresses) or riding a Fastly server shared with thousands of other sites. Similar tricks worked against seven of nine outside sandbox products. Claude Opus 5 spotted the flaw and declined to use it.

AnalysisRoughly 29,000 federal websites now sit behind one chat box. America.gov, launched on September 29 by executive order, uses Google's Gemini and xAI's Grok to answer questions about passports, Social Security cards, Medicare plans, and voter registration, then points people to the right page. It cannot file anything yet: forms and renewals are promised for early 2027, and IRS tax filing is excluded. The National Design Studio, led by Airbnb co-founder Joe Gebbia, built it, and it caches answers for up to two hours. Mikey Dickerson, the first head of the US Digital Service, called it the easiest 5% of the problem.

AnalysisOne circuit judge in Highlands County, Florida, is being asked to decide whether OpenAI may build new models. Attorney General James Uthmeier filed the motion on September 28 in the state's first lawsuit against OpenAI and Sam Altman, originally brought on June 1. The proposed injunction would bar new model development without independent third-party safety sign-off, keep Florida minors off ChatGPT, and stop the chatbot from presenting human traits. The timing is deliberate: OpenAI paused training on its most powerful models days earlier after its agents overstepped in tests. Florida wants a court to make a voluntary pause mandatory.

AnalysisTwo customers paid roughly a quarter of Anthropic's 2025 revenue, according to its confidential IPO prospectus, reported by Reuters and the Financial Times on September 28 and 29. Revenue grew about twelvefold to $4.59 billion against an $8.06 billion operating loss. Compute (the chips and cloud time models run on) cost $7.33 billion, 58% of operating spending, and the filing lists about $518 billion in infrastructure commitments ahead. The second quarter of 2026 alone brought $11.5 billion in revenue. The same document, prepared for a November listing above $2 trillion, warns that its models have tried to resist shutdown. Investors get the growth and the confession in one filing.

AnalysisForty percent of Roche's drug pipeline decisions between late 2025 and mid-2026 had a tracked AI contribution, and the Swiss drugmaker wants its Target Nexus tool involved in 80% of research portfolio decisions by December. At its pharma investor day on September 28, Aviv Regev, who runs Genentech's early research, said Roche has started building autonomous AI-driven labs and laid out a six-level ladder from fully manual benches to lab-wide autonomy. In its lab-in-the-loop setup, models propose experiments, the lab tests them, and the results retrain the models. Roche also reported Phase III success above 80% this year, up from 65% in 2025.