On September 29 in San Francisco, OpenAI held DevDay 2026 and showcased more than 20 new features. The main ones are: resident Dots agents on the GPT-6 Astra model, the ChatGPT Pro 500 tier at $500 per month with premium Ultrafast speed, and Sign in with ChatGPT — logging into third-party apps without an API key, where AI requests are deducted from Plus or Pro subscription limits. The new GPT-6.1 Sol model costs about 22% of Astra's cost per task at nearly equal quality. The essence of the event: the ChatGPT subscription is turning into a payment mechanism for other people's apps, and the price of agentic intelligence is dropping sharply.

image
image
image

What happened

At the OpenAI DevDay 2026 conference, held on September 29 in San Francisco, more than 20 updates were announced. Dots — always-on agents on the GPT-6 Astra model — were opened; each has its own cloud computer and access to more than 4,000 apps, and the first dot is already included in the Pro and Business Premium tiers. The ChatGPT Pro 500 tier at $500 per month was presented: it includes a 25x Plus subscription limit and Ultrafast mode with a speed of up to 300 tokens per second, which works up to 8 times faster in Codex and up to 6 times faster via the API for coding tasks. The GPT-6.1 Sol model launched at a price of $2 per 1M input tokens, $0.10 per 1M cached input, and $10 per 1M output. Sign in with ChatGPT was launched: logging into third-party apps without an API key, where AI requests are deducted from ChatGPT Plus/Pro limits; at launch, 16 partners, including Devin, Notion, and Vercel. Also announced were the Agents API with computer use, tool search, and context compaction, the Decisions API based on the Luna model in limited preview status, Codex in the cloud, and the /agents command in Codex CLI.

Context

To understand the scale, it is worth separating the fact of the conference itself from the shift it marks. Previously, OpenAI's monetization was mainly through API keys and its own products; now a user's subscription becomes a credit quota that third-party services spend through Sign in with ChatGPT, and the list of 16 launch partners is a claim to a login standard. In parallel, internal agent infrastructure is being handed out: the Agents API is built on the same harness as Codex, and Dots is the first product where an agent constantly lives in the cloud with a large context. A separate economic background is cached input at $0.10 per 1M tokens: the economics of agents working around the clock is determined by the cost of repeated context, not peak generation. By methodological measures, this is an engineering-economic release, not a scientific breakthrough: the only verifiable capability figure — the Artificial Analysis index — only confirms that Sol is almost on par with Astra but significantly lags behind the leader Claude Opus 5.5.

Why this matters for the industry

For companies and developers, three layers are changing. The first is unit economics: GPT-6.1 Sol costs about 22% of Astra's cost per task at nearly equal quality according to Artificial Analysis, and cached input at $0.10 per 1M makes long-lived agents and high-frequency classifications cheap; the Decisions API with Luna is exactly about the second scenario. The second layer is distribution and billing: Sign in with ChatGPT gives an app a user's quota without an API key, so some SaaS can give up its own monetization of AI requests and build UX on top of someone else's limit, and competition shifts to the ability to allocate quota. The third layer is infrastructure: the Agents API with computer use, tool search, and context compaction shortens the path to an agent prototype, and the common harness with Codex means a single runtime. The flip side: if an app turns into a UX layer and compute is paid for by a user's subscription, owners of vertical data and deeply integrated workflows win, while thin API wrappers lose.

Why this matters for users

If you have a ChatGPT Plus or Pro subscription, you can already log into third-party apps with your ChatGPT account — at launch, this includes Devin, Notion, Vercel, and 13 other partners. AI requests are deducted from the included subscription limits, no separate API key is needed, and you can assign a weekly spending limit to each app so you don't accidentally spend the entire balance on one service. The ChatGPT Pro 500 tier at $500 per month opens up Ultrafast mode and the first dot — this price is available to few, for the rest it is more of a directional benchmark. A developer on a subscription should try the Sign in with ChatGPT quickstart, Codex in the cloud, and the /agents command in Codex CLI.

What is still unknown / limitations

Much will have to be verified independently. The 1-point difference in the Intelligence Index between GPT-6.1 Sol (52) and Astra (53) according to Artificial Analysis lies within the noise of the aggregated index, and without the task composition, number of repetitions, and confidence intervals, it cannot be considered either a regression or a confirmation of parity; the leader Claude Opus 5.5 scored 58. There are no publications about the Ultrafast mode compromises: the quality of generation at a speed of up to 300 tokens per second, behavior on long contexts, and consistency with normal mode are unknown. Sol is not available in the ChatGPT chat — only Work and Codex, and the hypothesis that it was tuned for agent and coding workloads requires user evals. The Decisions API with Luna remains in limited preview, there is no independent calibration of Dots yet, and the composition of the 16 launch partners for Sign in with ChatGPT may change.

Sources

Author

Look at AI, editorial team