OpenAI held DevDay 2026 and effectively rewrote the rules of the game: Dots agents work constantly and get their own cloud computer, the mid-tier model GPT-6.1 Sol is almost on par with the flagship in agentic tasks but costs five times less, Sign in with ChatGPT turns a subscription into a consumable compute limit for third-party apps, and Ultrafast sells high speed as a separate tier. The analysis of the announcements on October 1 was published by the channel 'Quest Theory of Casts and Roles'. The course is clear: ChatGPT is becoming a platform for resident agents and a 'universal AI wallet', not just a chatbot.

image
image

What happened

The OpenAI DevDay 2026 conference took place on September 29 in San Francisco; on October 1, the author of the channel 'Quest Theory of Casts and Roles' released an analysis of its 14 announcements. The main product novelty — Dots: always-on agents on the GPT-6 Astra model, each agency has its own cloud computer and access to more than 4000 applications via plugins. Dots are available on Pro and Business Premium tiers, with the first dot already included in the plan, autonomy is regulated by permission levels, and for Enterprise, Edu, and Healthcare a beta is available via an administrator. The second key announcement — GPT-6.1 Sol: a new mid-tier model, 'almost Astra' in agentic coding and computer use, but five times cheaper — $2 per 1M input tokens, $0.10 per 1M cached input (95% discount) and $10 per 1M output; Sol is currently only included in ChatGPT Work and Codex, not in regular chat. The third block — Ultrafast, a premium speed tier for GPT-6 Astra with output up to 300 tokens per second: up to 8x speedup in Codex and up to 6x via API; it is only available on the new Pro 500 tier at $500 per month with 25x usage volume compared to Plus, as well as in Enterprise. Separately, Sign in with ChatGPT was launched with 16 starting partners, including Devin, Notion, and Vercel. Developers also got Agents API with Computer Use, an updated Codex CLI with /agents command and worktrees, and Code Review mode.

Context

All three blocks — agents, cheap model, and subscription wallet — come together in one structure: OpenAI is moving ChatGPT from a chatbot to a platform for resident agents and a 'universal AI wallet'. Sign in with ChatGPT is modeled on Sign in with Google, but instead of logging into sites, third-party apps spend the user's subscription limits as currency; this is a new model of AI compute distribution and a direct challenge to Google and Meta. The economics of long-lived agents in this structure is determined not by model size, but by the price of cached context: $0.10 per 1M cached input tokens in 24/7 Dots cycles means that the cost of re-read context drops by an order of magnitude. Quality is currently the only point with an external benchmark: according to independent aggregator Artificial Analysis, GPT-6.1 Sol scores 52 on the Intelligence Index versus 53 for Astra at approximately 22% of the cost per task. If the mid-tier model has truly almost caught up with the flagship in agentic tasks, this is an argument in favor of distillation and inference optimization as a source of progress, not just model scaling.

Why this matters for the industry

For the industry, the main shift is distribution and unit economics, not new model capabilities. Sign in with ChatGPT allows building products on user subscription limits without an API key and without their own billing, so competition shifts to fighting for a place 'in the user's wallet': the ecosystem of third-party apps spending subscriptions is expected to grow beyond the starting partner list, and this is already a new monetization point for SaaS. GPT-6.1 Sol makes 24/7 agents unit-economically real: re-read context at cache price turns constant agents from an exotic feature into a line item that can be calculated and recalculated. If Artificial Analysis estimates are confirmed by independent benchmarks, competitors will have to respond with cache discounts and agentic tiers, 'almost flagship for a fifth of the price' may become the market norm for mid-tier models, and agent TCO recalculation — standard practice; followed by likely task routing between an expensive fast model and a cheap cacheable one as an engineering standard. For startups, a window of cheap agentic operation opens: a window where background agentic features should be recalculated on Sol and cache-aware prompts.

Why this matters for users

For users, the main thing is available immediately, not 'next quarter'. On Pro and Business Premium tiers, you can set up your first Dot and try the 'goal instead of task' pattern: the agent is given a goal, not a one-time command, and autonomy boundaries are set by permission levels — what it does on its own, what it approves, and what it never does; piloting a resident agent is currently more reasonable on non-critical tasks. Developers with access to ChatGPT Work and Codex should move agentic coding to GPT-6.1 Sol and record the economics before and after on their own set of tasks: with cache-aware prompts, long agentic sessions turn out to be several times cheaper. The subscription starts working as a common wallet: its limits can be spent in third-party apps from the partner list without setting up separate API keys. Speed is a separate purchase: Ultrafast is only sold as part of Pro 500 or via an Enterprise contract. Those who communicate with ChatGPT outside of work tools, Sol is currently unavailable.

What is still unknown / limitations

The 'platform shift' formula is an interpretation on top of the same pool of secondary sources: increased confidence in the wording does not add independent validation. The 52 vs 53 score — a one-point difference in an aggregated index without a named set of tasks, number of runs, and context settings, so 'almost Astra' remains an unconfirmed statement, and one source is not validation; independent measurements on open agentic setups are needed. All key prices — from the 95% discount on cached input to $500 for Pro 500 — are stated values from secondary sources dev.to and emergent.sh, not confirmed measurements. Sol is currently absent from regular chat and, according to expectations from the analysis, has not yet reached the API, meaning not all pipelines can verify it. Initial reviews of Dots are negative — according to the author of the analysis, agents 'stumble' and activate at the wrong time, so the problem of activation timing and autonomy boundaries still needs to be designed; this will determine whether the focus in competition will shift from the quality of a single model to the quality of orchestration and the price of cached context.

Sources

Author

Look at AI, editorial team