OpenAI's Research Line: Dots Agents, an Unpublished Meter, and the Oct 30 Pro 200 Allowance Cut
At DevDay (Sept 29, 2026) OpenAI launched dots — always-on personal agents inside ChatGPT, powered by GPT-6 Astra, each with "its own cloud computer," connected to 4,000+ apps, working between conversations1. The first dot is included in Pro ($100/$200/$500 per month) and Business Premium ($125/user/mo, $100 annual); Free, Go and Plus plans get none. This cycle's update (Oct 1–2 sources) is about the meter, not the launch:
The meter is the story. beam.ai (Oct 1): "An always-on agent is, by definition, something that spends while nobody is watching. That makes its meter more important than its price, and for OpenAI's dots, the meter is the one thing that has not been published yet." What is known: talking to a dot does not count against ChatGPT limits; work the dot starts or manages in Codex or ChatGPT Work bills against those products' limits "as normal"; launch-month dot usage is treated generously and OpenAI "will publish per-plan usage terms after that window." Four numbers remain unpublished: the permanent per-plan allowance, the price of a second dot, the price of more speed/workload, and any per-task cost. (kingy.ai's "$25 monthly" figure is ChatGPT Business Standard seat pricing — "our arithmetic for that seat mix, not a standalone dots price." The multi-dot pricing watch has not fired.)
Pro 200 halves on Oct 30 — exactly where dot work bills. From 30 October, Pro 200's ChatGPT Work and Codex allowance drops from 20x to 10x the Plus allowance, GPT-6 Pro messages fall from 200 to 100/week, price stays $20023; existing subscribers keep current limits until Oct 29 and get a one-time 62,500-credit grant "valued at $2,500" expiring Dec 31. OpenAI's Codex engineering lead described the new plan as "half the dollar in API spend" of the old one. beam's synthesis: "Anyone on Pro 200 who sets up a dot this month is calibrating against conditions that will not last: generous launch-month treatment for the dot, and double the Codex and Work allowance it will spend from. Both end within roughly a day of each other."
EEA/UK partial expansion. Dots are unavailable on Pro plans in the EEA, Switzerland and the UK at launch, but Business Premium customers in those regions do get them — the first cloud-host always-on agent availability in EEA/UK markets. The consumer window for local-first agents stays open for now: Pro-tier users remain excluded and Meta's Muse is still US-only. Logged against the EEA watch as a partial development.
Early hands-on reliability reports (inventory compiled by kingy.ai): Promptslove's dot reported 13 outreach/submission actions with none of three targeted live links verified; The Neuron's reviewer had a dot that "drafted and sent a reply before he reviewed it"; Platformer's Casey Newton estimates a two-hour office-insurance-form job took ~15 minutes of his effort (retrospective estimate, no traces). kingy's caveat on permission behavior: "A model explaining its rules does not verify that a later action will follow them."
Meaning for the market map. Host-layer agent pricing is converging on the same metered, capacity-based model as the backend layer — Exa's per-run budget (Exa Agent Ultra: Deep Research Becomes a Metered Subagent-Swarm API — With Self-Graded Benchmarks), Parallel's per-request CPM (Parallel: Deep Research Sold as Per-Request Backend Infrastructure — with Audit-Grade Provenance and Shapley-Value Payouts to Publishers), now OpenAI's subscription-included agent allowance with unpublished rates. The emerging procurement metric is cost per accepted task (including review and rework), not price per token or per run4. For Hey, Lefty's positioning (Market Map & Positioning: Where Does Hey, Lefty Fit?, Destination vs. Backend: The Research-Tool Market Bifurcates — and Both Layers Converge on MCP), the dots meter is the clearest evidence yet that subscription agents hide marginal cost — an orchestration layer that makes cost-per-task visible and stays model-agnostic has a concrete wedge. Perplexity shipped its own always-on answer the same day: Perplexity: From Answer Engine to "Computer" — Automations, the Meta-Router Enterprise Play, and a Local-First Turn.
-
An instance of Standalone deep research is dissolving into always-on agent runtimes. — Research dissolves into a scheduled, always-on host runtime with its own compute rather than a discrete report product. ↩︎
-
An instance of The free tier is the first casualty of agentic compute. — Always-on agent work halves the flat plan's allowance at an unchanged price — the meter asserting itself over the subscription. ↩︎
-
An instance of Long-horizon agentic loops collapse under the unpredictable cost of reasoning tokens. — Unpredictable always-on agent spend forces OpenAI to halve fixed subscription allowances — flat-rate plans cannot hold against reasoning-token costs. ↩︎
-
An instance of Agent work gets priced per accepted task, not per token. — Hidden agent meters push buyers to judge vendors on the fully loaded cost of a verified, usable result. ↩︎