On October 1, 2026, OpenAI unveiled Dots, a task-running AI agent powered by its new GPT-6 Astra model, at its DevDay conference in San Francisco. Pro and Business subscribers get access on October 15, Plus follows in November, and the Agent API is priced at $8 per million input tokens. It is the first time OpenAI has
DevDay 2026 opened on the morning of October 1 in San Francisco, and OpenAI chief executive Sam Altman used the first ten minutes to ship two things at once: a new foundation model, GPT-6 Astra, and an agent called Dots.
Astra carries a one-million-token context window, accepts text, images, audio and video natively, and was trained with reinforcement learning over large-scale tool-use traces. OpenAI's numbers: 82.4% on SWE-bench Verified, up from 74.1% for GPT-5.2, and roughly a 60% reduction in per-token inference cost.
Dots runs inside an isolated cloud sandbox with browser, terminal and filesystem access. A single task can run for up to eight hours, sub-tasks can be
OpenAI moved agents to center stage because Meta's Muse produced a growth curve it did not anticipate. Muse launched in April 2026 inside WhatsApp, Instagram and Messenger, free to use with no subscription, and reached roughly 400 million weekly users within six months. Meta does not charge for the agent; it extends conversation time, which feeds advertising and business messaging revenue.
Meanwhile the paid ceiling for chatbots has become visible. ChatGPT has around 800 million weekly users but has been stuck near 20 million paying subscribers, and ent
Liu Gong believes Dots is not really competing with Muse on capability. It is competing with free.
From a technical standpoint, the value of a long-horizon agent depends on its failure rate, not its peak score. Dots landing at 72% on OSWorld means one task in four still needs a human to take over; Meta's Muse sits at 61% on the same evaluation, an 11-point gap that leaves both well short of the unattended reliability enterprises require. Errors compound across long chains: a 20-step workflow with 95% success per step completes end to end only about 36% of the time. That arithmetic explains the eight-hour cap, the checkpointing and the replay tooling. A million-token context and cross-session memory solve forgetting, not misclicking, and sandbox design plus recovery speed will shape the user experience more than raw reasoning gains.
On the business side, the two companies are not selling the same product. Meta trades Muse for engagement and covers marginal cost with advertising, which lets it give the agent away to three billion monthly users. OpenAI has to convert agents into subscription and API revenue. Forty tasks a month on Plus is a deliberately calibrated number: about 1.3 tasks a day, enough to demonstrate value, not enough to replace the $200 Pro tier. That playbook worked once in coding, where Claude Code and Cursor proved developers pay for saved time. There is no comparable evidence that consumers will pay an extra $20 a month to have PDFs filed and notes synced. If Dots cannot turn task volume into enterprise contracts, OpenAI bleeds inside a price war it started.
For the industry, Dots shifts competition from model scores to the task entry point. Whoever owns the surface where a user says handle this for me owns the next distribution layer. Meta's advantage is that the surface already exists; OpenAI's is task reliability plus a developer ecosystem, with the Apps SDK turning third-party software into callable actions and AgentKit turning evaluation into repeatable engineering. That squeezes the middleware tier: once a user can say enter this invoice into the system, subscription form tools and light RPA lose pricing power. Liu Gong expects the casualty over the next year to be neither Meta nor OpenAI but single-point automation vendors with no model and no distribution of their own.
- Long-horizon tasks: Dots scores 72% on OSWorld, Meta Muse 61%, and GPT-5.2 Agent Mode 54% on the same evaluation, an 18-point gain over the previous generation.
- Coding benchmark: 82.4% on SWE-bench Verified, up 8.3 points from GPT-5.2, and still below the 85.6% Anthropic reported for Claude 5 Opus.
- Cost: per-token inference cost down about 60% versus GPT-5.2; the Agent API lists at $8 per million input tokens and $30 per million output tokens, roughly 40% of the previous generation's price.
- Distribution and monetization: Muse reaches roughly thre
For the industry, within three months procurement teams will start writing agent task-completion rates into RFPs, and Anthropic, Google and the Chinese platform players are expected to answer with their own long-horizon launches in the fourth quarter. Over a year, the market likely splits into two tracks: free, embedded assistants from Meta and Google, and metered productivity agents from OpenAI and Anthropic. The work being displaced is not whole jobs but hours inside jobs, covering form entry, cross-system copying and first-pass analysis.
For ordinary
Meta's response was restrained. Its AI lead wrote on Threads that Muse would not change its pricing because a rival shipped a new model, noting that WhatsApp carries hundreds of millions of business conversations a day, a corpus and context OpenAI cannot easily match.
Google DeepMind researchers focused on reliability, publicly questioning how far OSWorld and similar benchmarks track real office work and arguing that a 72% demo number usually discounts by a third under enterprise audit. Analysts at SemiAnalysis noted that passing the full 60% cost reduc
This item comes from The Verge AI (published 2026-10-01 14:36:50). Wenfeng Radar provides a Chinese/English summary and commentary on overseas AI developments; the original content is owned by its author.
Daily aggregation of overseas AI developments with in-depth analysis. Bookmark us to never miss an important signal; back to homepage.