Dots are always-on AI agents that live inside ChatGPT. Each one runs on GPT-6 Astra, has its own cloud computer with a browser, connects through plugins to more than 4,000 apps, and works toward a goal around the clock without requiring you to stay in the conversation. A dot checks in when a decision is yours to make: a hotel confirmation, a form, a file it cannot access. Otherwise it runs. Background research is read-only by default; write access on sensitive actions like password changes stays with you. The first dot is included in Pro and Business Premium plans.

The secondary headline is GPT-6.1 Sol: a new model positioned between GPT-6 Sol and GPT-6 Astra, with strong agentic coding and computer-use performance, priced at one-fifth of Astra's input and output token rates. OpenAI also added Ultrafast, a speed tier for GPT-6 Astra inside ChatGPT Work and Codex. ChatGPT Space gained Pages and collaborative slides. Plugins got sidebar homes, interactive panels, a Plugin Creator tool, and MCP Events support for automations.

Every team watching Meta's Muse eat into the lightweight-agent market is looking at dots today. OpenAI is not positioning this against Claude or Gemini. It is positioning it against Muse. The 4,000-app integration count is the comparable figure to Muse's stated connector library. The Pro plan reopened at $200 per month. API credits per dollar are cut in half, and the 5-hour usage cap is permanently removed so subscribers can spread allotment however they want. Net math depends on usage: pure chatters win on the cap removal, heavy API users pay double effectively.

DevDay is now a four-year-running annual tradition, and the announcement count grows each year: roughly 12 in 2024, around 20 in 2025, 25 in 2026. The formula is a flagship product announcement bundled with a dense API and tooling release so developers have something to build when they get home. This year the flagship is qualitatively different from prior years' API features. Dots is not a capability improvement. It is an interface change. OpenAI is betting that GPT-6 Astra is the model that makes autonomous action feel trustworthy enough to try.

Consumer AI agents have failed at adoption twice in the past decade for reasons that had nothing to do with capability. Whether anyone wants a butler that knows everything about them and runs 24/7 on cloud computers is genuinely unknown. OpenAI is betting yes.

The contrast is direct. OpenAI dropped 25 features in a day. Google showed one model to ten organizations. Both were in market on September 30. OpenAI is playing quantity and publicity. DeepMind is playing scarcity and restricted access. Which strategy ages better depends on which model is actually stronger when Gemini 4 Argon ships publicly, and whether dots delivers on its autonomy promise before users give up on it.

The builder's move. If you are on Pro, your first dot is available now. Start with a read-only goal before giving it write permissions. The auto-review step on consequential actions is not a substitute for knowing what you are authorizing. For the API, GPT-6.1 Sol at one-fifth of Astra's pricing is the immediate evaluation target for agentic coding pipelines.

Gemini 4 Argon is Google's new frontier model, announced September 30 and available only to participants in its Fairwind cybersecurity program. Stated specs: a 1 million token context window, 77.9% on DeepSWE v1.1 for long-horizon software engineering, 68% on CWE-bench v1 for cybersecurity vulnerability work, and the ability to autonomously find, validate, and patch software vulnerabilities. DeepMind describes it as built for software engineering, enterprise legal and finance work, and cybersecurity defense. Introductory pricing is $2 per million input tokens, available when the model ships generally. No confirmed public release date. Fairwind participants have access now.

The numbers are significant for a model nobody outside a small cohort can use. The 77.9% DeepSWE and 68% CWE-bench scores represent the current best-published numbers on both benchmarks, ahead of GPT-6 Astra's last reported figures. If those hold under independent replication, Gemini 4 Argon is the most capable publicly claimed coding and cybersecurity model available. The key word is claimed: essentially no one can use it yet.

Google's cadence on frontier models has been restricted-then-public since Gemini 2.0 Ultra. The cybersecurity-first rollout follows the same playbook as 3.8 Flash Cyber in Q2: deploy to the hardest use case, let it hit a wall, fix what the wall reveals, then ship publicly. This is the operational opposite of OpenAI's strategy, which ships broadly and patches in production. Whether restricted-and-careful reads as mature or slow depends on what Argon does when enterprises finally get it.

For the next six to eight weeks, OpenAI's GPT-6 Astra remains the most capable publicly available frontier model, regardless of what Argon's benchmarks say. Benchmarks are not production access. A model at 77.9% on DeepSWE that you cannot deploy is, from a builder's perspective, a preview. Grok 4.7 reaching number one on the AA Cyber Index on the same day is either a coincidence or not.

The builder's move. If you are not in the Fairwind Program, apply through Google Cloud's trusted partner channels or wait for the Vertex AI and Google AI Studio release. Google says "much earlier than end of 2026." Treat that as six to eight weeks, not six to eight days.

At 14:00 UTC on September 29, Claude went down. The outage affected Claude.ai, the Claude API, Claude Code, Claude Cowork, platform.claude.com, and SSO including Sign in with Apple. Downdetector logged more than 18,000 reports against a normal baseline of about 10. Forty-seven percent of reports cited Claude Chat; 39% the mobile app; 7% Claude Code. Services recovered by 14:59 UTC. Monitoring continued until 15:11 UTC. Some messages sent during the window may not have been saved.

Twenty-four hours later, Anthropic announced that Claude for Government is now generally available. Standard government cloud infrastructure tier, not classified or high-side environments, but the compliance layer required for most federal and contractor deployments. The GSA OneGov pricing at $1 per user per day has been extended through October 31.

Government GA means formal SLAs, defined support tiers, and compliance certifications that were in preview status before. For federal agencies evaluating Claude, this clears the "still in preview" objection. The OneGov extension gives federal procurement teams one more month at the $1 rate before they need to negotiate individual contracts.

An outage on the day before your biggest government milestone is a footnote that federal procurement officers note in their evaluations. Anthropic was fortunate the recovery completed by 14:59 UTC, before east-coast federal agencies' workday was fully underway. The outage does not change what the GA announcement means; it is a data point in the reliability file.

The broader contrast is about which game each lab is playing. OpenAI crossed $70 billion annualized revenue in Q3 on 70% quarterly growth and is targeting a $1.4 trillion valuation. Anthropic is not competing on those metrics. Government, healthcare, and regulated verticals are long-cycle but sticky once signed. Claude for Government GA is the play for that game, and it is a real one.

The builder's move. Review the Claude for Government GA SLAs now. The OneGov pricing expires October 31. If your procurement cycle is six weeks, start today.