OneHuman Watchdog: September 2026 — Two Labs Shipped Agents Before Shipping Rules

September 2026 is the first month OneHuman tracked nine AI tools instead of eight. Meta's Muse, first covered in BNA28 on September 22, joined the tracker on September 30, one day after OpenAI shipped an agent of its own. Two of the industry's largest labs put autonomous, spending-capable software into the market this month, and neither published a durable rulebook for what it costs to run one. That is the month's real story, ahead of the six OneHuman reports that ran the usual watchdog check on the rest of it: Gemini's expiring cheap-API rate (BNA24), Copilot's nine-day-old defaults (BNA25), Claude Code's fifth usage-limit extension (BNA26), xAI's per-post billing switch (BNA27), Meta Muse's launch (BNA28, corrected below), and OpenAI's tweet-versus-tier mismatch (BNA29).

Updated September 30, 2026: added Anthropic's September model releases.

What Happened (This Month)

Six BNAs published, numbered 24 through 29. Gemini (Sep 2): Gemini 3.7 Flash's launch API pricing doubles January 1, 2027, and Gemini's consumer usage cap is now a multiplier with no stated base. GitHub Copilot (Sep 7): pricing held nine days before six default settings changed under it. Claude (Code) (Sep 12): the predicted fifth extension of a "temporary" usage boost arrived, but the accompanying "permanent 25% increase" turned out to be a 17% cut from the already-boosted level, disclosed only on X. Grok (Sep 17): xAI's X Search API moved from billing per call to billing per post and per profile, effective September 21; a call returning 1,000 profiles goes from $0.005 to $10. Meta Muse (Sep 22): the agent's US launch reportedly outpaced ChatGPT's early download curve, per an Apptopia estimate; see the correction below. ChatGPT (Sep 28): a viral tweet named three models powering ChatGPT Voice's "global rollout," but on Plus none reach the ordinary Chat window. OpenAI's own help center shows Chat access to Astra also extends to Business Standard, not only Business Premium.

The rest of September moved faster than the BNA cadence could cover. In OpenAI's DevDay week alone (Sep 29), the company introduced GPT-6.1 Sol (Work and Codex only, "not yet available in Chat," per its own announcement and TechCrunch), a $500/month Pro 500 tier, a reopened $200 Pro tier with allowances OpenAI says are "changing" with no published figure, and dots: always-on agents built on GPT-6 Astra, each with its own cloud computer, rolling out to Pro and Business Premium users 18 and over, excluding the EEA, Switzerland, and the UK at launch, with usage exempted from plan allowances "for the next month" and nothing published after that. Meta extended Muse to small businesses on the same $0/$20/$100 structure (Sep 29) and discounted its API's "Contributor" tier in exchange for letting Meta use submitted prompts and outputs. DeepSeek retired V4-Flash for V4.1-Flash, now split into peak and off-peak rates, undated on DeepSeek's own pricing page, so OneHuman treats it as current-as-of-this-sweep, not a dated September event. GitHub Copilot's code-review default moved from Lite to Balanced and its mobile Chat retention switched from 28-day auto-delete to permanent, both Sep 28. Gemini's Ultra tier and Grok's lineup also show new price points this sweep, but neither vendor dates the change to September, so OneHuman isn't calling those September events either. Claude's subscription prices, along with Mistral's and Perplexity's, held unchanged through this sweep, even as Claude's own API model lineup turned over. For Mistral and Perplexity, which moved often earlier this year, "nothing changed" is itself worth stating.

Anthropic shipped three model changes in September: Fable 5.1, alongside a restricted-access Mythos 5.1 (Sep 1), with prompt-cache reads cut 75% to $0.25 per million tokens; Opus 5.5 (Sep 22), priced at $4/$20 per million tokens, down from Opus 5's $5/$25; and Sonnet 5.5 (Sep 28), at an unchanged $2/$10, which Anthropic says runs 30%+ faster and costs up to 30% less per task through fewer tokens used, not a lower list price.

What This Means

Two patterns and one correction define September.

Agents arrived before the rules. Muse (launched Sep 8, extended to small business Sep 29) spends against an unquantified free-tier limit and paid allowances denominated in tokens per week, not messages. OpenAI's dots (Sep 29) has its own cloud computer, lets a user define what it can do on its own and bring results back for review, and carries no published usage terms once its first allowance-free month ends. Two of the largest AI labs shipped autonomous, spending-capable agents in the same month, and neither shipped a durable price sheet alongside the product.

Where a model lives now matters more than what it is called. GPT-6 Sol and Luna sit in ChatGPT Work and Codex only; none reaches Chat on any tier, by OpenAI's own account. GPT-6.1 Sol is rolling out to the same surfaces, Pro users first, expanding to Plus, Business, Enterprise, and Edu, also never in Chat. Astra reaches Chat only on Pro, Business (Standard and Premium alike, at different quotas), and Enterprise; Plus users get Astra in Work and Codex alongside the others. chatgpt.com/pricing marks all four models with a checkmark next to Plus, and that checkmark means "available somewhere in the product," not "available in Chat." Plan, surface, and increasingly region decide what a subscription actually includes.

Units that don't translate. A Muse token has no published exchange rate against a ChatGPT message. Grok's X Search API now bills per post and per profile instead of per call. Gemini's consumer usage cap is now a multiplier with no stated base, and paid usage resets on a rolling five-hour clock against a weekly ceiling, with new purchasable credits on top. Copilot's allowances run on AI credits. Claude Code's weekly limit only makes sense against a boost extended five times, then replaced Sept 14 by a "permanent" level 17% below it. Claude Sonnet 5.5 held its list price and offers "up to 30% less" per task instead, a savings claim that arrives as an efficiency promise, not a price cut, and isn't verifiable against any published benchmark. None of these figures compares cleanly to what came before it, or to a competitor's version of the same idea.

The correction: BNA28 (Sep 22) reported that no Meta Muse tier had published a usage cap. That was wrong. Meta's Help Center had listed Power's and Maximum's weekly token caps by at least September 9. OneHuman's research tooling read the page as blank because it renders only in a browser, and a plain fetch returns nothing. Found and corrected in the open on September 30, the same practice as August's Perplexity correction: no spin, the figure now stands corrected on the article itself.

Consumer Protection Q&A

Q: My plan says GPT-6. Can I use it in my regular ChatGPT Chat? A: On Plus, no. Astra, Sol, and Luna already live in Work and Codex on Plus; GPT-6.1 Sol is rolling out there too, Pro users first, expanding to Plus, Business, Enterprise, and Edu. None of the four reaches Chat on Plus. Chat access to Astra requires Pro, Business (either tier), or Enterprise.

Q: What is a Muse token actually worth? A: Meta hasn't published an exchange rate against messages or requests. Power ($20/mo) buys 500 million tokens a week; Maximum ($100/mo) buys 3 billion. Neither figure translates into "how many things I can ask it."

Q: Do agents like dots or Muse cost extra later? A: Unclear by design. Dots usage doesn't count against plan allowances "for the next month," with no published terms after that. Muse's free tier has a usage limit Meta has not quantified.

Q: Is the cheapest API automatically the cheapest to use? A: Not necessarily. Muse's API "Contributor" tier prices at $0.10/$0.20 per million tokens, a fraction of the standard rate, in exchange for letting Meta use the prompts and outputs sent through it.

Q: Did my Copilot chats stop auto-deleting? A: Yes, as of September 28. Chat retention on github.com/mobile switched from a 28-day auto-delete to permanent storage, with no partial opt-out published.

What Happens Next

30 days: whether OpenAI publishes dots' usage terms once its allowance-free month ends, and whether GPT-5.5's widely-reported, still-unconfirmed October 14 retirement gets an OpenAI first-party notice. 90 days: whether Meta quantifies Muse's free-tier cap or ships outside the US, and whether OpenAI sets a firm date for the custom-GPT-to-plugin migration it announced Sep 11. 6-12 months: whether opaque, allowance-based pricing, tokens per week, AI credits, rolling compute windows, replaces fixed message counts industry-wide, the direction every tool that changed pricing this month moved in.

What to Do Next

• If you are paying for anything branded GPT-6 on Plus, check which surface, Chat, Work, or Codex, actually has it before assuming access. The pricing page's checkmarks won't tell you.
• If you use Meta's Muse, keep purchase and spend approvals switched on; it can act before you review each step. If you use OpenAI's dots, limit what it can do on its own. Both launched without durable usage terms.
• If you are in the EEA, Switzerland, or the UK, note that dots' Pro-tier rollout excludes you at launch.
• If you compare AI plans on sticker price alone, compare the allowance behind it instead, tokens per week, AI credits, or a weekly quota. The sticker no longer tells you the ceiling.

Bottom Line

Nine tools now on OneHuman's tracker, six reports published, and the month's clearest signal came from what two of the largest labs didn't publish rather than what they announced. Muse and dots both put spending-capable software into users' hands before either company wrote down what using it will cost past the first month. Everything else this month, from Gemini's capless multiplier to Grok's per-post billing to Claude Code's fifth extension, fits the same shape: the number that determines what you pay keeps moving to a unit that doesn't compare to last month's. One of OneHuman's own six reports needed a correction this month too. That's the standard applied to everyone else, applied here.

Sources

• OpenAI, ChatGPT release notes and Help Center, "ChatGPT Business models and limits" and "Introducing GPT-6 Sol and Luna," fetched September 30, 2026
• TechCrunch, September 29, 2026, on GPT-6.1 Sol: "not yet available in Chat"
• chatgpt.com/pricing, fetched September 30, 2026
• Meta Help Center, "About Muse subscriptions"; Meta Newsroom, September 8 and September 29, 2026, fetched September 30, 2026
• Axios, September 29, 2026, on Muse for Small Business
• OpenRouter, meta/muse-spark-1.3 listing, fetched September 30, 2026
• DeepSeek, api-docs.deepseek.com/quick_start/pricing, fetched September 30, 2026
• Gemini, gemini.google/us/subscriptions, fetched September 30, 2026
• xAI, grok.com/plans and docs.x.ai/developers/pricing, fetched September 30, 2026
• GitHub, Copilot changelog (August 28, 2026) and github.com/mobile retention change, effective September 28, 2026
• OneHuman, BNA24 through BNA29 (content/news/posts/, September 2-28, 2026)
• Anthropic, "Claude Fable 5.1 and Mythos 5.1" (anthropic.com/claude-fable-and-mythos-5-1), "Claude Opus 5.5" (anthropic.com/claude-opus-5-5), and "Claude Sonnet 5.5" (anthropic.com/claude-sonnet-5-5), fetched September 30, 2026

Verified by OneHuman · September 30, 2026

Independent. AI-assisted. Human-verified. No ads. No affiliates. No investors.