What each lab shipped, August–October 2026
Dates are the vendors’ own (newsrooms, release notes, help centers), re-read on October 26. Benchmark figures are self-reported by the vendor unless stated otherwise; we quote them only to show what each lab chose to emphasise.
| Lab | Models | Products and plans | Enterprise move |
|---|---|---|---|
| OpenAI | GPT-6 Sol / Luna (September 22), GPT-6.1 Sol and GPT-6 Astra (DevDay, September 29); GPT-6 with Intelligent UI for every ChatGPT plan (October 7); GPT-5.5 retired from ChatGPT on October 14 | Pro 100 / 200 / 500; dots always-on agent; Codex cloud; ChatGPT Slides and a parallel-agent GPT-6 Pro (October 16); ads on Free and Go in 89 countries (October 21) | Models coming to Google Cloud (October 21); Uber rolls ChatGPT Work out to 30,000+ employees (October 22) |
| Anthropic | Fable 5.1 and Mythos 5.1 (September), Opus 5.5 (September 22), Sonnet 5.5 (September 28), Haiku 5.5 (October 7) | Cowork merged into Claude; Claude Design, Slides and Docs; Claude Code mods; monthly API credits on Max; Claude Duet preview for US Pro and Max (October 21) | Team seats at 1.25x / 6.25x Pro per session; $100M Claude Frontier Academy (October 2); Agent Skills donated to an open foundation (October 22) |
| Gemini 4 Argon: announced September 30 for trusted cyber defenders, generally available October 13; output limit raised to 1M tokens | Argon in the Gemini app and AI Mode for AI Pro and Ultra; chat-to-tasks toggle; 40+ connected apps; Antigravity inside the Gemini macOS app | Argon API at an introductory $2 / $10; will host OpenAI models in its cloud | |
| Meta | Muse Spark powers Muse; Muse Wonder and Wonder Full (October 12) via the Meta Model API and Muse Code | Muse agent (free with limits; Power $20, Maximum $100); limited testing in the US and Canada; Wonder to reach Muse subscribers first | Muse for Small Business (September 29) with connectors such as Shopify, Slack and QuickBooks |
| SpaceXSI (formerly SpaceXAI / xAI) | Grok 4.6 (August 12), Grok 4.7 (September 21), Grok Voice 3.0 (October 14) | Grok Bot and Team Bots; The X Subscription (October 12): Grok, Grok Bot, Cursor and X Premium in one plan with a shared usage pool | Partner Network with 130+ early-access companies (October 22); Grok 4.6 on Bedrock, Foundry and Google’s agent platform |
| DeepSeek | V4-Pro GA (August 13), V4.1-Flash with a new natively multimodal architecture (September 10), V4.1-Flash-1017 with open weights (October 17) | Same model in the DeepSeek web app and app; Search and Fetch APIs at $0.75 per 1,000 calls | Lower API prices with an off-peak half price; Anthropic-format endpoint for coding tools |
| Alibaba Qwen | Qwen3.8-Max (2.4T parameters, 95B active, early August; weights on Hugging Face from August 8); Qwen3.8-Omni-Flash (September 18) | QwenWork workplace agent platform; Qwen-Image-2.1 open-sourced | API on Alibaba Cloud Model Studio |
Price ladders side by side
Official monthly list prices in USD, before tax. ClaudeMax sells only some of these (see the last section); the rest are here for comparison.
| Vendor | Entry | ~$20–30 | $100 rung | $200 rung | Above $200 |
|---|---|---|---|---|---|
| ChatGPT | Go $8 | Plus $20 | Pro 100 | Pro 200 | Pro 500 |
| Claude | Free | Pro $20 ($17 annual) | Max 5x $100 | Max 20x $200 | — |
| Google AI | Plus $4.99 | Pro $19.99 | Ultra 5x $99.99 | Ultra 20x $199.99 | — |
| Meta Muse | Free (limited) | Power $20 | Maximum $100 | — | — |
| X Subscription | Plus $8 | Premium $30 | Super $100 | Ultra $200 | — |
Three things stand out. First, the vocabulary has converged on multiples: 5x / 20x at Anthropic and Google, 4x / 15x / 50x Plus at SpaceXSI, and OpenAI’s three Pro tiers (OpenAI’s Help Center gives no ratios; press reports put them at about 5x / 10x / 25x Plus). Second, only OpenAI has gone above $200 for one user. Third, the bottom of the ladder is where business models diverge: OpenAI funds Free and Go partly with ads, SpaceXSI bundles X features, Google bundles storage, and Meta keeps Muse free with a limit. For teams, Claude Team Premium seats ($125 a month) and ChatGPT Business add seat-based plans on top.
Agent strategies compared
The clearest convergence of the quarter: each lab now has an agent that works in the background, on its own cloud computer or browser, and stops for approval before sensitive actions.
| Agent | Who gets it | Notable design choices |
|---|---|---|
| OpenAI dots | Pro and Business Premium in eligible markets; Enterprise beta | GPT-6 Astra, own cloud computer; chatting with it does not count toward limits, the Codex and Work tasks it starts do |
| Claude Duet | Pro and Max in the US, waitlist research preview | Built around Opus 5.5; morning brief; checks connected apps every few hours to save usage; refuses trading securities and moving money even with approval |
| Meta Muse | Free with limits, Power and Maximum; US and Canada, limited testing | Muse Spark today, Muse Wonder rolling out to subscribers first; nothing published, sent or paid without approval |
| Grok Bot / Team Bots | From the X Subscription Premium plan; enterprise plans | Own computer shared by your Bots; Team Bots shared across a company; usage pooled with Cursor |
| Gemini tasks | Google AI Pro and Ultra | Toggle from chat to tasks; works across Google apps and 40+ connected apps |
The differences are in gating and metering. OpenAI and Anthropic put agents behind their higher paid tiers and meter background work against the plan; Meta lets anyone in its two launch countries start free; SpaceXSI pools agent and coding usage across Grok and Cursor. None of them publishes how much agent work a given plan buys, which is why we expect usage terms (trend 2) to keep changing.
Consumer vs enterprise focus
- OpenAI plays both ends hardest: 1.2 billion weekly users, ads on the free tiers, and large enterprise rollouts of ChatGPT Work such as Uber’s.
- Anthropic leans enterprise and developer: Claude Code, Team seats, an engineer-training academy and API price cuts, with Duet as its first proactive consumer-style app.
- Google wins on distribution: Search AI Mode, the Gemini app, Workspace and a cloud that will now also resell OpenAI models.
- Meta is consumer first, with small businesses as the bridge, and ships its new models to developers through its own API.
- SpaceXSI bundles consumer (X, Grok) with developer tools (Cursor) and builds an enterprise channel through partners.
- DeepSeek and Qwen compete mainly on open weights and API price, which matters most to developers and companies hosting their own models.
Seven trend calls for the next 6–12 months
These are our analysis, not vendor statements. Each call lists the signals we observed, a confidence word (likely / possible / speculative) and what would prove it wrong. None of this is financial or investment advice.
1. The $100 / $200 usage-multiple ladder becomes the industry default
Our read: Paid AI plans are now priced by how much compute you may use, not by which features you get. Expect more vendors to add a rung above $200 for single heavy users. Confidence: Likely.
- OpenAI added Pro 500 above Pro 100 and Pro 200 on September 29; only the top tier gets Ultrafast.
- SpaceXSI’s X Subscription (October 12) sells Premium, Super and Ultra at $30 / $100 / $200 as 4x / 15x / 50x the Plus allowance.
- Google AI Ultra comes as 5x and 20x AI Pro at $99.99 and $199.99; Anthropic Max as 5x and 20x Pro at $100 and $200.
What would prove it wrong: A major lab drops its $200 tier, or goes back to one flat “unlimited” price.
2. Included allowances tighten; overflow moves to credits and pay-as-you-go
Our read: The headline price stays put while what it includes is re-cut, and heavy months are billed separately. Confidence: Likely.
- New Pro 200 subscriptions get a lower allowance than before at the same $200; grandfathered ones switch after October 29 (OpenAI Help Center).
- OpenAI bills Fast and Ultrafast as multiples of included usage; Anthropic sells usage credits on top of Pro and Max; Claude Duet’s background checks count toward plan limits.
- The X Subscription bills third-party models in Cursor at API list prices through top-ups.
What would prove it wrong: Two or more labs raise included usage at the same price without adding a new paid add-on.
3. The always-on agent becomes the main reason to pay
Our read: Within 6–12 months the flagship paid feature is an agent that keeps working while you are away, with its own cloud computer, approval gates and connectors, not a smarter chat box. Confidence: Likely.
- OpenAI dots (September 29): included in Pro and Business Premium, powered by GPT-6 Astra.
- Anthropic Claude Duet (October 21), SpaceXSI Grok Bot and Team Bots, Meta Muse, and Gemini’s new chat-to-tasks toggle (October 13) all ship the same pattern.
- Duet and Muse both ask before paying and use Link by Stripe; Duet refuses some actions even with approval, such as trading securities.
What would prove it wrong: By mid-2027 these agents are still waitlisted previews or limited to one country, or their usage terms make daily use impractical.
4. API prices per unit of capability keep falling, subscriptions do not
Our read: Developers get the same capability for less every quarter; consumer list prices hold and the savings show up as more usage or new features instead. Confidence: Likely.
- Haiku 5.5 runs about 75% cheaper than Haiku 4.5 and Sonnet 5.5 cache reads were halved (October 7); Opus 5.5 costs 40% less than Opus 5 on typical work.
- Gemini 4 Argon launched at an introductory $2 / $10 per million tokens; GPT-6.1 Sol is priced at one fifth of GPT-6 Astra.
- DeepSeek’s flash model costs $0.30 / $1.20 at peak and half that off-peak.
What would prove it wrong: A frontier lab raises the API list price of an existing model, or ends an introductory price without a cheaper successor.
5. Models are sold across rival platforms; lock-in moves to agents and data
Our read: Which model you use matters less than whose agent, connectors and memory you live in. Confidence: Possible.
- OpenAI models are coming to Google Cloud and, later, Google Antigravity and Gemini Agent (October 21); Haiku 5.5 launched on AWS, Google Cloud and Azure.
- Anthropic donated Agent Skills to the Agentic AI Foundation (October 22) after 70+ platforms, including ChatGPT, adopted it; DeepSeek’s API also speaks the Anthropic format.
- Grok 4.6 is on Amazon Bedrock, Microsoft Foundry and Google’s agent platform; media reported in October that Grok Bot routes some tasks to Claude Opus 5.5 (9to5Mac, Implicator).
What would prove it wrong: New exclusive cloud deals, or a lab pulling its models from a rival’s marketplace.
6. Chinese labs compete on open weights and price, and close in on agentic coding
Our read: DeepSeek and Qwen keep publishing weights and cutting API prices; on their own benchmark tables the gap to the US frontier on agentic coding is narrowing. Confidence: Possible.
- DeepSeek V4.1-Flash-1017 (October 17) shipped with open weights; DeepSeek reports 77.0 on DeepSWE v1.1, while Google reports 77.9% for Gemini 4 Argon (different harnesses, both self-reported).
- Qwen3.8-Max (2.4T parameters, 95B active) had weights on Hugging Face from August 8.
- DeepSeek opened its own Search and Fetch APIs at $0.75 per 1,000 calls, usable from Claude Code and Codex.
What would prove it wrong: Their release cadence slows, or independent leaderboards show the gap widening again.
7. The most capable models arrive in stages, gated by trusted-access programs
Our read: New frontier models will reach security researchers and vetted organisations first, and consumers weeks later. Confidence: Likely.
- Gemini 4 Argon went to Fairwind cyber defenders on September 30 and to everyone else on October 13.
- Anthropic expanded its Cyber Verification Program (October 6) and runs a Life Sciences Verification Program; OpenAI gates GPT-Rosalind behind trusted-access approval.
- Google released Argon without cyber guardrails only to trusted defenders.
What would prove it wrong: The next flagship from two of these labs launches to all paid users on day one.
What to buy today, given these trends
If allowances keep being re-cut and plans stay monthly, the sensible move is to buy for the next month, not the next year, and to choose by how you work. Live prices from our order system:
| If you… | Buy | Price here (30 days unless noted) |
|---|---|---|
| use ChatGPT daily, mostly chat, research and some Codex | ChatGPT Plus top-up | ¥185 ≈ 27.58 USDT |
| code or run agents most of the day in Codex and ChatGPT Work | ChatGPT Pro 200 top-up (ready-made also available) | ¥1,300 ≈ 193.75 USDT; ready-made ¥1,349 ≈ 201.05 USDT |
| run out of Pro 200 early every week, or need Ultrafast | ChatGPT Pro 500 top-up | ¥3,299 ≈ 491.66 USDT |
| hit the limit only in some months | Codex credits on top of Plus or Pro (500 credits per unit) | ¥180 ≈ 26.83 USDT |
| work mainly in Claude and Claude Code | Claude Pro, Max 5x or Max 20x top-up | Pro ¥249 ≈ 37.11 USDT; Max 5x ¥999 ≈ 148.89 USDT; Max 20x ¥1,999 ≈ 297.92 USDT |
| want more than Max 5x per session in a separate account | Claude Max 6.25x ready-made: a Claude Team-plan Premium seat, 6.25x Pro per session (Anthropic support article 9266767) | ¥1,379 ≈ 205.52 USDT |
| want to try an always-on agent and have no Muse account | Muse AI Access (includes 1 billion Muse tokens, one-time purchase) | ¥99 ≈ 14.76 USDT |
| already use Muse and hit the free limit | Muse AI Power or Maximum on your own account | Power ¥200 ≈ 29.81 USDT; Maximum ¥760 ≈ 113.27 USDT |
For the reasoning behind each vendor, read our deep dives on OpenAI’s October 2026 progress and Anthropic’s October 2026 progress. Google AI plans, SuperGrok and the X Subscription are not sold here.
How buying here works: top-ups are added to your own account and your sign-in data (Session JSON for ChatGPT, a cookie export for Claude and Muse) is checked before you pay; if the check fails you pay nothing. Ready-made accounts are hand-delivered in about 30 minutes to 4 hours after you verify a delivery email; Muse AI Access is delivered automatically, usually in 10–30 minutes. Pay in USDT on TRC20 or BEP20. Ready-made accounts are not refundable once delivered; if anything goes wrong, contact the bottom-right live chat or Telegram @ccnb1937 promptly.
Risk, briefly: each vendor’s terms apply per user, and access from unsupported regions may be restricted. ClaudeMax is independent and is not OpenAI, Anthropic, Meta or an authorized reseller; vendors set the allowances and may change them again.