● What is contested Whether valuations are a bubble. Serious people disagree, and this page does not settle it.
● What is not contested Vendor pricing. In six weeks OpenAI halved an API rate, Cursor’s entry price went $20 to $8, and Google made a live voice model free.
● The real buyer risk Not a crash. Churn - three model retirements in five weeks, all silent.
Every figure below was read from the vendor’s own page, not from market commentary.
Most writing about the AI bubble is about share prices, and almost none of it is checkable by the people who actually buy AI tools. So this page does something narrower and more useful: it looks at what the major vendors charge, how that has moved, and what a buyer should do about it - whichever way the valuation argument resolves.
What actually changed: prices went down, fast
Between late September and early November 2026, the published cost of using these models moved in one direction:
| Change | Detail | When |
|---|
| OpenAI halved a tier | GPT-6.1 Terra costs 25 input and 125 output credits per million, against Sol at 50 and 250 - exactly half on every line | Late Oct |
| Cursor cut its entry price | Entry tier moved from Pro at $20 to Plus at $8. The $20 and $60 tiers are now legacy, closed to new signups | 12 Oct |
| Google made live voice free | Gemini Live models are free of charge for input and output on the free tier. Paid is about $0.023/min combined | Current |
| xAI cut image generation | The new default is $0.01 per image cheaper at every resolution than the model it replaced | 2 Nov |
| Anthropic held price, cut usage | Claude Fable 5.1 held at $10/$50 per million, but Anthropic estimates typical workloads cost ~25% less, and agentic ones up to ~45% less, because it uses fewer tokens | Sept |
Whatever is happening to valuations, the cost of using frontier models is falling quickly. For anyone buying rather than investing, that is the headline, and it is the opposite of a squeeze.
The spread matters more than the cuts. On OpenAI’s own credit table, GPT-6 Astra costs a hundred times what GPT-6 Luna costs for the same work on the same plan. Most teams overpay by defaulting to the strongest model, not by picking the wrong vendor. Model selection is a bigger lever than anything in the bubble debate.
Tiers are being restructured, not just repriced
Three of the big consumer plans changed shape in October, and the shape change tells you more than the price:
- Cursor is now sold as The X Subscription - Free, Plus $8, Premium $30, Super $100, Ultra $200 - and the weekly usage pool is shared across Cursor and Grok. Entry at $8 buys Cursor models only, with no Auto, Cloud Agents or third-party models.
- Anthropic renamed Max 20x to Ultra at $200 and Max 5x to Max at $100. Verified Ultra accounts lost the five-hour limit entirely.
- OpenAI prices Pro at $100, $200 or $500 a month, and states Pro plans currently have no five-hour limit.
The common thread is that vendors are now metering you across products rather than per tool. Claude Code draws on your Claude allowance, Codex draws on your ChatGPT allowance, Cursor draws on a pool shared with Grok. That is what cost discipline looks like from the supplier side, and it changes how you should budget. Our cost comparison across the three coding tools works through the arithmetic.
The risk that is real for buyers is churn, not collapse
If you are paying for AI tools, the thing most likely to disrupt you in the next quarter is not a market event. It is a model being retired under you. xAI alone has run four API changes in five weeks:
| Date | Change | How you find out |
|---|
| 2 Oct | grok-voice-transcribe-1.0 end of life | Silently routed to 2.0 |
| 14 Oct | grok-voice-2.0 deprecated | Alias repointed |
| 2 Nov | grok-imagine-image-quality retired | Redirected to a lower quality default |
| 7 Nov | Text output watermarking | On by default, new header only |
Not one of these produces an error. Your code keeps running and what comes back quietly changes. That is a real operational risk and it is entirely independent of what happens to anyone’s share price. Details in our guide to the 7 November watermarking change.
What each side of the argument actually claims
Both cases are worth stating properly, because the honest answer is that this is unresolved.
The case that it is a bubble
- Capital expenditure is running far ahead of revenue from AI products, so the returns have to arrive on a schedule nobody has demonstrated yet.
- Infrastructure is being financed with debt secured against hardware whose useful life is uncertain.
- Falling prices can be read as competition for a market that is smaller than the buildout assumes, not as efficiency gains.
- Comparisons are drawn to the dot-com buildout and to the 19th-century railway booms, where the infrastructure proved useful and most of the investors still lost money.
The case that it is a reset, not a burst
- Anthropic reports run-rate revenue above $30 billion, up from about $9 billion at the end of 2025, with more than 1,000 business customers spending over $1 million annualised.
- Compute commitments are being signed for capacity that only comes online from 2027, which is a long-lead-time supply problem rather than idle capacity.
- Falling prices alongside rising usage is what efficiency gains look like, and the token cost of a given task has dropped faster than headline rates.
- Anthropic has filed an S-1, which means these numbers move from blog posts to audited disclosure.
Where we stand: we are a tools site, not a markets desk, and we are not going to tell you whether equities are overvalued. What we can say is that the figures in the bullish column above come from company filings and company pages, and much of what circulates in the bearish column comes from video summaries and forum threads. That asymmetry is worth noticing without mistaking it for an answer - a claim being badly sourced does not make it wrong.
What to actually do, either way
These five steps are worth doing whether valuations correct or not, which is why they are the right response to an argument you cannot settle:
- Audit which model your defaults point at. A hundredfold spread between the cheapest and most expensive model on the same plan dwarfs any vendor-level saving.
- Stop pinning model IDs you do not monitor. Put vendor retirement and migration pages on a calendar, and log the
model field from responses rather than the one you sent. - Know which pools are shared. If your coding tool and your chat tool draw on one allowance, budget for the pair.
- Keep a second provider warm. Not as a bet against anyone - as insurance against a deprecation notice with 60 days on it.
- Measure cost per completed task, not per token. The cheapest rate and the cheapest outcome are frequently different models.
If you want to work through vendor exposure properly, our vendor stability due diligence prompts and dependency audit prompts are built for exactly this.
FAQ
Is the AI bubble bursting?
That is an open argument about valuations, and this page does not claim to settle it. What is documented is that the cost of using frontier models fell sharply between September and November 2026, and that vendor revenue at Anthropic has grown from about $9 billion run-rate at the end of 2025 to over $30 billion.
Are AI tools getting more expensive?
Not currently. OpenAI halved an API tier, Cursor cut its entry price from $20 to $8, Google offers live voice models free on the free tier, and xAI’s new image default is a cent cheaper per image than the model it replaced.
Should I wait before committing to an AI tool?
Waiting mainly costs you the use of the tool. A more useful hedge is to avoid annual prepayment on anything you have not run for a month, keep a second provider tested, and track model retirement notices.
What happens to my tool if a vendor runs out of funding?
For the large vendors, the more likely disruptions are repricing and model retirement rather than shutdown - both of which have happened repeatedly in the past six weeks. For smaller tools, export your data periodically and know what you would switch to.
Why are AI prices falling if the buildout is so expensive?
Both explanations are live. One is efficiency - newer models complete the same task in fewer tokens, so cost per job falls faster than headline rates. The other is competition for a market that may be smaller than the buildout assumes. The pricing data alone cannot distinguish them.
Related reading
Sources
Every price and revenue figure on this page was read from a vendor pricing page or company announcement, on 2 November 2026. Where this page describes an argument about valuations, it describes it as an argument rather than as a finding. AIToolsRecap does not provide investment advice.