Cursor's new $30 plan leaves out Claude

What changed at Cursor, and why Vercel matters too

On 12 October, xAI, the company that owns the AI code editor Cursor, started selling Cursor inside one bundle called The X Subscription (opens in a new tab). The bundle also includes xAI’s Grok chatbot and X Premium, the paid subscription on X. New customers now start on the bundle, and the pricing page (opens in a new tab) on the Cursor site shows the bundle’s plans instead of the old ones.

The bundle comes in several plans. The one that matters for most of this story is called Premium. It costs $30 a month, and it’s the plan Cursor lists as the replacement for its old $20 Pro plan.

What you get for $30

You get a weekly allowance for Grok, the AI model xAI makes. You can use it in Cursor or in the Grok apps, and it resets every week.

Claude works differently, and Cursor’s help page (opens in a new tab) explains how: models from other companies, such as Claude and GPT, are billed “at API list prices from the first use”. In plain words, you pay the normal per-token price that the model’s maker publishes (what Anthropic charges for Claude, for example), and you start paying with your first request, because the plan includes none of it.

That’s a real change, because the old Pro plan did include Claude.

What it costs if you want Claude

Let’s do the math that Cursor didn’t publish.

Until last week, Cursor Pro cost $20 a month. When Cursor moved off per-request pricing in July 2025, it described Pro (opens in a new tab) as including $20 of usage on top models, such as Claude, billed at the model makers’ normal prices. So if you spent that $20 of usage on Claude, the $20 fee was the whole bill.

Cursor’s page on its old plans (opens in a new tab) lists Premium as the plan that replaces Pro, and Premium doesn’t include Claude. So if you move, the same month of Claude work costs the $30 plan plus about $20 of Claude at the normal price: about $50 in total, two and a half times what you paid before.

Keep two things in mind. First, this is our own math, based on Cursor’s pages, because Cursor hasn’t published an example. Second, people on the old plans can keep them for now. A Cursor staff post says they keep their included usage of other companies’ models until they switch, but Cursor hasn’t said how long the old plans will last.

Cursor users have noticed. On Cursor’s own forum thread (opens in a new tab), the most-liked reply complains that the new plans charge the normal API price for Claude and GPT “from the very first request”. It calls this the end of Cursor’s “model neutrality”, the idea that Cursor favours no one company’s models.

The plan also hides how much you get

There’s a second, smaller problem. The pricing page says the Premium plan gives four times the usage of Plus, a cheaper plan in the same bundle. But it never says how much usage Plus gives. Four times an amount you don’t know tells you very little.

Elena Verna, who leads growth at Lovable, an AI app builder, made the same complaint in her newsletter on 16 October:

Premium gets 4x more usage than Plus. Super gets 15x more. Except nowhere does it explain how much usage Plus actually includes.

She adds that her own company has the same problem: “I lead growth at Lovable, and we have the same problem.”

Vercel made the same kind of deal

The same week, Vercel launched Token Pass (opens in a new tab). Vercel runs AI Gateway, a service that lets developers reach many AI models with a single key. Token Pass costs $30 a month and gives you $80 of credit to spend on AI Gateway. So it looks like a discount, but it comes with a catch: the credit works only on some models.

Those models are GLM (made by Z.ai), Kimi (made by Moonshot) and, as of this week, a model from DeepSeek. They are open-weight models: their makers publish the model itself, so any company can host it and set its own price. Claude and GPT are left out.

Vercel names these models in the first line of its announcement, so you know the limit before you sign up. The terms (opens in a new tab) add a few things worth knowing:

Another company runs the models. Vercel sells the credit, but a company called Wafer currently runs the models and answers the requests.
When the credit runs out, requests stop. You aren’t charged more, but requests made with your Token Pass key simply fail until your next billing period (opens in a new tab) starts.
It’s not for traffic that has to stay up. Vercel’s docs say to use a regular key “for production traffic that has to stay up”.
Vercel can end it at any time, with a pro-rata refund of that month’s fee.

Is the $80 of credit really worth $80? You can now check. The docs say each request is charged at the price the model’s maker sets, and the page shows the price table. (When we first read the docs, on 20 October, requests were charged at the price Wafer set, which only subscribers could see.)

Why discount these models in particular? Vercel published a report on its gateway (opens in a new tab) the same day. By its own estimate, open-weight models made up about 23% of what customers spent there in September, up from about 5% in June, though this report counts more models as open-weight than the last one did. So the discount covers a part of the business that is growing. (These are Vercel’s own figures, from its own traffic.)

Others have drawn the same line

Cursor and Vercel aren’t the first. Two other companies follow the same pattern this month, and one goes the other way.

DigitalOcean discounts the open models it runs itself. We covered it earlier this month, and its pricing page (opens in a new tab) says that Claude and GPT “aren’t part of the discount”.
Anthropic now includes credit for its API (opens in a new tab), the way developers build Claude into their own software, in its Max and Team plans. The credit goes to Claude, because Claude is the only model Anthropic sells.
Warp, the developer terminal we wrote about last week, does the opposite: it adds a markup. It resells models from several companies, Claude included, and on its pay-as-you-go plan, where you pay for what you use, it charges 20% more (opens in a new tab) than the model makers do.

Our take

At Cursor, Vercel and DigitalOcean, the discount covers the models the seller makes itself or open models any company can host, and Claude is billed at its normal price. At Anthropic the credit covers only Claude, because that’s all it sells, and Warp adds a markup instead.

So when you compare AI plans, check two things on the pricing page: which models the discount covers, and what you pay for the ones it leaves out.

To be fair to xAI, the bundle is a good deal for some people. If you already paid separately for SuperGrok, Cursor Pro and X Premium+, xAI says those cost “$90/month combined” (opens in a new tab), against $30 for the Premium plan. The most-liked complaint on the forum comes from someone who wanted Cursor with Claude and GPT, and that buyer may not be the one xAI built this bundle for.

Our hunch is that the lists of included models will keep changing. Vercel says it will add more (opens in a new tab), and its pages already list DeepSeek as a third model. We’re watching for the first company to discount a model it has to buy from someone else, like Claude, because that would break the pattern.

Gauge changed its slogan and cut its Growth plan at the same price

Gauge sells AEO, short for answer-engine optimisation: helping a brand show up in the answers that AI assistants like ChatGPT give. One of its studies was part of our lead story two weeks ago.

On 17 October, Gauge changed the headline on its home page (opens in a new tab) to “The Agent-Led Growth Platform”. It describes agent-led growth as winning customers and users when AI agents find, recommend, choose and use products on a person’s behalf.

On the same day, Gauge’s Growth plan got smaller at the same price (opens in a new tab). The plan tracks how your brand shows up in AI assistants’ answers. It now tracks three assistants instead of six (ChatGPT, Google AI Mode and Microsoft Copilot), tracks half as many answers each month, and includes fewer seats, meaning user logins. The other three assistants, Google AI Overviews, Gemini and Perplexity, moved to the Enterprise plan, which has a custom price. (We know what the plan looked like before because we keep our own record of the page.)

Gauge didn’t invent the term, and its own post (opens in a new tab) doesn’t claim to. Since last winter, companies that sell AI sales agents have used “agent-led growth” (opens in a new tab) to mean a sales process run by bots. Gauge means something different: coding agents choosing which tools to use. So far we’ve found no other company using it that way.

Our take: a word on one company’s website is marketing. It becomes a category only when a second company sells under the same word, in the same sense. And Gauge’s new headline arrived on the same day as a cut to what the Growth plan gives you.

The ChatGPT link tag went missing for about a week

When ChatGPT links to a website, it usually adds a small label, called a tag, to the end of the link: utm_source=chatgpt.com. That tag is the main way your analytics tool, Google Analytics for example, knows a visitor came from ChatGPT.

From about 8 October, a growing share of links lost the tag. Profound, a company that tracks AI search, says that at the worst point only about 1 in 20 (opens in a new tab) ChatGPT answers with links still carried it. (These are Profound’s own numbers, and it hasn’t published its method.)

There was a second problem. Cedric Huesler, a senior director at Adobe, found that the links had also stopped passing along the page the visitor came from, according to LLM Pulse (opens in a new tab), another tracking company. With both the tag and that information gone, a visit from ChatGPT looks like someone typed the address by hand. LLM Pulse says ChatGPT visits to the sites it monitors fell by almost half over one weekend, while visits from Perplexity and Claude stayed normal.

Then the tag came back. Huesler says OpenAI told him it was a bug, and that a fix began rolling out on 15 October. Within hours, tagged visits were back (opens in a new tab) to roughly the level of the week before.

OpenAI itself has said nothing in public, so the only word from OpenAI comes secondhand, through one person, Huesler. And nobody has said whether the links again tell websites where their visitors came from.

If you report on traffic from AI, your number depends on a tag that OpenAI adds and can remove. This is the third week running that AI citations, the links an AI answer gives to your site, have come up here. Two weeks ago a study showed they don’t last long, and last week researchers paid $14 to get a planted article cited by one AI search engine within the hour. This week, the way you count them broke for about a week.

Briefly

Discounts with an end date. Three companies put end dates on discounts this week. DeepSeek made some of its web search and page fetch calls free until 31 October (opens in a new tab), and published the price after that. Google gave its Gemini 4 Argon model a launch price that doubles (opens in a new tab) in December. Vercel gave an 80% discount for 30 days (opens in a new tab) on Drama 3, a model from Fish Audio, but didn’t say what it will cost afterwards. A discount is only useful if you know what you’ll pay when it ends.

Keenable charges agents more than labs. Keenable, a search service that AI programs call, now shows two plans (opens in a new tab). Developers building AI agents pay for what they use, at four times the price per request of the tier for AI labs and companies that host models, which starts at 100 requests a second. Both prices existed before, so what’s new is showing them as two plans for two kinds of buyer.

OpenAI says a ChatGPT plugin tripled sign-ups for Stack Overflow. Stack Overflow for Agents is the Q&A site’s product that coding agents connect to. OpenAI published a case study (opens in a new tab) saying the ChatGPT plugin tripled its sign-ups. OpenAI’s explanation is that setting it up inside ChatGPT takes fewer steps than connecting through MCP, the usual way coding agents connect to outside tools. Because OpenAI is writing about its own product, trust the direction, more sign-ups, more than the exact number.

Rewritten on 23 October 2026 to make it easier to read. Three facts changed since it was first sent. The Token Pass docs at Vercel now charge each request at the model maker’s normal price and list a third model, so the $80 of credit can be checked. Links in ChatGPT answers lost their tag gradually from 8 October, with the worst drop on 14 October, rather than all at once. And the free DeepSeek calls run until 31 October, not 30 October.

Everything else this week

Launch · 3

Release · 9

News · 7

Campaign · 4

Sources, check them yourself

Every load-bearing claim above links to one of these. A single-sourced claim says so in the text.