Cursor's New Plans Bundle Grok: Claude and GPT Now Billed at API List Price

Cursor's subscription system has undergone a structural change. The plan new users land on by default only covers Grok models, while Claude and GPT—though still callable on Premium and above—are billed at API list price from prepaid credits starting with the very first request. For developers who rely heavily on Claude, this means the monthly fee and the model bill are now formally decoupled.

What Exactly Changed in the New Plans

According to Cursor's help pages, third-party models (such as Claude or GPT) are no longer included in plan credits. New users start on The X Subscription from the moment they sign up—a bundled subscription combining Grok, Cursor, and X Premium, whose credits cover only the Grok family of models.

This bundle was announced by SpaceXSI (what TechCrunch calls "the company now known as SpaceX Super Intelligence") on x.ai, and Cursor posted a parallel announcement on its community forum the same day. SpaceX had previously agreed to acquire Anysphere, the company behind Cursor, for $60 billion.

Pricing and limits by tier are as follows:

Plan	Price	Credits	Claude/GPT
Free	$0	Limited	Not listed
Plus	$8/month (promotional)	Basic	No, Grok only
Premium	$30/month	4x Plus	Yes, usage-based billing
Super	$100/month	15x Plus	Yes, usage-based billing
Ultra	$200/month	50x Plus	Yes, usage-based billing

One detail worth noting: on paid plans, Grok and Cursor share the same weekly credit pool—a Grok conversation and a Cursor agent run draw from the same quota. But the Cursor pricing page describes usage only in "multiples of Plus," without converting to tokens or request counts, so users can't yet translate "4x" into a concrete number of agent sessions.

What Billing at API List Price Means

Premium and above offer no credit subsidy for third-party models. You pay for exactly what you use, with an optional spending cap.

So the plan fee and the Claude bill become two separate expenses. Anthropic's official pricing: Sonnet 5.5 is $2 per million input tokens and $10 per million output tokens; Opus 5.5 is $4 input and $20 output. Take a sample session—3 million uncached input tokens plus 150,000 output tokens—Sonnet 5.5 costs $7.50 and Opus 5.5 costs $15.

A coding agent resends the entire conversation history on every turn, which is the fundamental reason input tokens accumulate so quickly. Anthropic's cache read pricing is far lower—Sonnet 5.5 at $0.10 per million tokens—so cache hit rate significantly affects the total bill. But the Cursor page only says "API list price" without explaining how caching is billed, so the calculations above should be treated as estimates rather than quotes.

Set your spending cap before your first Claude request, while prepaid credits are still untouched.

On "uncached input tokens" vs. "cache reads": LLM API billing typically splits input tokens into two categories. Prompts appearing for the first time count as "uncached input" (cache write) and are billed at the standard input rate; identical prefixes repeated in subsequent requests, if they hit the provider's prompt caching, are billed at the "cache read" rate—often a fraction of the former. A coding agent must carry the full conversation history on every turn, so if the historical portion hits the cache, costs drop substantially. But caches have a lifespan (usually minutes to an hour); too long a gap invalidates them and they revert to uncached pricing—which is why "many rapid turns in a row" is cheaper than "many turns with gaps." The Cursor page only says "API list price" and doesn't clarify whether cache reads and writes follow Anthropic's official rates, so your actual bill may be higher or lower than an estimate based purely on uncached pricing.

The relationship between spending caps and prepaid credits: In usage-based billing, prepaid credits are a balance you top up in advance and that is drawn down per call, separate from the monthly fee. Setting a spending limit is like adding a gate to that balance: once cumulative charges hit the threshold, subsequent requests are rejected rather than continuing to overdraw or auto-recharge. For teams, set the cap to an acceptable monthly figure before the first Claude/GPT call, and periodically check whether "the rate of prepaid balance decline" matches "actual token usage"—so you don't burn through an entire month's budget in a few days due to cache misses or long-context resends.

Who's Affected, Who Isn't

Existing subscribers: Current Cursor, SuperGrok, or X Premium plans keep their prices and features unchanged unless you actively switch or cancel. The forum announcement explicitly states that existing Cursor subscribers retain their included third-party model credits until they switch.

New users: They start on The X Subscription, including "users with only a free X account." The forum post's author notes that the new Ultra and the old Cursor Ultra are different plans, and the old one is no longer available to new users.

Plan changes: Existing users can still switch between current Cursor plans without migrating, but plans billed through the Apple App Store must complete migration first.

Teams and Enterprise: Existing teams can add seats or adjust seat tiers without migrating. Enterprise customers without an active Grok account can migrate with help from the customer team; other enterprise accounts are not yet supported.

One clause bears repeating: "Migration is permanent, cannot be undone, and support cannot undo it for you." If your Cursor plan came from the App Store, you must cancel it yourself in your device settings after migrating, or Apple will keep charging you. Migration prorates a refund for the old plan and offers a limited-time discount.

If most of your team's Cursor work runs on Claude or GPT, calculate last month's actual usage at list price before deciding.

Where the Bundle Makes Sense

If Grok models can handle your work, this is a good deal. SpaceXSI says Premium bundles SuperGrok, Cursor Pro, and X Premium+, which separately total $90/month, for a bundled price of $30. It also claims each X Subscription tier "provides at least twice the coding usage of the corresponding Cursor plan," and that each tier delivers "more usage per dollar than the plan it replaces."

But early feedback skews negative. Multiple replies in the Cursor forum thread are critical, with one writing: "Cursor's users came for the freedom to choose their model." The forum sample is small and self-selected, so treat it as a sentiment signal only.

If you haven't used Grok in Cursor yet, the most worthwhile test is this: run the same task with Grok and with your current go-to model, and compare diff quality and the bill.

Interpreting claims like "at least twice the coding usage": SpaceXSI claims each X Subscription tier "provides at least twice the coding usage of the corresponding Cursor plan," but the unit of usage isn't disclosed as tokens, request counts, or agent sessions. When credits are described in "multiples" rather than absolute units, users can't independently verify them or convert them directly against token-billed Claude/GPT. For developers, the only reliable approach is to run the same batch of real tasks with Grok and with your current model, record completion time, retry counts, and diff quality, and translate the "multiple" into your own workload terms. Only if Grok's output is still acceptable under that measure is the bundle price genuinely a good deal.

Keep Model Choice in Your Own Hands

Another thing happened the same day: Meta announced that its larger Muse Wonder Full model would be available only to Muse Code subscribers, not via API. Within a single day, two products were using plans to determine which models you can reach.

Three things you can do this week:

1. List the models your team uses in Cursor, and calculate what that usage would cost at list price.
2. If you're on the new plans, enable a spending cap for third-party model usage.
3. Keep at least one agent setup that brings its own credits for your target models, so price changes at other vendors don't directly interrupt your work.

The author develops Hermes IDE, which can run command-line agents like Claude Code, Codex, Copilot, and OpenCode side by side, each in its own terminal and worktree, using its own login credentials. Users can pick the agent, model, effort, and account per task, and when one agent hits its usage limit, they can move the task to another to continue. It doesn't make tokens cheaper; what it changes is how much a single vendor's pricing change affects you.

Why "bring-your-own-credits" agent setups reduce lock-in risk: Multi-agent parallel tools (like the Hermes IDE mentioned in the article) return model choice to the user: each CLI agent uses its own login credentials and draws from its own subscription or API credits. When one vendor adjusts plans, restricts models, or raises prices, what's affected is that agent's channel—not your entire workflow. You can move the same task to another channel that still has credits and keep going, without reconfiguring the project or migrating data. It doesn't lower the per-token price, but it downgrades "a single vendor's pricing change" from a business interruption risk to a task-routing adjustment.

Will You Migrate?

If you use Cursor daily, does $30 of bundled Grok usage beat paying list price for Claude or GPT tokens? Or is a plan covering only a single vendor's models simply unacceptable? The answer depends on your model preferences and actual usage structure.

One-line Summary

Cursor's new-user plans cover only Grok, with Claude and GPT billed at API list price from prepaid credits.

Paragraph Summary

Cursor's subscription system has seen a structural adjustment: new users default to The X Subscription, whose credits cover only Grok models. Claude and GPT remain callable on Premium and above, but no longer receive plan subsidies—instead, they're billed at API list price from prepaid credits starting with the first request. This formally decouples the monthly fee from the model bill. Existing subscribers keep their plans and third-party model credits, but migration is permanent and irreversible; teams and Enterprise have varying migration conditions. On paid plans, Grok and Cursor share a weekly credit pool, and the official description gives usage only as "multiples of Plus." SpaceXSI claims the bundle is worth $90 but costs only $30, though early forum feedback skews negative. The author recommends setting spending caps in advance, calculating actual usage at list price, and keeping a bring-your-own-credits agent setup to spread single-vendor pricing risk.