Usage hits 100% with fewer tokens; GPT/Claude vanish; dashboard hides $ cost

Support Bug Reports
1

Where does the bug appear (feature/product)?

Cursor IDE

Describe the Bug

On Pro Plus, usage hit 100% mid-cycle after ~1.17B tokens, while the previous cycle used ~2.20B tokens without exhausting usage.

Root cause appears to be dollar-cost metering (not raw tokens), but the self-serve Usage/Spending UI no longer shows $ cost breakdown, so users cannot see why fewer tokens burn the quota faster.

Additionally, after hitting the limit:

  1. GPT/Claude disappear from the model picker (AvailableModels returns 0 Claude/GPT/Gemini; only Auto/Composer/Grok/Kimi/GLM remain).
  2. On-demand is disabled (noUsageBasedAllowed=true), so third-party models stay locked.
  3. Heavy cursor-grok-4.6-high-fast usage appears to spill into / exhaust the Other Models (API) pool, then GPT/Claude become unavailable.

PlanUsage snapshot: autoPercentUsed=100, apiPercentUsed=100, totalPercentUsed=100, includedSpend limit=$70, remainingBonus=false. Message: “You’ve used 100% of your included API usage”.

Steps to Reproduce

  1. Use Pro Plus with heavy Agent usage on cursor-grok-4.6-high-fast during a billing cycle.
  2. Compare Dashboard Usage token totals vs previous cycle (previous ~2.2B tokens OK; current ~1.2B hits 100%).
  3. Open Spending/Usage pages and try to reconcile burn rate by $ cost — Cost/$ columns are missing for self-serve plans.
  4. After usage hits 100%, open model picker / AvailableModels.
  5. Observe GPT/Claude gone; only Cursor models + Kimi/GLM remain; on-demand off blocks third-party usage.

Expected Behavior

  1. Dashboard should show included usage burn in $ (or equivalent credits) and per-model cost share for self-serve plans, not only opaque token counts + percentage.
  2. Fast vs non-Fast pricing impact should be visible before users exhaust the cycle.
  3. Cursor Models spillover into Other Models / API pool should be clearly labeled in UI.
  4. When API pool is exhausted, model picker should explain why GPT/Claude are unavailable (quota vs regional filter), instead of silently removing them.
  5. If regional filtering also applies, surface that explicitly.

Operating System

MacOS

Version Information

Cursor IDE Version: 3.19.19
VSCode Version: (bundled)
OS: macOS 15.6.1 (24G90)
Plan: Pro Plus (individualMembershipType=pro_plus)
Billing cycle: 2026-09-02 to 2026-10-02

For AI issues: which model did you use?

Primarily cursor-grok-4.6-high-fast; also composer-2.5-fast / Auto. GPT and Claude previously available, now missing from picker.

For AI issues: add Request ID with privacy disabled

N/A — billing/usage/model-catalog issue (not a single failed AI request). Account userId owningUser=159188509.

Additional Information

Measured aggregates from DashboardService APIs:

Previous cycle (~Aug 2–Sep 2):

  • ~2.199B tokens, ~$1178 equivalent cost
  • Mix: grok-4.6-high-fast (~$525 / 436M), grok-4.6-high non-fast (~$330 / 865M), default/Auto (~$256 / 710M)
  • Effective ~$0.54 / M tokens

Current cycle (Sep 2–Sep 10):

  • ~1.173B tokens, ~$1311 equivalent cost, usage 100%
  • Almost entirely cursor-grok-4.6-high-fast (~$1277 / 1089M, ~97% of cost)
  • Effective ~$1.12 / M tokens (~2.1x previous)

Current policy status:

  • GetHardLimit: noUsageBasedAllowed=true
  • GetUsageLimitPolicyStatus: LIMIT_HIT_STAGE_SLOW_POOL, allowedModelIds=[“default”] only for slow queue
  • AvailableModels: 23 models, 0 Claude/GPT/Gemini

Ask:

  1. Restore transparent $ / credit burn UI for self-serve plans
  2. Clarify Fast vs non-Fast cost impact in product UI
  3. Explain Cursor Models spillover into Other Models
  4. Don’t silently drop GPT/Claude from picker without reason (quota vs region)

Does this stop you from using Cursor

Sometimes - I can sometimes use Cursor

6

Hey @cscs
Included usage is billed by request cost, not token count, and this cycle was almost entirely Grok 4.6 Fast, which is 2x the standard per-token rate (pricing). Last cycle leaned on Auto, non-Fast Grok, and the launch discount, so the same tokens went further.

A few options until your Oct 2 reset:

  • Switch Grok 4.6 to the standard (non-Fast) variant in the model picker (~half the cost).
  • Stick with Auto, which stays available.
  • Turn on on-demand usage with a spend limit if you want to keep going now.

On GPT/Claude/Gemini missing: availability varies by provider and region, set by the providers, not Cursor. Auto and Cursor’s models stay available. Details: Regions.

7

That effective $/M token doubling when almost everything is grok-high-fast is the smoking gun. Until the $ breakdown is back, a practical hedge is pinning a non-fast model for routine edits and keeping a hard on-demand cap at $0 so the picker cliff does not surprise you mid-task.

10

Adding to this discussion, I am experiencing a very similar issue on the standard Pro plan. Last month, I utilized approximately 334M tokens while staying around 30% utilization. This month, maintaining the exact same workflow and habits, my usage reached 460M tokens and 65% utilization in just 11 days.

As noted regarding request-cost billing, it appears that Auto’s underlying routing behavior has shifted, causing quotas to drain significantly faster. Because self-serve dashboards lack a detailed cost breakdown, it is difficult to track what is driving this increased burn rate. Greater visibility into model-weight routing or cost-per-request tracking would help clarify these sudden quota discrepancies.

11

It just vanish with my token all of the sudden went from 40% to 70% of auto and 49% to 91% fo api without using any api or extra model and in on ultra that a lot of money of mine vanish