I am seeing the same as Spencer too. Hopefully, someone from Cursor will respond to which one it is.
@Colin we are all waiting on your response…
Also worth mentioning, On-Demand is billed correctly (Opus 4.8 prices listed above).
Meaning, if you:
- Spend $200 On-Demand, then you get $200 in Other Model usage
- Spend $200 on Ultra subscription, you now get $125 worth of Other Model usage
Which means it is more cost-effective for users to:
- Cancel Ultra ($200/mo), switch to Pro ($20/mo), and use On-Demand so we get what we pay for
- Or switch to VS Code since Cursor is built on VS Code
Same gap on Pro+ ($60/month), cycle 20 Sep–20 Oct 2026. This is not Ultra and not a pool spillover: Cursor Models is 20.9M tokens and 1.5%. The only Other Models row is claude-opus-5-thinking-high, 9.1M tokens, shown as 31.6%.
Tooltip:
- Cache Read 8,501,197
- Cache Write 555,028
- Input 68
- Output 38,929
- Total 9,095,222
At the published Opus 5 rates ($5 / $6.25 / $0.50 / $25 per million) that is about $8.69. Against the documented $70 Other Models pool that is about 12.4%, not 31.6%. 31.6% of $70 would be about $22. Uncached input is 68 tokens, so the cache was used.
Could someone from Cursor say whether this 31.6% is correct, and what the Other Models percentage is actually divided by? The list-price total and the meter still do not match, and there is no published formula.
I’ve been a Cursor only user for nearly 2 years now. But sorry guys, I got Claude Code the other day and I’m blown away. The usage you get for the $200 account is laughable compared to even what I used to get with Cursor.
I’m still using Cursor with Claude Code as a plugin, so I have the best of both worlds. I really wanted to love Cursor as the only tool I need, but with the cut 3rd party usage, that can’t be the case. So Grok for debugging and simpler tasks (I have a whole Cursor toolset I’ve built!) and Claude Code for the main coding brain.
And don’t get me wrong I really love what they’re doing with Grok! And it may one day compete with the frontier models and I hope it does and I’m totally rooting for them. However as Musk eluded to: the other guys basically got a 3 year head start on them.
I wish I found this early, I have been double stung this week.
I couldn’t afford Ultra for 3 days making ends meet, I thought no worries I’ll down grade Pro Plus on 28th September and then upgrade, I have done this once in the past.
The utilisation changes I would assume came into play on the 1st of October. I upgraded 1st October, with 22% other model usage and 3 days into the period showing on 30th September under Pro Plus.
I upgraded to Ultra on 1st October knowing I would get some refund. I got only $20 refunded for only 22% other model usage whilst on Pro Plus. Now I don’t expect 100% refund but C’mon why only $20. I emailed hi, no qualification just saying it was pro rata. however the numbers didn’t add up. This was very different to the last time I upgraded mid subscription.
I’m assuming, the pro rata rate was based on the updated on the 1st October. And yeah my cost analysis for just 4 hours and 8 events looks like $100-150 USD of other models could be the new ball park - including bugbot - which is hard to even estimate cost. By my estimation if I continued, I won’t make more than 53 prompts and 25 PR reviews before running out on the other model usage. I use other frontier models for review the work from cursor models. ~53 potential is a big differences from 386 prompts and bugbot events during 28th Aug-27th Sep. I have not touched my account pending response from support.
I looked at my first 2 and only prompts on 1st Oct with Opus 5.5 and Sonnet <1 hour with my upgraded subscription. $3.9 USD on pricing based on documented cost with token details from cursor and my other model usage was 8% (6 events was from bugbots!) from reset on Ultra!
I thought it was an issue with the upgrade reset and immediately sent a query to hi. The same AI responses about not publishing how it is calculated.
I have sent numerous emails to hi, initially for 3 responses from AI and the scripted response. Now my emails are ignored, sent a survey, ignored.
I love cursor for the flexibility upto now. I requested a refund since I could not get a transparent response if this reset was a hiccup or actual, before finding this thread. Looks like it is the new norm and a change without confirmation based on this thread. Cursor have reverted with a ‘once you use it no refunds’ for my request for a refund.
Now for a new user asking for a refund, fair enough, for someone that upgraded and was twice stung, for $3.9 and 4 hours usage - for 2 prompts (6 bugbots), when changes were made without announcement that is financially straining myself - that is a tough pill. I hope there is something that can arranged.
It just hurts and as great as Cursor is, it is disappointing that support have not been any to respond to even explain if this is a hiccup or are real changes. $330 AUD is not a small amount for Ultra for me, I have shifted budgets to leverage and learn AI, and $400 AUD is what I have been stung with this month for maybe $125 USD of other models usage. (down from ~$450 USD).
Hopefully we get some transparent details of current state with usage henceforth. I’m aware this should not be a medium to attempt to contact Cursor, I agree, I wish this wasn’t something I had to do to be heard.
Hi everyone, and hopefully someone from the Cursor staff (@deanrie, @kevinn).
I’m an Ultra plan subscriber ($200/mo) looking for official clarification regarding the actual limits of the “Other Models” pool (third-party API usage like Claude Opus).
Historically, the Ultra plan was known to include a $400 allowance for third-party models. The pricing page currently advertises “20x Pro limits on Agent”, and since the Pro plan historically included $20 for other models, the logical expectation remains $400 (20 x $20). In fact, querying my account details via the API still returns included: 40000 (which implies $400).
However, a reconstruction of recent updates points to a completely different reality.
The Timeline of Undocumented Changes: Over the last few months, tracking our actual usage has become increasingly difficult due to a specific sequence of updates:
- Until July 2026: Cursor docs (“Models & Pricing”) explicitly stated that Pro includes $20 of “Other Models” usage and Ultra includes $400. The dashboard clearly displayed our costs in dollars.
- July 31, 2026: The “Cost” column was removed from the Usage page for individual plans, and dollar amounts were zeroed out in the dashboard API. Users can now only see percentages, not dollar values.
- August 24, 2026: Cursor announced the end of flat-rate Auto pricing and stated it was “increasing included limits” for Cursor Models. However, no announcement was made regarding the Other Models pool, and the specific dollar amounts ($20, $400) were quietly removed from the public documentation.
- September/October 2026: After noticing my percentage meter spiking unusually fast after a few Opus 5.5 queries, I pulled my raw usage data from the dashboard to run the numbers manually.
The Math (from the internal API data): The underlying math proves that the denominator for the Other Models pool is currently set to $125, not $400.
- Total spent on Opus 5.5: $11.38 (verified via Anthropic’s standard token pricing)
- Dashboard
apiPercentUsed: 9.104% - Calculation: 11.38/0.09104=∗∗11.38 / 0.09104 = **11.38/0.09104=∗∗124.99** (If the pool were $400, $11.38 would represent 2.84%, not 9.1%)
For cross-verification (Cursor Models Pool):
- Total spent on Grok/Auto: $533.79
- Dashboard
cursorPercentUsed: 17.793% - Calculation: 533.79/0.17793=∗∗ 533.79 / 0.17793 = **~533.79/0.17793=∗∗ 3,000**
- Total Weighted Usage: (533.79+533.79 +533.79+11.38) / (3000+3000 +3000+125) = 17.445% (which matches my dashboard’s
totalPercentUsed: 17.44544%exactly).
The math is undeniable: the billing system is currently measuring our third-party model usage against a $125 cap, representing a ~68% reduction from the expected $400.
I understand that the $3,000 limit on first-party models is a generous increase, but reducing the third-party pool without an explicit changelog or email notification—especially right after removing dollar-value visibility from the UI—feels like a severe lack of transparency.
My questions for the team:
- Is the $125 cap a bug, or is this the intended new limit for the Ultra plan’s third-party models?
- If this is intentional, what exactly does “20x Pro limits” mean on the current pricing page?
- Can we get back the dollar-value visibility on the dashboard so professionals can predictably track their $200/month investment?
I love Cursor and use it daily, but we need transparent cost structures. Looking forward to some clarity on this!
Thank you.
Wow, thank you for this! Amazing work. I, too, eagerly await a response from Cursor regarding this.
I also have the same problem, I upgraded my account to Ultra in the previous cycle and it was amazing, I barely used Other Models, despite using them almost everyday, and also worked a lot with grok. And in this billing cycle only after 1 week I see that almost all my Other models limits are used, without using habit changes. I downloaded csv-s, analyzed them and get to the same results as above - seems now instead of $400 on other models the limit is only ~$125, and it was not communicated with me, and not what I subscribed too. It would be nice to get an explanation because it looks like if On Demand usage is billed correctly it makes sense to return back to $20 subscription with $200 of on demand instead of what I have now.
I am seeing the same issue on a directly paid Cursor Ultra subscription, not a SuperGrok-linked plan.
My Ultra subscription renewed for $200 on Oct 4, 2026.
Based on my exported Usage CSV and Cursor’s published model rates, my current third-party model usage is approximately $24.60. However, the Spending dashboard already shows 20% of Other Models used.
That implies an effective allowance of approximately:
$24.60 / 20% ≈ $123–$125
This closely matches the ~$125 figures reported by other paid Ultra users in this thread, rather than the previously documented $400 API usage allowance.
I contacted support under ticket T-G60954 and explicitly requested human billing review twice. Both responses came from “Cursor’s AI Support Assistant” and repeated only that subscriptions with usage are not refundable. Neither response addressed:
- when the Ultra allowance changed from the previously documented $400;
- how existing subscribers were notified before renewal;
- why my current meter mathematically implies ~$125;
- my request for human review.
@Colin @deanrie — could a human billing specialist please review T-G60954?
I am requesting a full refund of the Oct 4 renewal, because I renewed based on the previously represented Ultra allowance and the materially lower allowance appears to have been applied without clear notice before renewal.
I am not requesting bonus usage or a manual quota refill. I am requesting review of the billing terms applied at renewal.
Same meter, different failure mode. On my Pro account the third-party pool sits at
0% for the whole cycle, not because I did not use it, but because every model behind
it returns “This model provider doesn’t serve your region”. Auto and Composer, both
Cursor-side, still work, and that is where my 3% went.
So the allowance I paid for is metered at 0% and is also the part I cannot reach. If
the meter is the basis for pro-rated refunds, that reads in my favour, yet the ticket
answer was “does not meet refund eligibility”. Worth flagging here since this thread
is already about reconciling the meter with reality.
I hope a human billing specialist can review my ticket T-G60970.
Sooo… no answer from Cursor? I suspect they’re going to announce something soon. Probably something emphasizing how amazing Grok is and to use Grok Bot as they step away from 3rd party models. After using Claude Code exclusively for a week, I’ve decided to part ways with Cursor. It doesn’t mean I won’t revisit. I love the competition and if they’ll keep going with Grok, it will catch up. The hard thing for Cursor seems to be them getting side tracked with features engineers don’t need. I don’t want to be locked into someone’s ecosystem for anything. And with how good AI is getting, I don’t have to be. I already have my own rulesets and workflow that I’ve created and use daily. If I need something else, I just build it.
Cursor: Lean into Xai! Make the Grok model better (or at least on par with!) than Claude and I will come running back. Cursor, nearly two years ago, you turned this engineer from an AI skeptic to a full time AI user. We deserve better.
It makes sense. Their mistake was advertising a dollar amount of tokens in the first place. Initially, it was a good way to attract a user base, but (and this is pure speculation), as 3rd party providers upped their pricing agreements, Cursor would have been forced to increase price or to transparently disclose that 3rd party model token grants would be reduced. This may have all been coming to a head in May/June… Cursor was never going to survive once the bigger players (Anthropic, OpenAI) simply “caught up”, unless they had a frontier model of their own.
Enter xAI who needed users AND a shot in the arm with coding tools/harness. Now the race is on: can Grok catch up and offer competitive price/perf?
Same problem. I bought Ultra on October 1, 2026 — six days ago — for $200/month.
My usage export for October 1–6, priced at Cursor’s own published Opus 5.5 rates ($4 / $5 / $0.20 / $20 per million), comes to $29.69 for Claude Opus 5.5 Max. The dashboard calls that 23.5% of the Other Models pool. $29.69 / 0.235 = $126. A second line, Opus 5.5 High at $1.22 and 1.0%, gives $122. Against a $400 pool the Max usage would be 7.4%, not 23.5%.
So the allowance I am actually getting is about a third of what it used to be. No email, no changelog, and the dollar column is gone, so the only way to notice is to rebuild the bill by hand. I am extremely disappointed: a shadow cut of the limit by 3x, applied to a subscription purchased this month.
Another Ultra user here, subscribed since April.
My plan renewed yesterday, and just a few queries using Claude and GPT consumed over 20% of my monthly allowance.
I was shocked and initially thought it was a bug, until I found this thread.
I think Cursor owes its Ultra subscribers a clear explanation. What’s the point of paying for Ultra if it ends up being more expensive than on-demand usage?
Paid Ultra, billed by Cursor directly at $200/month. This is not a SuperGrok-linked or complimentary plan.
This cycle started on 2026-10-03. Other Models is already at 100%. At Cursor’s published API rates, the third-party requests that filled it total $125.37: Claude Opus 5.5 xhigh $117.63, GPT-5.6 Sol $6.13, and Claude Opus 5.5 medium $1.62.
The account API still returns an included limit of $400. The Other Models percentage is not measured against that $400. It matches a $125 pool. Total included spend divided by ($3,000 Cursor Models + $125 Other Models) equals totalPercentUsed exactly.
Through August, the documented Ultra Other Models allowance was $400. On this full-price renewal it is $125. The current pricing page only says “Included,” so the cut is not stated there. Cutting the third-party allowance from $400 to $125 on the same $200 plan is a serious downgrade.
Is $125 the intended Other Models allowance for paid Ultra, and when did it change from $400?