Signals
Is Claude Ultra Worth It? Theo Says $200 Buys $14,662 of Opus 5.5 but Only $3,021 of Fable 5.2. We Checked the Math, What the Number Is Made Of, and Who Should Actually Pay
By the Top5Apps editorial team · Published November 10, 2026 · Updated November 10, 2026 · 15 min read
Short answer: against an API key, yes, by a mile. Against the $100 Max plan, only if you run out of week or need to work past the five-hour limit. And if you want it for Fable 5.2, know that half your week is the most Fable you can have.
On the evening of November 9, hours after Fable 5.2 shipped, Theo Browne (@theo) posted that the $200 Claude Ultra plan 'gets you $15,000 of Opus 5.5 in a month,' but only around $3,000 of Fable. The post has 374,000 views. We rebuilt his arithmetic from his own tables and Anthropic's price list, and it holds. Then we looked at what the number is made of, what Anthropic's plan pages actually promise, and who the second $100 is for. (Disclosure, per The Receipts Standard: this site is produced with Claude-family tools.)

Theo - t3.gg@theo
The $200 Claude Ultra plan gets you $15,000 of Opus 5.5 in a month, but only around $3,000 of Fable 5.5. Even ignoring that Fable is restricted to 50% of your weekly limit, it burns faster per $ than Opus does. More info below
View original post ↗
The short version
- The claim: one $200 Ultra account, emptied every week, is worth $14,662 a month at API list prices if you use only Opus 5.5, and $3,021 if you use only Fable 5.2. Measured by Theo from 1.1 million of his own Claude Code requests.
- The math checks. Every headline figure rebuilds from his tables. His token counts for a full Fable 5.2 limit, priced at Anthropic's list rates, come to $703 against his $705.
- 60% of the Opus figure is cache reads. $2,060 of the $3,421 week is 10.3 billion tokens of Claude re-reading context it has already seen. Output is 16%. The number is real at list price, and it is the price of working as if tokens were free.
- Fable is capped at half the week, and Anthropic says so. Its pricing page lists Fable on Max and Ultra at '50% of weekly limits.' Per output token, Fable 5.2 used the week 2.9 times as fast as Opus 5.5. A Fable-only week was about 300 turns.
- Every plan gives the same usage per dollar. Max is 5x Pro a week for $100. Ultra is 10x for $200. What Ultra adds is no five-hour limit for verified accounts and $200 a month in API credits.
- The caveats: one account; Claude Code only; Fable 5.2 had been public for six hours and its row is projected from 47% of one limit. Theo runs T3 Chat, which sells access to Claude models.
- Our read: worth it if you hit Max's weekly limit or need runs longer than five hours. Not worth it for Fable alone, and not for typical use.
What Theo measured
Theo Browne is the CEO of T3 Chat and T3 Code and has 426,000 followers on X. His method is simple and unusually good for a viral post. 'I obsessively log all the requests I make in claude code,' he wrote, and the charts are built from 1.1 million of those requests, captured through a proxy between August 11 and November 9 on one $200 account. The method, as far as the charts show it: for each stretch of work he has the tokens that went through and the share of the weekly limit it used up. Price the tokens at Anthropic's API rates, divide by the share of the limit, and you get what a full week is worth.
| If you spend the week on | API value, one week | API value, 30 days | Turns a week | Threads a week | Share of week used |
|---|---|---|---|---|---|
| Opus 5.5 only | $3,421 | $14,662 | 1,218 | 230 | 100% |
| Fable 5.2 first, then Opus 5.5 | $2,416 | $10,353 | 909 | 247 | 100% |
| Fable 5.1 first, then Opus 5.5 | $2,223 | $9,528 | 700 | 180 | 100% |
| Fable 5.2 only | $705 | $3,021 | 300 | 132 | 50% |
| Fable 5.1 only | $513 | $2,197 | 90 | 64 | 50% |
One slip to clear up first. The post says 'Fable 5.5.' The chart under it says Fable 5.2, and nine minutes later he corrected himself: 'Fable 5.2*. Man this is going to be SO f*ing annoying.' There is no Fable 5.5. Anthropic's newest models are Opus 5.5, Sonnet 5.5, Haiku 5.5 and Fable 5.2, and it is easy to see how the numbers blur.

Does the arithmetic hold? Yes
We rebuilt every headline number from his own table, and they all close.
- The month. Thirty days is 30 ÷ 7 = 4.29 weekly limits. $3,421 × 4.29 = $14,661. His chart says $14,662. Fable 5.2: $705 × 4.29 = $3,021, exactly.
- The mixed week. Empty the Fable limit ($705), then spend the other half of the week on Opus ($3,421 ÷ 2 = $1,710.50). Total $2,415.50. His table says $2,416.
- The half. He says Fable 5.2 uses 10.1% of the week per million output tokens, and a Fable limit holds 4.9 million. 10.1 × 4.9 = 49.5%. That is the 50% cap Anthropic publishes, recovered from two numbers on different charts.
- The dollars. His table gives the tokens in a full Fable 5.2 limit: 4.9 million output, 20.8 million cache-write, 869.2 million cache-read. At Anthropic's list prices of $40, $16 and $0.20 a million, that is $196 + $333 + $174 = $703. He reports $705. The Opus week lands between $3,218 and $3,575 depending on which of two cache-write rates applied; his $3,421 sits inside.
So the dollar figures are what his token counts cost at list price. What nobody outside can check is the token counts themselves and the percent-of-limit readings. The logs are his. The math is sound; the sample is one person.
What the $14,662 is made of
This is the part the headline hides. Take the full Opus week of $3,421 and split it by token type. $2,060 of it, 60%, is cache reads. That is 10.3 billion tokens of Claude re-reading the conversation and files it has already seen, at $0.20 a million. Output, the code and prose Claude actually wrote, is $564, or 16%. The rest is cache writes and fresh input.
Put another way: for every token Opus wrote that week, it re-read 365. That is how an agent works. Every step of a long task sends the whole thread back through the model, and a long Claude Code session is hundreds of steps.
It isn't a trick. Anthropic bills cache reads on the API, so a key doing identical traffic would get this bill. The catch is 'identical.' On a meter, people compact long threads, start fresh ones, and hand small jobs to cheaper models. A flat plan removes every reason to. $14,662 is the list price of how a heavy user works when the next token is free. Nobody would run that traffic on a key, so nobody is really being saved that much.
And list price is not cost. One reply asked how Anthropic could be profitable at this rate. Theo's answer: 'Because the api prices are insanely inflated lol.' That is his opinion, and Anthropic doesn't publish what a token costs it to serve. But it is the right caution. 'Worth $14,662' means 'at list.'
For scale, 1,218 turns a week is 174 finished replies a day, seven days a week, counting the ones subagents send. That is what it takes to empty an Ultra week on Opus.
How far to trust it
Theo was careful to say what the data can't do. 'These are still early numbers,' he wrote. His own footnotes, and a few things we would add:
- One account, one person's work. A different mix of long and short threads would give a different dollar figure for the same limit. The cache-read share above is why.
- Six hours of Fable 5.2. Anthropic announced the model at 18:57 UTC on November 9. The thread went up at 01:09 UTC, 6 hours and 11 minutes later. The Fable 5.2 row is, in his chart's words, 'projected from 47% of one limit, about ±10%.' The Fable 5.1 row is the median of six emptied limits and is the firmer number.
- Different work on each model. He notes that 'a ton of my usage has been benchmarking the new model, so this might change over time.' The chart says the 5.2 and 5.1 sessions 'were different work' and that for the same work you should expect a gain of 1.3 to 1.5 times.
- Claude Code only. These are coding-agent requests. Chat, research and file work in the Claude apps draw on the same weekly limit but weren't measured.
- Two names, one plan. Some charts are labelled 'Claude Ultra' and some 'Claude Max 20x.' It is the same $200 plan, renamed on October 30. Most of the data was logged under the old name.
- Who is measuring. T3 Chat sells access to AI models, Claude's among them, so Theo competes with the product he is pricing. Nothing in the thread reads as a pitch, and the finding flatters Anthropic's plan more than it hurts it. Readers should still know.
Against all that: Anthropic doesn't publish its limits in tokens or dollars. It describes Ultra as '10x Pro usage every week' and leaves the rest to a usage bar. Until Anthropic says more, a careful log from one heavy user is the best public measurement there is.
The Fable catch

If you are thinking of Ultra because you want Anthropic's best model all week, this is the section that matters.
Anthropic says so in plain words, though not on the Ultra page itself, which mentions a separate Fable limit and stops there. The pricing page lists Fable on Max and Ultra at '50% of weekly limits,' and the Fable help article says you can use 'up to 50% of your weekly usage limits on Fable models at no extra cost.' Theo's logs agree: 'Fable can only be used for 50% of your weekly limit.' Use nothing but Fable and the other half of your week goes unspent. Past the cap you can 'keep using Fable models with usage credits, or switch to another Claude model,' and usage credits are 'billed at standard API rates.'
Inside that half, Fable burns fast. A million output tokens of Fable 5.2, with the cache traffic that comes with them, used 10.1% of his week. The same on Opus 5.5 used 3.5%. That is 2.9 times as fast. On the price list Fable 5.2 is only twice Opus ($8 and $40 a million against $4 and $20), and cache reads cost the same $0.20 on both. The plan charges Fable more heavily than the price list does. By our division, 1% of an Ultra week buys $34.21 of Opus at list and $14.10 of Fable 5.2.
In work terms: a week of Fable 5.2 alone came to 300 turns, about 43 a day, across 132 sessions. Then Fable is gone until the reset, with half a week of Opus still in the tank. That is a real improvement on Fable 5.1, which managed 90 turns from the same limit. It is still a ration.
Theo - t3.gg@theo
The best way to measure IMO is "work done per week" If I exclusively used Opus, I'd be able to do ~1,200 turns and 230 threads in a given week If I exhaust my Fable limit then move to Opus, I get ~900 turns but slightly more threads? This is partly because, from my experience, Fable is more likely to "finish" work in a single prompt.
View original post ↗
That last line is the fair counterweight. Fewer turns isn't less work if each turn finishes more, and by his count the Fable-first week closed slightly more sessions (247 against 230), though his own footnote says a batch of short benchmark sessions pushed that count up. One of the most-liked replies drew the practical conclusion anyway. 'Opus 5.5 is good enough,' it said. 'Not seeing myself use fable other than review and hard debug.' On these numbers that is the sensible way to spend an Ultra week: Opus for the volume, Fable for the hard parts.
It also puts a number on yesterday's story. The '$642 iceberg' run we checked would be 91% of one week's Fable 5.2 limit on Ultra by Theo's measure, and that poster's four runs together, $798.51 at list, would be 113%. One subscription could not have covered them in a week.
What Anthropic says you get
| Pro | Max | Ultra | |
|---|---|---|---|
| Price | $20 a month, or $200 a year | $100 a month | $200 a month |
| Usage | The baseline | 5x Pro, every five hours and every week | 10x Pro, every week |
| Five-hour limit | Yes | Yes | None for verified accounts |
| Fable models | Paid with usage credits (included up to 50% of the week during launch week, Nov 9 to 16) | Included, up to 50% of the weekly limit | Included, up to 50% of the weekly limit |
| Monthly API credits | None | $100 | $200 |
| Annual billing | Yes | No | No |
Three things in that table are easy to miss.
- The old name no longer fit. When Max launched in April 2025, the $200 tier was sold as '20x more usage than Pro.' The weekly figure Anthropic publishes today is 10x. With the five-hour limit gone for verified accounts, 10x a week is the only multiple left.
- The API credits equal the price. Ultra includes $200 a month in API credits, and Max includes $100. They are for building on the Claude Platform. They can't be spent on interactive Claude Code or on extra usage in the apps, they expire each billing cycle, and you have to claim them by linking a Console account. If you already spend that much on the API for your own projects, the plan is close to free. If you never touch the API, they are worth nothing to you.
- There is no token count anywhere. The pricing page's footnote is the whole specification: 'How fast you reach a limit depends on the model you use.' That gap is why a thread like Theo's gets 374,000 views.
Ultra or Max? That is the real $100 question
Nobody deciding on Ultra is choosing between it and an API key. They are choosing between $200 and $100. Anthropic's own multiples make that comparison simple. Max is five times Pro each week. Ultra is ten times. Twice the price buys twice the week. There is no bulk discount.
Scale Theo's measured week by those multiples and you get the other two plans. This is our arithmetic, not his and not Anthropic's: a Max week would be worth about $1,710 of Opus 5.5 at list, or $7,331 over 30 days, and a Pro week about $342, or $1,466. Every plan returns roughly 73 times its price. Ultra is not a better deal per dollar than the plan below it. It is a bigger tank.
So what does the second $100 buy besides volume? Two things.
- No five-hour limit. Max still meters you in five-hour windows. A verified Ultra account has only the weekly limits. If you run long unattended jobs, overnight agents or the multi-agent workflows we covered yesterday, that is the difference between a job that finishes and one that pauses until the window resets.
- Another $100 of API credits, if you would use them.
There is one independent check on the scale of all this. In June, SemiAnalysis bought every Anthropic and OpenAI plan and 'randomly ran long horizon coding tasks until we exhausted the weekly limit.' It noted that a $200 plan was 'widely believed' to max out near $2,000 a month at API prices, and its table put the real ceiling for Claude's $200 tier at about $8,000. Theo's figure, five months later on a newer Opus, is $14,662. The methods and models differ, so the two numbers can't be compared closely. They agree on the point that matters: these plans return dozens of times their price at list.
Who should actually pay $200
- You hit Max's weekly limit most weeks: yes. The alternative is usage credits at API rates, and by Theo's numbers a tenth of an Ultra week of overflow would cost $342 at list. The upgrade is $100.
- You run jobs longer than five hours: yes, for that alone. Check that the account is verified first. In the Claude apps, open Settings, then Usage. Anthropic's help page says a verified account shows no 'Current session' bar.
- You want Fable 5.2 as your daily model: not for that reason. Half the week is the most Fable either plan gives you. Ultra's half is twice the size of Max's: about 300 turns a week on Theo's log, so something like 150 on Max by the same scaling. If that difference decides it, fine. Plan to do most of your work in Opus 5.5 either way.
- You are a typical developer: probably not. Anthropic's Claude Code cost page says that across enterprise deployments the average is 'around $13 per developer per active day,' and under $30 for 90% of users. Five days at the average is $65. That is 1.9% of Theo's Ultra week. Five days at $30 is $150, or 4.4%. Weekly volume is not what most people run out of.
- You pay by API key today: move to a plan, but not necessarily this one. The same page puts the average developer at '$150-250 per developer per month' on the API. Max costs less than that and, on the scaling above, holds about 30 times as much. Keep the key for products and automation, which plans don't cover.
- You mostly chat: there is no evidence here. Everything Theo measured is Claude Code. Start on Pro and move up when the usage bar tells you to.
The test costs nothing. Spend two weeks on Max and watch two things.
Before you upgrade
# in Claude Code: how much of your plan is left
/status
# what the session would have cost at API list price
/usage
# in the Claude apps
Settings > UsageIf the weekly bar reaches 100% before it resets, or a five-hour stop breaks work you care about, Ultra pays for itself. If neither happens, it doesn't. Each of those is a real reason to upgrade. 'It is worth $14,662' is not one.
The risk: a deal this good can be repriced
A plan that returns 73 times its price at list is either a large subsidy, a sign that list prices sit far above cost, or both. In each case Anthropic holds the dial, and it has turned it before, in both directions.
In July 2025, three months after Max launched, Anthropic announced weekly limits for Pro and Max, saying they would 'apply to less than 5% of subscribers.' Its thread gave the reasons: people running Claude Code 'continuously in the background, 24/7,' account sharing and reselling, and a case where 'one user consumed tens of thousands in model usage on a $200 plan.' This October it moved the other way and removed the five-hour limit for verified Ultra accounts.
Jack Vinijtrongjit@juminoz
@theo So based on this, I'm currently spending $105,572/mo in equivalent API credits across Claude and ChatGPT. They better not end the subscriptions any time soon or there is no way this is sustainable.
View original post ↗
That reply, with a figure that is the poster's own, is the honest reaction to the chart. If your workflow only makes sense at 1% of list price, the price is a dependency. Two things limit the exposure. Max and Ultra are monthly only, so you are never locked in. And every major lab now sells a tier at this price: OpenAI has ChatGPT Pro at $100, $200 and $500, Google has AI Ultra at $199.99, and xAI has an Ultra plan at $200. SemiAnalysis's June table put ChatGPT's $200 tier at about $14,000 of API value. The flat $200 plan is how the industry sells heavy use, and the labs are competing on how much it holds.
Our read
Claude Ultra is worth it for people who run out of Max, and for almost nobody else. Theo's math is right, his number is real at list price, and his thread is the best public measurement of what the plan holds. But $14,662 is the size of the tank, not the amount you will save. Three fifths of it is the model re-reading its own context. The Fable half is a ration. And every Claude plan gives the same usage per dollar, so the question is never whether Ultra is good value. It is whether you need the volume or the unbroken hours.
Theo deserves credit for the method, for footnoting his own charts, and for saying 'early numbers' in the thread instead of leaving readers to find out. What would settle the rest is Anthropic publishing its limits in tokens, or a second and third account repeating the measurement. We will update this page if either happens. Rankings: no movement. Claude and Claude Code keep their places, and both reviews will pick up the renamed plans at the next scheduled refresh. To price your own prompts at the rates used here, the Token Counter carries Opus 5.5 and Fable 5.2.
Where these apps rank
