What a $20 Claude Pro plan buys across the 5.5 lineup
Saturated-use token allowances for Opus 5.5, Sonnet 5.5, and Haiku 5.5, rebuilt from public rate cards and measured usage. One formula, two anchor measurements, and every assumption left on the record.
Opus 5.5
3.05B
tokens per month on the $1,280 anchor
Sonnet 5.5, post-cut
6.11B
after cache reads were halved on October 7
Haiku 5.5, short prompts
83.5B
prompts at or under 100k tokens
These are ceilings estimated from saturated use, not quotas published by Anthropic. Two independent pool measurements disagree by about 35%; this page leads with the lower one and shows both in the results.
One formula explains every number
A subscription is treated as buying a fixed dollar amount of list-price usage. Divide that pool by a model's blended price per million tokens, and the result is the allowance estimate used on this page.
The weights come from measured Claude workloads: agentic coding re-sends large context, so about 97% of billed tokens are cache reads, 2.5% are new content written into cache, and 0.5% is model output. Fresh uncached input sits at 0.1% or less and folds into the write share.
Your own mix will differ. Because cache reads dominate, cache-read prices move the estimate the most. Sonnet 5.5's cache-read cut on October 7 is the clearest recent example.
Two pool measurements, about 35% apart
The pool is the soft spot in every figure. Two samples of saturated Pro usage exist, both built from public usage panels by the Real API Pricing project. They disagree, and the project still treats the reconciliation as open.
| Anchor | Derived from | Monthly pool | Token multiplier |
|---|---|---|---|
| $1,280 (used here) | Opus 5.5 sample: merged panels at $3.20 per weekly percentage point | $1,279.63 | 1.00× |
| $1,984 | Opus 5 sample: two accounts, September 14 to 22, at $4.96 per point | $1,983.82 | 1.55× |
The two pool measurements on one scale
This report leads with the lower, newer measurement because it is the conservative basis for planning. The results table shows both columns, so no figure depends on that choice alone.
Why they differ: the samples measure different model generations in different months, and panel readings are recorded to whole percentage points. The gap is larger than any other uncertainty in the model.
The 5.5 lineup on $20 per month
Six rows cover the family as it shipped: Opus 5.5, Sonnet 5.5 on both sides of the October 7 cache-read cut, and Haiku 5.5 on both sides of its prompt-length threshold, plus a half-and-half mix for a workload that crosses it.
Tokens per month on Claude Pro, $1,280 anchor
| Model and tier | Blended $/MTok | Tokens/month, $1,280 | Tokens/month, $1,984 | Effective $/MTok |
|---|---|---|---|---|
| Opus 5.5 | 0.419 | 3.05B | 4.74B | 0.00655 |
| Sonnet 5.5, pre-cut | 0.3065 | 4.17B | 6.47B | 0.00479 |
| Sonnet 5.5, post-cut | 0.2095 | 6.11B | 9.47B | 0.00327 |
| Haiku 5.5, over 100k | 0.076625 | 16.7B | 25.9B | 0.00120 |
| Haiku 5.5, 50/50 mix | 0.045975 | 27.8B | 43.1B | 0.00072 |
| Haiku 5.5, up to 100k | 0.015325 | 83.5B | 129.4B | 0.00024 |
- Pre-cut and post-cut refer to Sonnet 5.5 cache reads: $0.20 per million tokens through October 6, $0.10 from October 7. The cut alone raised the token estimate about 46%.
- Haiku 5.5 is priced by prompt length. Up to 100k tokens it charges $0.10 input, $0.50 output, and $0.01 cache reads; past 100k the same prompt pays five times those rates. Anthropic says roughly 90% of requests to the previous Haiku were short.
- The 50/50 mix assumes half of usage on each side of the threshold, so its blend is the average of the two Haiku blends.
The rate cards behind the numbers
All four cards are published Anthropic rates, per million tokens. Cache writes are the 5-minute tier; batch processing and prompt-caching discounts can move individual requests further.
| Model and tier | Input | Output | Cache write, 5 min | Cache read |
|---|---|---|---|---|
| Opus 5.5 | $4.00 | $20.00 | $5.00 | $0.20 |
| Sonnet 5.5 | $2.00 | $10.00 | $2.50 | $0.10 |
| Haiku 5.5, up to 100k | $0.10 | $0.50 | $0.125 | $0.01 |
| Haiku 5.5, over 100k | $0.50 | $2.50 | $0.625 | $0.05 |
Each blend applies the mix to its card. Haiku's short-prompt card, for example: 0.97 × $0.01 + 0.025 × $0.125 + 0.005 × $0.50 = $0.015325 per MTok, which converts a $1,280 pool into 83.5 billion tokens.
What the tokens are made of
Every total carries the same composition, just scaled. Cache reads do almost all of the counting; output is half a percent; uncached input is near zero.
| Model and tier | Total | Cached reads | Cache writes | Output |
|---|---|---|---|---|
| Opus 5.5 | 3.05B | 2.96B | 76M | 15M |
| Sonnet 5.5, pre-cut | 4.17B | 4.05B | 104M | 21M |
| Sonnet 5.5, post-cut | 6.11B | 5.92B | 153M | 31M |
| Haiku 5.5, over 100k | 16.7B | 16.2B | 417M | 83M |
| Haiku 5.5, 50/50 mix | 27.8B | 27.0B | 696M | 139M |
| Haiku 5.5, up to 100k | 83.5B | 81.0B | 2.09B | 418M |
Two recorded panels keep the convention honest: measured cache writes ran 1.3% to 2.0% of tokens and output ran 0.4% to 0.8%, with uncached input at or under 0.14%. The 2.5% and 0.5% weights sit close to that evidence rather than being invented.
Reading the meter: sessions and percentages
Plans meter saturated use as a percentage of a weekly pool, and a month is four weeks. That gives two useful conversions: divide a monthly figure by four for the weekly pool, and by 400 for the tokens behind one percent of the weekly meter.
| Model and tier | Per month | Per week | Tokens per 1% of the weekly meter |
|---|---|---|---|
| Opus 5.5 | 3.05B | 0.76B | 7.6M |
| Sonnet 5.5, pre-cut | 4.17B | 1.04B | 10.4M |
| Sonnet 5.5, post-cut | 6.11B | 1.53B | 15.3M |
| Haiku 5.5, over 100k | 16.7B | 4.17B | 41.7M |
| Haiku 5.5, 50/50 mix | 27.8B | 6.96B | 69.6M |
| Haiku 5.5, up to 100k | 83.5B | 20.9B | 208.7M |
One recorded session shows the shape of real use. A 62-minute Opus 5.5 run at high effort logged 1.03B tokens total and consumed about 75% of a five-hour window on its plan:
- Cache reads: 1.00B, or 97.0% of the session
- Cache writes: 21.1M, or 2.0%
- Output: 8.0M, or 0.8%
- Uncached input: 1.4M, or 0.1%
A session that moves the weekly meter by 10% is therefore about 76M tokens on Opus 5.5, or about 2.1B on Haiku 5.5 at short prompts.
What changed, and when
| Date | Change | Effect here |
|---|---|---|
| September 22 | Opus 5.5 ships at $4 / $20 with $0.20 cache reads; Pro, Max, and Team five-hour limits rise. | Creates the newer pool measurement, $1,280 per month. |
| September 28 | Sonnet 5.5 ships at $2 / $10, the same card as Sonnet 5, needing fewer tokens per typical task. | Adds the Sonnet 5.5 rows. |
| October 7 | Haiku 5.5 ships at $0.10 / $0.50 for short prompts; Sonnet 5.5 cache reads are halved to $0.10. | Raises the Sonnet 5.5 estimate about 46%; adds the Haiku rows. |
The Sonnet 5.5 cache-read cut, in tokens
What would change the answer
- Saturated use is a ceiling, not a quota. Anthropic does not publish token allowances; these estimates describe a price floor.
- The anchor gap: on the higher pool, multiply every figure by 1.55.
- Effort setting matters more than the model name. Independent testing put Sonnet 5.5 at about $0.59 per index task at medium effort and $7.60 at max effort, above Sonnet 5's $5.09.
- Haiku's threshold is a cliff, not a slope: one request past 100k tokens pays five times rates on the whole prompt.
- Plans pool usage across models and features, so a mixed workload lands between rows, never at their sum.
- Token counts are not comparable across generations. Models from Claude 4.7 onward count roughly 30% more tokens for the same text than earlier models.
- The mix is a measured convention. A lower share of cache reads raises the blend and shrinks the allowance.
Sources
- Real API Pricing: methodology and interactive data, and the open dataset and decision log.
- Anthropic: Introducing Claude Opus 5.5, Claude Sonnet 5.5, and Claude Haiku 5.5.
- Anthropic: Claude pricing documentation, the rate cards used in every blend.
- Claude (@claudeai): the October 7 post on halving Sonnet 5.5 cache reads.