How many tokens does a subscription buy?
For 31 days we recorded every request AGNT sent on one $200 Claude Max plan: 38,805 requests and 15.87 billion tokens. A plan is a spending budget, not a token allowance, so the tokens it buys depend on the tool. At Claude Opus 5.5 prices that budget tops out near 97.5 billion tokens a month, for any tool. AGNT's real month works out to 42.3 billion on a full plan, and its long uninterrupted runs reach 63.7 billion. Without AGNT's caching and compression, the same month of work would have needed 4 plans instead of 1.
What a $200 plan buys in a month
Billions of tokens, at Claude Opus 5.5 prices
The setup
One account, one plan, 31 days, every request.
AGNT logs each request it sends to the model: how many tokens were read from the prompt cache, how many were written to it, and how many came back as output. This report uses every request from one heavy user's account, a person working in AGNT all day, between October 18 and November 17, 2026.
The account never reached a limit. Its busiest week used 53% of the weekly budget, so the per-plan figures here are AGNT's measured mix of tokens scaled to a full plan, not a plan run to its limit.
# what was recorded planClaude Max, $200 a month windowOct 18 to Nov 17, 2026 requests38,805 in 287 conversations tokens15.87 billion modelClaude Opus 5.5, 95.4% of tokens limit errorsnone
Where the tokens go, and where the plan goes
AGNT's recorded month, by kind of token. Select a kind to trace it.
Share of plan usage
The price sheet: Claude Opus 5.5, per million tokens
- $0.20cache read
- $4plain input
- $55-minute cache write
- $81-hour cache write
- $20output
List prices read on November 17, 2026. Colours match the chart above. AGNT's cache writes are almost all at the 1-hour price, and the two grey prices matter for other tools. Every request in this report is priced on this one sheet, whichever model served it.
A plan is a budget
The cheaper the average token, the more tokens a plan buys.
Anthropic shows a plan's limits as percentages and never as a number. We treat the plan as a budget that each request draws down in proportion to what it would cost on the API. The tokens a plan buys in a month are then the budget divided by the average price of the tokens a tool sends.
the ceiling $19,507 ÷ $0.200 per million = 97.5 billion AGNT's month $19,507 ÷ $0.461 per million = 42.3 billion
Try it: what a plan buys at any mix of tokens
Start from a measured mix, then move the controls.
billion tokens a month
- Average price per million tokens
- Output tokens a month
- The $200 plan, per million tokens
Where AGNT sits
In real use AGNT is ahead of Claude Code, and too close to call against the best of the rest.
One clean session is any tool's best case: it starts once, runs without a break and ends before anything expires. A month of real use has breaks, expired caches and rebuilt context. Setting one tool's best case beside another's ordinary month is not a comparison, so we pair real use with real use and clean with clean.
One $200 plan, compared like for like
AGNT in green, Claude Code in dark grey, other tools in light grey. The dashed line is the ceiling for any tool. Select a row to see how it was measured.
An independent check: seven weeks of Claude Code
The best outside evidence for Claude Code is a run we did not make. Horizon Labs, a research group, let Claude Code work on its own for seven weeks on Claude Opus 5.5 and one Team Premium seat, and published the whole transcript. We totalled its 4,055 requests on the same price sheet.
| On the same price sheet | AGNT, recorded month | Claude Code, 7-week run |
|---|---|---|
| Requests | 38,805 | 4,055 |
| Tokens processed | 15.87B | 1.91B |
| Input read from the cache | 97.76% | 96.38% |
| Saved by caching, as recorded | 88.7% | 87.2% |
| Tokens a month on a full $200 plan | 42.3B | 37.9B |
| Output tokens per plan dollar | 9,520 | 3,283 |
| Usage that is reads, writes, output | 42% / 39% / 19% | 37% / 56% / 7% |
Saved by caching compares each record with the same tokens at the plain input price, on this price sheet.
The two agree on caching. Scaled to a $200 plan, that run is the highest of our four real-use estimates for Claude Code, and AGNT's month is 12% above it.
They differ on where the usage goes. In that run, writing to the cache is 56% of usage, against 39% for AGNT. 50% of the run's usage is the extra cost of writing context a second time, usually after a pause: 93 of its 94 pauses longer than an hour missed the cache.
The run was on a different plan and did a different kind of work: an agent that mostly waits, against a person working all day. So read the gap in output per dollar as workload as much as tool.
Lost caches
More than a fifth of the month's usage went to rebuilding lost caches.
With its cache intact, AGNT's traffic costs $0.353 per million tokens. Over the whole month it averaged $0.461. The gap is the price of writing context to the cache again after an entry is lost: about 22% of the month's usage, plus 1.6% for the first request of each conversation, which has to write its prompt once.
What the month's usage paid for
Shares of the month's $7,309 of usage at list prices. Select a part. Shares are rounded, so they can add to a little under the whole.
A fix for the largest loss was implemented and verified offline on November 17, 2026, after the recording ended. The replays in the next two sections use the corrected cut.
Caching and compression
Switch them off and the same month needs 4 plans instead of 1.
We replayed the recorded month with AGNT's own mechanisms switched off. Caching keeps the prompt stable so the model re-reads it at the cheap rate. Compression caps oversized tool results, moves large ones out of the prompt and lets stale ones age out. Flip the switches.
The same month of work, with AGNT's mechanisms on or off
Each tile is one $200 plan. The fill is how much of that plan the work uses.
- Plans needed
- Subscriptions a month
- Usage, as % of one plan
Caching cuts plan usage by 88.5%, compression by 17.5%, and the two together by 90.3%. The two overlap, so the figures do not add up. None of them depends on how large the plan's budget is. The count of plans does.
A second method agrees on caching. The account's usage ledger, with no replay involved, shows the recorded cache re-use making the month 88.4% cheaper than the same tokens at full price.
This compares AGNT with its own mechanisms switched off. Other tools cache too, and the comparison with them is in section 03.
Tokens are not work
Carry less context and a plan does more work with fewer tokens.
A plan's token count rises when a tool re-reads more and falls when it gets leaner, which makes it a poor score on its own. Cut a long conversation back harder and each request carries less. The plan then buys more requests and more output, with fewer tokens. Drag the slider to cut deeper.
Tokens a month against output per plan dollar
AGNT in green, Claude Code in dark grey, other tools in light grey. The dotted line is AGNT as it cuts deeper. No tool can pass the dashed ceiling.
- Tokens a month
- Requests a plan buys
- Output tokens per plan dollar
- Mean context, all requests
Most conversations never grow long enough to be cut, which is why the mean context across all requests sits below the amount a long conversation keeps.
The paper also derives the cheapest point to cut for a given amount kept. With AGNT's measured numbers, 3,385 new tokens a request and 455K kept, it is 806K. AGNT's cut was observed to trigger at 750K.
Limitations
The budget is a floor. A higher real limit lifts every token count together.
Every token count here scales with one number we cannot measure directly: the plan's budget. We use the lowest value consistent with one public report. Move the slider to see what a higher budget changes, and what it leaves alone.
If the plan's real budget is higher than the floor
Token counts scale with it. Ratios and percentages do not.
- Does not move
- 90.3% less usage with caching and compression
- 7% to 27% ahead in clean sessions
- 1.1 to 2.2 times Claude Code in real use
- Weekly budget
- Ceiling, any tool
- AGNT, real month
- Without caching or compression
- The budget rests on one report.A single public account of 21.4 billion tokens at 95% of the weekly limit. We could not verify it.
- How Anthropic meters AGNT is not measured.We assume usage is counted in proportion to API price. Public measurements say requests through the Agent SDK use the five-hour window two to three times faster. Where AGNT's traffic falls is unknown.
- AGNT never reached the limit.Its busiest week used 53% of the weekly budget, so the per-plan figures are its measured mix scaled up.
- The other tools' numbers come from mixed sources.One clean session each for Claude Code and the Agent SDK. Router averages of API-key traffic, converted from a cache-hit rate. Older logs. One autonomous run on a different plan.
- One account, one month.A single heavy user's mix of work. Other work will have different shares of new content, output and idle time.
- The switch-off results are a replay.The replay matches the record within 2.4% on conversations that never reach a cut. The corrected cut had not yet run in production. As recorded, the month used 37% of a plan, not 33%.
Paper and data
The paper has the definitions, the proofs and all eight tables.
This page and the paper are generated from the same computed figures. The data file holds every figure shown here, with how each was measured. It does not include the raw requests. All figures are as of November 17, 2026.
- how-many-tokens-does-a-subscription-buy.pdf
The paper, 14 pages. - how-many-tokens-does-a-subscription-buy.tex
Its source, which compiles on its own. - token-yield-data.json
Every number on this page. - SHA256SUMS.txt
SHA-256 checksums of the source, the data file and the README.
Related research: The Cache Wars 2, a controlled measurement of prompt-cache efficiency in six agent harnesses. All AGNT Labs research.
Cite
@techreport{annie2026tokens,
title = {How Many Tokens Does a Subscription Buy? Theoretical Maxima,
a Like-for-Like Comparison of Agent Harnesses, and the
Value of Caching and Compression in AGNT},
author = {Annie},
institution = {AGNT Labs},
year = {2026},
month = nov,
url = {https://agnt.gg/whitepapers/how-many-tokens-does-a-subscription-buy}
}
This is what the product is for
AGNT Labs publishes its working out. The same discipline runs through the product: every agent run produces a receipt showing what it did, what it read, and why — so a result can be checked rather than trusted. AGNT runs on your own machine and is free to use.