Anthropic says Fable 5.1 pricing works out about 25% lower to run, but the saving lands only on the part of your bill that comes from cache reads.
- Fable 5.1 pricing cut one rate, cache reads, from $1 to $0.25 per million tokens. Input and output prices did not move.
- Your saving is roughly 75% of your current cache-read share. A bill with little caching barely changes.
- Seat subscriptions sit outside the cut as announced. It applies where usage is billed by token.
Fable 5.1 pricing, announced by Anthropic on September 1, 2026, keeps input at $10 and output at $50 per million tokens and cuts one line: cache reads, now $0.25 per million tokens. Anthropic’s headline figure, an estimated 25% saving for typical workloads, is an average across its own customers. Whether your company sees 25%, 5% or 45% depends on a number most finance teams have never pulled from the usage report.
That number is the share of your Claude bill spent on cache reads. Pull it before you budget the saving.
What moved the money in Fable 5.1 pricing?
One rate moved and everything else stayed put. Anthropic’s launch post says Fable 5.1’s pricing is otherwise the same as Fable 5’s, and the platform release notes put cache reads at 0.025 times the base input price, against 0.1 times on other models. On a $10 input price, that is $0.25 instead of $1.
A cache read is what you pay when the model reuses context it already processed, such as a long system prompt, a contract it has already read, or a codebase an agent keeps returning to. Chat-style use with short prompts creates few of them. Agents that loop over the same files for hours create a lot.
Anthropic is open about where its averages come from. Anthropic’s cost chart draws on a month of real customer traffic from August 2026, run at default effort and covering Claude Enterprise, Claude Code and the API. That is a reasonable sample of Anthropic’s customers. It is not a sample of yours.
Show Me the Invoice
We can work backwards from Anthropic’s own numbers to see what they assume, then forwards to a bill of your own.
| Input | Value | Type |
|---|---|---|
| Cache-read price cut | 75% ($1 to $0.25 per million tokens, USD) | Fact, Anthropic release notes |
| Typical workload saving | About 25% | Vendor claim, Anthropic |
| Highly agentic workload saving | Up to about 45% | Vendor claim, Anthropic |
| Monthly token bill | $5,000 USD | Illustration |
| Cache-read share of that bill | 10% | Illustration |
Step 1, what Anthropic’s figures imply. If only cache reads got cheaper, the total falls by 0.75 times the cache-read share. A 25% total saving means 25% divided by 75%, or about 33% of typical spend was cache reads. A 45% saving means 45% divided by 75%, or 60% for the agentic workload.
Step 2, an illustrative bill. At $5,000 USD a month with 10% in cache reads, $500 goes to cache reads. Multiply $500 by 0.75 and the saving is $375 a month, or 7.5% of the bill. That is well short of 25%.
Sensitivity. If cache reads are 60% of the same bill, the saving becomes $2,250 a month, or 45%. If the bill is ten times larger at the same 10% share, the saving is $3,750, still 7.5%. The share decides the percentage. The size of the bill only decides the dollars. (We re-ran these sums before publishing.)
Who captures the value?
Companies running long, context-heavy agents capture most of it, because cached context is most of what they pay for. Anthropic’s own split puts cache reads at about a third of spend for typical use and three fifths for agentic work. A team sending short, one-off prompts will see a much smaller change.
Companies running long, context-heavy agents capture most of it, because cached context is most of what they pay for. Anthropic’s own split shows the difference: about a third of spend for typical use against three fifths for agentic work. A team that sends short, one-off prompts through the API will see a much smaller change, even though the headline applies to the same model.
Anthropic gains something too. A cheaper cache read makes it less painful to keep huge contexts loaded, which encourages exactly the long agent runs that consume the most tokens overall. Cheaper per unit can still mean a larger monthly total if usage grows (a familiar pattern in metered software).
Why it matters in Canada
Canadian companies buying the Claude API are billed in US dollars, so the saving and the exchange rate both land on the same line of the budget.
The second change in the announcement may matter more to some buyers here: Anthropic says its Enterprise Frontier Safeguards, which keep data in cloud infrastructure the customer controls, will reach enterprise customers in phases beginning later this fall. Until then, the release notes say Fable 5.1 requires 30-day data retention unless Anthropic expressly authorizes zero retention. Anthropic has not said which cloud regions the customer-controlled option will support, so a buyer with Canadian residency requirements should treat that as unknown.
The bull case
The strongest case for budgeting the full 25% is that Anthropic measured real usage, not a lab estimate, and that most serious deployments are moving toward agents.
If your team is shifting work from chat to Claude Code or long-running tools this quarter, your cache-read share will rise, and the saving will rise with it. On that view, the 25% is a floor for where usage is heading. We take that seriously. It is still a forecast about your workload, and only your usage report can confirm it.
What would change our view
If Anthropic publishes the distribution behind its averages, showing most token-billed customers saving close to 25%, our caution about the headline would be wrong for most readers. A second change to input or output prices would also reset this arithmetic.
What should you do this quarter?
Work out your own saving before you budget for one. Export a month of Anthropic usage by line item, divide cache-read spend by total spend, and multiply that share by 0.75. That is your budget figure for Fable 5.1, not the headline 25%. Then divide the new total by finished tasks to see the real cost.
Our earlier work on cost per successful task and the Claude Opus 5 cost per task shows why the next step is to divide the new total by finished tasks, because a cheaper token can still sit inside an expensive task.
If the share comes back under 10%, leave the forecast alone and spend the time on something else. Price moves elsewhere, such as Nvidia’s server price rise, will move your AI costs more. More pieces on spending sit in our AI in Business section, and our guide to AI for business in Canada covers where these costs fit in a wider plan.
Related: if agents run on these models, set a monthly token budget for each AI agent before usage-based bills arrive.
FAQ
What is Fable 5.1 pricing?
Fable 5.1 costs $10 per million input tokens and $50 per million output tokens, the same as Fable 5. Cache reads cost $0.25 per million tokens, 75% less than before.
Will my company save 25% on Fable 5.1?
Only if about a third of your token spend is cache reads. Your saving is roughly 0.75 times your cache-read share of spend.
Do Claude seat plans get cheaper?
Anthropic ties the price cut to usage billed by token. Flat seat prices were not part of the announced change.
