Anthropic Cut Cache Reads 75%. Opus 5 Still Undercuts It.

One line on your invoice now decides which Claude model is cheaper, and it is not the one your router reads. On 1 September 2026 Anthropic shipped Claude Fable 5.1 and left its sticker price alone — $10 per million input tokens and $50 per million output, identical to Fable 5. It cut exactly one rate: cache reads, from $1.00 to $0.25 per million tokens.

That single change inverts a cross-model ranking most enterprises have hard-coded. Fable 5.1 reads from cache at $0.25/MTok while Claude Opus 5 charges $0.50/MTok for the same operation — the flagship is now half the price of the cheaper model on that line, despite a base input rate twice as high. But the inversion only decides your bill above a threshold that Anthropic's own savings estimate does not reach. The number that matters is the share of your spend sitting in cache_read_input_tokens, and almost nobody has it to hand.

---

What Anthropic Actually Changed, and What It Left Alone

Anthropic changed one multiplier and nothing else. Cache hits and refreshes on Fable 5.1 and Mythos 5.1 are priced at 0.025x the base input price, where every other Claude model uses the standard 0.1x. A footnote on the pricing page carries the entire change.

Everything else on the Fable line is untouched. Base input stays $10/MTok. Output stays $50/MTok. The 5-minute cache write stays at $12.50/MTok (1.25x base) and the 1-hour write at $20/MTok (2x base). Batch processing is still $5 in and $25 out. The 512-token minimum cacheable prompt length is unchanged, and so is every rate on Opus 5, Sonnet 5 and Haiku 4.5.

Anthropic estimates the cut reduces typical bills by about 25% and highly agentic ones by up to approximately 45%, and publishes no methodology behind either figure. Independent write-ups have repeated those figures without testing them: VentureBeat noted the $0.25 rate is half Opus 5's cache read price despite a base input price twice as high, and MarkTechPost framed the benefit as accruing to context-heavy agentic workloads. Both are true. Neither tells you whether to move traffic.

Those two percentages are more useful than they look. A 75% cut applied to one line item can only produce a 25% total reduction if that line item was a third of the bill, and can only produce 45% if it was 60% of it. So the estimates imply a cache-read share of the Fable 5 bill even though Anthropic never names one — provided the token mix is held constant. That proviso is doing more work than it looks, and the rest of this article tests it.

---

The Test: Cache Reads Must Exceed a Third of Your Bill

Fable 5.1 beats Opus 5 on price only when cache reads are more than one-third of your Fable 5.1 bill. That is not a rule of thumb. It falls out of the price table exactly, because Opus 5 is priced at precisely half of Fable on every line except one.

Check it against the published table: base input $5 against $10, 5-minute cache write $6.25 against $12.50, 1-hour cache write $10 against $20, output $25 against $50. Every one is a clean 0.5x. Only cache reads break the pattern, and there Opus 5 is 2x Fable 5.1 rather than 0.5x.

So call your Fable 5.1 cache-read spend R, and everything else on the invoice X. Fable 5.1 costs R + X. The identical token mix on Opus 5 costs 2R + 0.5X. Fable 5.1 wins when R + X is less than 2R + 0.5X — which simplifies to R greater than half of X, or cache reads above 33% of the Fable 5.1 total.

Now translate that back into a number you already have. Run the same mix through the old Fable 5 rate and the crossover lands at a 50% saving. If moving from Fable 5 to Fable 5.1 cut your bill by more than half, cache reads are dense enough that Fable 5.1 undercuts Opus 5. If it cut your bill by less than half, Opus 5 is the cheaper model for that workload.

Anthropic's own ceiling is 45%. Its "highly agentic" case implies cache reads at roughly 27% of the new bill — close to the line, and on the wrong side of it, with Opus 5 landing about 9% cheaper on the identical token mix. Its "typical" case puts cache reads near 11%, where Opus 5 is roughly a third cheaper. The crossover is real. It sits just past the edge of the range the vendor claims.

---

Two Bills, One Price List, Opposite Answers

The same price table produces opposite recommendations depending on the shape of the traffic, so here are both, worked end to end.

A long-running agentic coding session. A 200,000-token prefix — repo context, tool definitions, system prompt — written once to the 1-hour cache, then 60 turns that each re-read it, add 1,000 tokens of fresh uncached input and emit 2,000 output tokens.

Line item	Fable 5.1	Opus 5
1h cache write (0.2 MTok)	$4.00	$2.00
Cache reads (12 MTok)	$3.00	$6.00
Uncached input (0.06 MTok)	$0.60	$0.30
Output (0.12 MTok)	$6.00	$3.00
Total	$13.60	$11.30

Opus 5 is 17% cheaper. Cache reads are 22% of the Fable 5.1 bill — under the 33% line. On Fable 5 the same session cost $22.60, so the upgrade saved 40%: a healthy number, comfortably inside Anthropic's agentic band, and still not enough to cross over.

A fan-out over a cached corpus. The same 200,000-token prefix, but 2,000 short extraction calls against it, each contributing 300 tokens of uncached input and 150 tokens of output. Fable 5.1 costs $125.00; Opus 5 costs $212.50. Fable 5.1 is 41% cheaper, cache reads are 80% of its bill, and the move from Fable 5 saved 71%.

Both numbers come off the same price list. The variable is the read-to-output ratio, and nothing in a per-model rate card exposes it.

There is a catch worth stating plainly, because it undercuts the excitement: the workload shape where Fable 5.1 beats Opus 5 is the shape where the mid-tier models beat both. Sonnet 5 reads from cache at $0.20/MTok and Haiku 4.5 at $0.10/MTok — both cheaper than Fable 5.1's new $0.25. That fan-out job costs $85.00 on Sonnet 5 and $42.50 on Haiku 4.5 — and the Haiku column if anything understates the gap, because it still uses the older tokenizer, against which Fable 5.1's produces roughly 30% more tokens for the same text. Fable 5.1's cache read is not the cheapest on the menu; it is the third cheapest, sitting between Sonnet 5 and Opus 5. If your traffic is read-heavy and output-light enough to trigger the crossover, the honest next question is whether you needed a frontier model for it at all. We have argued before that per-token rate cards are the wrong unit for model selection; this release is the cleanest example yet.

---

Fable 5.1's Own Behaviour Pushes Your Mix the Wrong Way

Three documented behaviour changes in Fable 5.1 add output tokens and turns — the line items that did not get cheaper — which moves a workload away from the crossover rather than toward it. Anthropic lists all three in its own migration notes.

The first is batching. Fable 5.1 may issue one tool call per turn where Fable 5 batched several, which Anthropic says shows up in custom coding agents and bash-and-editor harnesses. More turns means more cache reads, but also more output tokens and more round trips at $50/MTok.

The second is file editing. When editing text files, the model is more likely to rewrite an entire file than make a targeted edit — the result is usually the same, but Anthropic notes the rewrite costs more output tokens and time. A one-line change that emits the whole file is billed at the rate that did not move.

The third is effort. The API default is high, and effort affects all tokens in the response including thinking, tool calls and function arguments. Anthropic's migration checklist explicitly says to re-tune effort from that default. Adaptive thinking is always on for Fable 5.1 and cannot be disabled, so the floor under output spend is higher than on a non-thinking model.

None of these are defects. They are trade-offs Anthropic documented honestly, and the capability gains behind them are real — Fable 5.1 scores 52.6% on Terminal-Bench-Science 0.1 against 24.7% for Fable 5 and 29.0% for Opus 5, and 55.8% on Terminal-Bench 4.0 against 42.0% for its predecessor. Those are vendor-reported, not independently replicated. But the point stands: a model that finishes a task in one session beats a cheaper model that needs three attempts, and no rate-card arithmetic captures that.

The point is narrower, and it has now been measured by someone other than the vendor. Artificial Analysis — which discloses that it supported Anthropic with pre-release evaluation of the model — puts Fable 5.1 at maximum effort at $3.76 per Intelligence Index task, 20% more than Fable 5, driven by roughly 1.7x the output tokens. The cache cut is doing real work inside that number: without it the same run would have cost about $5.16. But the behaviour change more than swallowed the discount, and the realised direction was a cost increase, not a 25-45% saving. Opus 5 scored 63 to Fable 5.1's 66 at maximum effort, for $2.34 per task.

One benchmark is not your invoice, and an intelligence index is output-heavy by construction — precisely the workload shape the crossover says should lose. That is the point rather than an objection to it. If you budgeted the 25-45% saving and then adopted the model's default behaviour, your realised saving will land under the estimate, because the estimate holds the token mix constant and Fable 5.1 emits a different one.

---

A Cache Miss Now Costs You Fifty Hits, Not Twelve

Cutting the read price without cutting the write price quadrupled the penalty for a cache miss. On Fable 5 a wasted 5-minute cache write cost 12.5 reads' worth of tokens ($12.50 against $1.00). On Fable 5.1 it costs 50 ($12.50 against $0.25). On the 1-hour cache the ratio went from 20 to 80. Opus 5 still sits at 12.5.

Your cache hit rate was always worth managing. It is now worth four times what it was, and the operational traps are unchanged and easy to trip:

• The 5-minute TTL runs from the start of the request, not the end of the response. Anthropic's caching documentation is explicit: if a response takes four minutes to stream, the follow-up must start within about one minute of it completing. A Fable 5.1 turn at default high effort with adaptive thinking on can burn most of that window before the user has read anything.
• Cache invalidation cascades in one direction: tools, then system, then messages. Editing a tool description invalidates everything. Rebuilding the system prompt between requests invalidates the messages below it.
• Changing top-level effort invalidates the message cache. The per-message effort beta is the exception — it changes level mid-conversation without restarting the cache, which is a genuine cost feature dressed as an ergonomics one. It is not a reason to pick Fable 5.1, though: Opus 5 supports it too.
• Caching is silent when it fails, and the floor is not the same on every model. A prompt under Fable 5.1's 512-token minimum is processed without caching and no error is returned. That minimum is 1,024 tokens on Sonnet 5 and 4,096 on Haiku 4.5 — eight times Fable 5.1's — so the mid-tier move recommended below can quietly stop caching a prefix that cached fine on the flagship, with nothing in the response to say so. If both cache_creation_input_tokens and cache_read_input_tokens come back zero, you paid full input rate.

That last one is the audit. Those two fields, plus input_tokens, are the whole diagnostic, and they are already in every response your application has ever received.

---

Three API Changes Will Break You Before the Bill Does

Fable 5.1 ships three breaking changes, and the one that matters most to a routing layer fails silently. Anthropic's migration guide lists all three.

Forced tool use returns a 400. tool_choice set to {"type": "any"} or {"type": "tool", "name": "..."} is rejected with tool_choice: type "tool" and "any" are not supported for this model. auto and none are unchanged. The documented replacement is auto plus strict tool use, or structured outputs. If you rely on forced tool calls for schema-valid JSON — most extraction pipelines do — that is a code change, not a model-ID swap, and it will also invalidate your message cache once when you make it.

Thinking blocks travel in one direction only. Fable 5.1 reads thinking blocks from earlier models; no earlier model reads Fable 5.1's. When a request carries a block the target model cannot read — a router or fallback switching models mid-conversation, exactly the architecture a multi-model cost strategy produces — the API drops the block before the model sees it. Anthropic is direct about the consequence: without the thinking-binding-controls-2026-08-01 beta header, the drop is silent. Your router keeps working, your bills keep arriving, and the reasoning continuity you were paying for quietly stops crossing model boundaries. Anyone who built a multi-model router around cache economics needs that header on before the first failover.

Editing conversation history invalidates thinking blocks. Modifying the system prompt, the tools array, or an earlier message before a Fable 5.1 thinking block produces a 400 reading The block is bound to a different conversation. The check is enforced for accounts created on or after 31 August 2026 — one day before launch. Older accounts get the mismatch recorded but not acted on unless they opt in, which means a company with a 2024 account and a company that signed up last week can run identical code and get different behaviour. Injecting a per-request reminder into an earlier turn and removing it next request is the specific pattern that breaks; turn-scoped system messages are the supported replacement.

One more constraint sits outside the API. Fable 5.1 and Mythos 5.1 carry 30-day data retention and are not available under zero data retention unless expressly authorised by Anthropic. They are Covered Models, as Fable 5 and Mythos 5 already were. If your ZDR commitment is what got the model past your security review, the cache-read discount does not change that answer.

---

What to Do About It

This Week:

1. Pull cache_read_input_tokens, cache_creation_input_tokens and input_tokens for the last 30 days from your gateway or Claude console, per workload, not in aggregate. Multiply each by its rate. If cache reads are under a third of the Fable 5.1 total, that workload is cheaper on Opus 5 today.
2. Grep your codebase for tool_choice with type any or tool. Every hit is a 400 the moment the model ID changes. Decide now between strict tool use and structured outputs.
3. If you route or fail over between Claude models, turn on the thinking-binding-controls-2026-08-01 beta header and log input_transformations. You need to know whether blocks are being dropped before you need to explain a quality regression.

This Month:

4. Re-derive your saving rather than adopting the 25-45% band. Compare your own August Fable 5 invoice against the same token mix at Fable 5.1 rates. If the delta is under 50%, Opus 5 is the cheaper model for that traffic and the routing policy should say so.
5. Instrument cache hit rate as a first-class metric with an alert, not a dashboard tile. At a 50:1 write-to-read ratio, a regression in prefix stability is now a budget event.
6. Run the fan-out workloads against Sonnet 5 and Haiku 4.5 before assuming Fable 5.1 is the cost play. Their cache reads are $0.20 and $0.10. Check the cacheable minimum when you move — 1,024 tokens on Sonnet 5, 4,096 on Haiku 4.5 — and assert both cache usage fields are non-zero in the A/B, or you will price a run that never cached. Any gateway that can enforce a spend cap can also A/B this.

Before Renewal:

7. If you hold a usage commitment, model it at both rates. A commitment sized on Fable 5 cache-read volume is now oversized in dollars while your token volume is unchanged, and that gap is negotiating leverage that expires when you sign.
8. Confirm whether your account is on the pre- or post-31-August-2026 preserved-thinking enforcement path. It determines whether a history-editing bug is a hard 400 in production or a latent defect.

---

The Bottom Line

A price cut that changes which model you should pick, while changing no headline number, is now a normal event rather than a surprising one. The rate card is no longer the interface to the cost. Multipliers, cache mechanics, tokenizer differences and behaviour defaults are, and they move independently of the two numbers a procurement spreadsheet has a column for. Anthropic's own guidance still says to start with Opus 5 for most workloads and reach for Fable 5.1 for demanding long-horizon work — which is a capability argument, correctly, not a price one.

The 75% cut is real and it is generous. It is also load-bearing on a variable nobody publishes and most teams cannot produce on demand: what fraction of your tokens are re-reads. Anthropic is better placed to see that fraction than you are, and a 25% and a 45% estimate quietly assume it. It called both figures estimates and showed no working.

Your router is making that call every second on numbers that are now wrong. Go find the fraction before your CFO does.

Continue Reading

• GPT-5.6 Sol Is $20 Until Nov 21. Budget Both Rates.
• Best LLM Gateways for Cost Control: Self-Host First
• Haiku Burned More Tokens Than Sonnet. Spec It in Code.
• Anthropic Bid $7B for Cheaper Inference. Don't Fix Your Rate.
• What RAG Actually Costs: $1,308 a Month at 10M Tokens/Day
• Google Bills Agent Memory Now. Revisions Count as Storage.