September 2, 2026 ·

anthropicclaudepricingagentscodingstrategyeu-ai-act

Anthropic did not cut the price of Claude Fable 5.1 — it cut the price of one line item, and only your workload decides whether that helps

TL;DR: On 1 September 2026 Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 — the same underlying model behind two different safeguard layers. Coverage led with a 75% price cut. The pricing table is narrower than that: base input is still $10/MTok, base output is still $50/MTok, and both cache write tiers are unchanged. Exactly one number moved — the cache read, from $1.00 to $0.25 — and Anthropic broke its own uniform pricing rule to move it, applying a 0.025x multiplier on these two models where every other Claude uses 0.1x. Why it matters: that is not a discount on the model, it is a discount on re-reading context you already paid to send. The advertised “~25% typical, up to ~45% agentic” saving is a claim about the shape of your traffic. The arithmetic: a cache-dominated agent turn gets close to 50% off versus Fable 5; a chat-shaped one gets under 5%. And Fable 5.1 only undercuts Opus 5 once cached tokens exceed 20× fresh input + 100× output — roughly 220,000 cache hits on a normal turn. For you: read cache_read_input_tokens in the responses you are already getting, then decide. And budget migration time — forced tool use now returns a 400.

What shipped

Anthropic released two models on 1 September. Claude Fable 5.1 is generally available as claude-fable-5-1 on the Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry. Claude Mythos 5.1 is the same model with a lighter safeguard layer, restricted to vetted US organisations through Project Glasswing. Both carry a 1M-token context window, 128k maximum output tokens, and always-on adaptive thinking.

The capability jump is real and concentrated in agentic and scientific work. Anthropic’s numbers put Fable 5.1 at 52.6% on Terminal-Bench-Science 0.1 against 24.7% for Fable 5 and 29.0% for Opus 5 — better than double its predecessor. Terminal-Bench 4.0 goes from 42.0% to 55.8%, with Mythos 5.1 at 60.9%. AutomationBench roughly doubles, 17.1% to 31.4%. Humanity’s Last Exam lands at 60.9% without tools and 65.0% with them; CursorBench 3.2.0 at 73.4%; OSWorld 2.0 at 41.7% under strict scoring.

Those are vendor-run benchmarks with a stated standard error of 3.5 to 4.5 points, which is wide enough to swallow most of the smaller gaps. Treat the Terminal-Bench-Science doubling as a signal and the two-point differences as noise. Independent replication is what settles these, and it has not arrived yet.

But the benchmarks are not the interesting part of this launch. The pricing table is.

The price that moved, and the one that did not

Here is the comparison that most coverage compressed into a single percentage:

Line item	Fable 5	Fable 5.1	Opus 5
Base input	$10 / MTok	$10 / MTok	$5 / MTok
Output	$50 / MTok	$50 / MTok	$25 / MTok
Cache write (5 min)	$12.50 / MTok	$12.50 / MTok	$6.25 / MTok
Cache write (1 hour)	$20 / MTok	$20 / MTok	$10 / MTok
Cache read	$1.00 / MTok	$0.25 / MTok	$0.50 / MTok

Four of the five rows are identical. The model did not get cheaper. Reading from cache got cheaper, by 75%.

What makes this more than a routine discount is that Anthropic broke a rule to do it. Across the entire Claude range, a cache hit has always cost 0.1x the base input price — a clean multiplier that holds for Haiku, Sonnet, Opus and Fable alike. Anthropic’s own documentation now carries a footnote carving out an exception: cache hits on Fable 5.1 and Mythos 5.1 are priced at 0.025x base input. Two models in the entire lineup are off the ladder.

A vendor does not break its own pricing schema casually. This is a targeted subsidy for one workload shape — the long-running agent that keeps a large context resident and re-reads it across many turns — and it tells you where Anthropic thinks the revenue is going.

What the saving is actually worth to you

Anthropic’s stated figures are roughly 25% on typical workloads and up to about 45% on highly agentic ones. Both are true and neither is about your bill. Here are two turns, priced out.

A cache-dominated agentic turn — 220,000 cached input tokens, 1,000 fresh input tokens, 2,000 output tokens:

Fable 5: $0.220 + $0.010 + $0.100 = $0.330
Fable 5.1: $0.055 + $0.010 + $0.100 = $0.165 — a 50% cut

A chat-shaped turn — 10,000 cached, 5,000 fresh, 2,000 output:

Fable 5: $0.010 + $0.050 + $0.100 = $0.160
Fable 5.1: $0.0025 + $0.050 + $0.100 = $0.1525 — a 4.7% cut

Same models, same price list, and a saving that varies by an order of magnitude depending entirely on what you send. The number you are looking for is already in your logs: every API response returns cache_read_input_tokens, and the fraction of your total input that field represents is your discount, scaled.

This is the durable point, and it generalises past this launch. Frontier pricing has stopped being a single number you compare across a row. It is now a structure — separate rates for fresh input, cached input, cache writes at two durations, output, batch, US-only inference at a 1.1x multiplier — and vendors increasingly compete by discounting the line item that suits their infrastructure rather than the one that dominates your bill. We saw the same pattern when the price war flipped in August and again in the open-weight price floor moves. A headline percentage is now a workload assumption wearing a number.

The comparison nobody ran: Fable 5.1 against Opus 5

Because the cache cut is steep enough to cross a tier boundary, it is worth asking directly whether the expensive model is now the cheap one.

Fable 5.1 charges $0.25 per million cached tokens. Opus 5 charges $0.50 — twice as much — while being half the price on every other line. So for a sufficiently cache-heavy workload, the flagship undercuts the mid-tier. Setting the two totals equal gives the crossover:

Fable 5.1 costs less than Opus 5 only when cached tokens > (20 × fresh input tokens) + (100 × output tokens).

Check it against the cache-dominated turn above: 220,000 cached, 1,000 fresh, 2,000 output. Opus 5 comes to $0.110 + $0.005 + $0.050 = $0.165, precisely level with Fable 5.1. Below that cache volume Opus 5 wins; above it Fable 5.1 does.

That threshold is high but not theoretical — it describes a real agent working a large repository inside the 1M-token window. What it rules out is the casual assumption that the cache cut has made Fable 5.1 a sensible default. For anything conversational, Opus 5 is still roughly half the price, still the model Anthropic reports the lowest misalignment scores for, and still outside the covered-model retention regime that constrains the Fable and Mythos class. If you are picking a model for Claude Code or a coding agent, our AI coding tools shortlist and the DeepSeek V4 Pro versus Claude Opus 4.8 comparison are both better starting points than a headline percentage.

Three things that break

The upgrade is not a drop-in swap, and the breaking changes sit in orchestration code rather than prompts.

Forced tool use is gone. Requests setting tool_choice to "any" or to a named "tool" now return a 400 error. Any agent that guaranteed a structured call by constraining the model must now handle a model that may reply in prose. This is the change most likely to break something in production.

Thinking blocks are model-bound. Earlier Claude models can read Fable 5.1’s reasoning; Fable 5.1 will not accept thinking blocks produced by a different model. If you route one conversation across a model ladder to control costs, that pattern needs revisiting.

Editing earlier turns invalidates their thinking blocks, enforced for organisations created after 31 August 2026.

Alongside these, the safeguard layer was retuned rather than loosened wholesale. Cyber safeguards now permit vulnerability discovery while continuing to block exploit development — Anthropic reports roughly 60% fewer interventions in Claude Code — and biology safeguards fire about 85% less often on benign requests. That is a direct answer to the false-positive complaints that followed the reward-hacking disclosure, and it lands in the same week, which is unlikely to be coincidence.

The compliance small print

Two items matter disproportionately for European and regulated buyers.

Both models embed the statistical text watermark Anthropic began applying in August, with C2PA credentials on generated files, satisfying the Article 50 transparency obligations that went live on 2 August. The detection API, however, is in private preview — limited to regulators, law enforcement, media, fact-checkers, independent researchers, educational bodies and EU civil society groups, with no announced timeline for wider access. The mark exists; you cannot check it. Your own disclosure obligation is unaffected by that asymmetry.

On retention, the default remains 30 days, with zero data retention requiring express authorisation. Fable-class models have carried tighter terms than the Opus line all year. If you operate under ZDR or a BAA, confirm in writing which models your teams actually call rather than assuming the 5.1 release reset anything — and note that Anthropic’s willingness to change commercial terms mid-flight is established, having withdrawn its Priority Tier, the only tier with a published uptime target, from sale in August.

What to do

Measure before you migrate. Pull cache_read_input_tokens from a representative week of responses and compute cache hits as a share of total input. Under ~50%, the cut is close to rounding error and this launch is a capability decision, not a cost one.

Price the crossover, not the headline. If you are on Opus 5, apply the rule above before assuming Fable 5.1 is now competitive. For most traffic it is not.

Budget for the 400. Audit for tool_choice: "any" before you switch a production agent.

Do not build a pipeline that only pays on one vendor’s exception. The 0.025x multiplier exists on two models from one provider. Designing around it is a single-vendor bet, and this year has already shown what those cost when OpenAI cut Cursor’s model access. A neutral gateway keeps the option open.

The bottom line

Fable 5.1 is a genuine capability release — the Terminal-Bench-Science jump is the most substantial single-benchmark move Anthropic has posted this year, and Claude remains the quality default for work that ships. But the pricing story is narrower than the coverage, and narrower in a way that matters: Anthropic did not make its best model cheaper. It made one billing line cheaper, on two models, by breaking its own multiplier, in a direction that rewards exactly one architecture.

If your workload has that shape, this is one of the largest effective discounts of the year. If it does not, you are looking at the same $10 and $50 you were paying yesterday, plus a migration. The difference between those two outcomes is not in the announcement. It is in a field your API is already returning.

Update, 3 September 2026: the access half of this launch turned out to matter more than the pricing half. Anthropic’s announcement states that Fable 5.1 and Mythos 5.1 “are the same model, but with different levels of safeguards” — Mythos is not a separate model but a permission set, reachable only through the Cyber Verification Program and the new Life Sciences Verification Program, currently limited to US organisations while Anthropic coordinates with the US government on wider access. Within 48 hours OpenAI confirmed Astra at the Critical cyber threshold behind Daybreak Blue, and Google shipped Gemini 3.8 Flash Cyber behind Fairwind. Read alongside the retuned cyber and biology safeguards described above, the pattern is that the safeguard configuration — not the weights — has become the product being tiered. Three labs, three vetting regimes, one week.

Update, 4 September 2026 — a competitor matched the sticker price exactly. OpenAI released GPT-6 Astra on 3 September at $10 per million input tokens and $50 output — the identical headline rate Fable 5.1 charges. The two price sheets diverge underneath it. Anthropic’s discount lives in the cache-read line, which is why this article argues the saving depends on your workload shape; OpenAI’s Astra keeps cached input at $1 per million but adds a hard threshold at 272,000 input tokens, above which the entire request bills at $20/$75. If you are comparing the two flagships on headline price you will find no difference at all, which makes the cache and context-length columns the only ones that actually decide the bill. The Astra price sheet and the arithmetic behind its cost-per-task claim.

Frequently asked questions

Is Claude Fable 5.1 actually cheaper than Claude Fable 5?

Never more expensive, and the size of the saving is entirely a function of how much of your input is served from cache. Base input ($10 per million tokens), base output ($50), and both cache write tiers ($12.50 for the five-minute cache, $20 for the one-hour cache) are identical between the two models. The single changed number is the cache read, which drops from $1.00 to $0.25 per million tokens. So if none of your traffic hits the cache, Fable 5.1 costs you exactly what Fable 5 cost. If your traffic is dominated by cache hits — the normal profile for a long-running agent that keeps a large codebase or document set resident in context across many turns — you approach a 50% reduction on the input side of the bill. Anthropic's own framing of roughly 25% on typical workloads and up to about 45% on highly agentic ones is consistent with that arithmetic. It is a statement about a distribution of customer traffic, not a guarantee about yours. The way to know your number is to read the cache_read_input_tokens field your existing API responses already return and work out what fraction of your input it represents.

When does Fable 5.1 become cheaper than Opus 5?

Later than most people expect, because the cache cut works against a model that is twice the price on every other line. Opus 5 is $5 input, $25 output and $0.50 cache read; Fable 5.1 is $10, $50 and $0.25. Setting the two totals equal and solving gives a clean rule: Fable 5.1 costs less than Opus 5 only when your cached input tokens exceed twenty times your fresh input tokens plus one hundred times your output tokens. Put concretely, a turn with 1,000 tokens of new input and 2,000 tokens of output needs roughly 220,000 tokens of cache hits before the two models cost the same, and more than that before Fable 5.1 wins. That threshold is reachable — it describes a genuine agentic loop over a large repository within the 1M-token context window — but it is not where most production traffic sits. For anything chat-shaped, Opus 5 remains substantially cheaper, and it is worth remembering that Opus 5 also posts the lower misalignment scores and is not subject to the covered-model retention rules that apply to the Fable and Mythos class.

What breaks if I upgrade from Fable 5 to Fable 5.1?

Three things, and the first is the one that will bite an existing agent. Forced tool use is gone: requests sending tool_choice set to 'any' or to a specific 'tool' now return a 400 error rather than complying, so any code that guaranteed a tool call by constraining the model needs reworking to handle a model that may answer in text instead. Second, thinking blocks are now bound to the model that produced them — an earlier Claude can read Fable 5.1's reasoning, but Fable 5.1 will not accept thinking blocks from a different model, which matters if you route a single conversation across a model ladder. Third, editing earlier turns mid-conversation invalidates the thinking blocks attached to them, and this is enforced for organisations created after 31 August 2026. None of these is exotic, but all three sit in exactly the kind of orchestration code that agent frameworks generate rather than developers write, so budget review time rather than assuming a drop-in swap.

Can I get Claude Mythos 5.1 in Europe?

No. Mythos 5.1 is the same underlying model as Fable 5.1 with a lighter safeguard layer, and access runs through Project Glasswing, a restricted programme for vetted cybersecurity and life-sciences organisations developed with the US government. European organisations cannot currently enrol, notwithstanding that Claude has been generally available in the EU for two years. Anthropic has said it is working to include international partners but has published no timeline. For EU buyers the practical position is that the generally available Fable 5.1 is the ceiling of what you can purchase, and any vendor pitching Mythos-class capability to a European entity is either describing a downstream product — the Claude Security vulnerability scanning route, for instance, which delivers findings rather than model access — or describing something they cannot currently sell you.

Both models carry an EU AI Act watermark. Does that let me detect Claude-generated text?

Not yet, and the gap between the two halves of that sentence is the part worth holding onto. Fable 5.1 and Mythos 5.1 embed the same statistical, machine-readable watermark Anthropic began applying in August, with C2PA credentials attached to generated files, which is how Anthropic meets the Article 50 transparency obligation. The detector, however, launched into private preview restricted to eligible organisations — regulators, law enforcement, media, fact-checkers, independent researchers, educational bodies and EU civil society groups. Anthropic says access will widen but has not said when. So the marking obligation is being met by the provider while the verification capability stays gated, and an ordinary buyer, employer or publisher cannot check a document today. That does not reduce your own disclosure duty by a word: if you deploy AI-generated content, the obligation to disclose is yours regardless of whether anyone can independently verify the output.

Does the cheaper cache change how I should build agents?

It changes the economics enough to be worth a design pass, but in a specific direction: it rewards keeping context resident rather than retrieving it repeatedly. At $0.25 per million tokens, re-reading a cached 200,000-token context costs five cents, which is genuinely close to free relative to the output tokens in the same turn. That makes the brute-force approach — load the whole repository or document set into the 1M-token window and keep it warm — cost-competitive with a carefully engineered retrieval layer that exists mainly to keep input bills down. Before you rebuild anything, check two things. Cache writes are unchanged and still cost more than base input ($12.50 for five minutes, $20 for an hour), so a workload that keeps rewriting the cache rather than reading it will not benefit and may cost more. And the discount exists on precisely two models from one vendor, so designing a pipeline around it is a bet on Anthropic specifically, which is the sort of single-vendor dependency that has already caught developers out elsewhere this year.

Sources

Related tool reviews

Questions or corrections? Email Pick Right. Want the full list? See all news.