Claude Fable 5.1 Pricing: What Anthropic's Frontier Model Actually Costs

What Is Claude Fable 5.1?

Claude Fable 5.1 is Anthropic's current frontier-tier model, released in September 2026 alongside a companion version called Claude Mythos 5.1. Both run on a 1 million token context window with a maximum output of 128,000 tokens, and both cost exactly the same to use: $10 per million input tokens, $50 per million output.

Here's the part that trips people up: Fable 5.1 and Mythos 5.1 aren't two different models with two different capability levels, the way Opus and Sonnet are. They're the same underlying model. The difference is safety configuration and who's allowed to use it. Fable 5.1 ships with standard safeguards around cybersecurity, biology, and chemistry topics. Mythos 5.1 has those safeguards relaxed, and access is restricted to vetted organizations through two specific programs: the Cyber Verification Program for defensive security work, and the Life Sciences Verification Program for legitimate biology and chemistry research. Both are part of Anthropic's Project Glasswing, currently limited to a set of US organizations while Anthropic works with the US government to expand it.

That access model isn't theoretical. Mozilla has credited an earlier version of Mythos with helping patch 271 Firefox vulnerabilities, the kind of large-scale defensive security sweep that's exactly what the Cyber Verification Program exists for.

Claude Fable 5.1 Pricing at a Glance

Fable 5.1 and Mythos 5.1 share one rate card. The headline input and output rates didn't move from the previous generation, the real change is in caching. 

One more rate to know: running Fable 5.1 with US-only inference (data that never leaves US infrastructure) costs 1.1x the standard rate, across input, output, cache writes, and cache reads alike. That applies whether you're calling Anthropic directly, through Claude on AWS, or through Claude in Microsoft Foundry.

The Cache Cut Is the Real Price Change

Cache reads dropped from $1 to $0.25 per million tokens, a 75% cut. That's a bigger deal than it sounds, because cache reads on Fable 5.1 now cost 2.5% of the base input rate, well below the 10% cache-read discount most other Claude models and GPT Astra offer.

Run the math on a single write-then-reuse cycle at 1 million tokens: a cache write costs $12.50, and reading it back costs $0.25, for $12.75 total. Sending that same million tokens fresh twice, with no caching at all, costs $20. That gap is the entire pitch for restructuring prompts around cache-friendly architecture, front-loading static content like system instructions and reference material, and putting the variable, per-request content at the end.

Same Sticker Price, Doesn't Mean the Same Bill

Here's the catch: Fable 5.1 doesn't actually cost less to run than Fable 5, even with the cache cut. Artificial Analysis, which ran pre-release evaluation on the model, found that Fable 5.1 at maximum reasoning effort costs $3.76 per task on their Intelligence Index benchmark, 20% more than Fable 5's $3.14, despite the 75% cheaper cache reads. The model uses more output tokens to reach a higher score (it topped the index at 66, the highest score Artificial Analysis has measured), and that extra output spend outweighs what caching saves.

Anthropic gives you a lever here that Astra doesn't have in the same form: five effort levels, from low to max, with a roughly 11x range in output token volume between them. At "xhigh" effort, Fable 5.1 scores 65 on the Intelligence Index at $2.72 per task, a dollar cheaper than max effort, while still beating Claude Opus 5's max-effort score of 63 at $2.34. If you're not chasing the absolute top score, that's real money on the table.

Fable 5.1 vs Opus 5, Sonnet 5, and GPT Astra

That Fable-to-Astra match isn't a coincidence worth glossing over. It means the two labs' flagship reasoning models now cost exactly the same per token, for the first time. If you've been running an Astra-vs-Claude cost comparison to pick a frontier vendor, the sticker price no longer settles it, you're back to workload fit, reasoning behavior, and benchmark performance for your specific task. We covered the Astra side of this in Finout's GPT Astra pricing breakdown.

On a pure cost-per-task basis, Fable 5.1 is the expensive end of Anthropic's own lineup: 60% more per task than Opus 5 at max effort ($3.76 vs $2.34), and roughly 8x Sonnet 5's per-token rate. That's not a knock on Fable 5.1, it's a reason to treat model selection as a routing decision per task rather than a single account-wide default.

Where You Can Run Fable 5.1

Like most frontier models, Fable 5.1 isn't a single-channel purchase.

Anthropic direct API

Buying directly from Anthropic gets you the standard rate card above, with global routing by default and the option to pin to US-only inference at the 1.1x premium.

Claude on AWS Bedrock

Fable 5.1 is available through AWS Bedrock, billed through your AWS account. The US-only inference multiplier carries over here too.

Claude in Microsoft Foundry

Azure customers can run Fable 5.1 through Microsoft Foundry, including Azure's US Data Zone Standard deployments, which also carry the 1.1x US-residency rate.

Where the Economics Work For You, and Where They Don't

The cache-read cut is the cheapest rate on any current Claude or GPT Astra flagship, and it costs nothing extra to benefit from if your workload already reuses long, stable prompt prefixes: system instructions, tool definitions, retrieved documents. The five effort levels also mean you're not locked into max-effort pricing for every call, tuning effort down is a legitimate lever most teams under-use.

Where it works against you: output tokens still carry no cache relief at $50 per million, so a workload that's genuinely output-heavy sees none of the cache savings. Fresh, uncached input is double Opus 5's rate and five times Sonnet 5's. And the savings from caching only materialize if your prompts are actually structured to hit the cache, most teams haven't built that discipline yet, so the advertised numbers and the numbers on your invoice can diverge substantially.

How to Control Fable 5.1 Spend

1. Tune effort level to the task

Don't default every call to max effort. The xhigh-vs-max comparison above shows a real quality-for-cost tradeoff sitting right in the API, not something you have to build yourself.

2. Build for cache hits deliberately

Structure prompts with static content first (system instructions, reference documents, tool schemas) and variable, per-request content last. That's what turns the $0.25 cache-read rate from a footnote into your biggest lever.

3. Route by task, not by default model

Reserve Fable 5.1 and Mythos 5.1 for work that actually needs frontier-level reasoning. Opus 5 and Sonnet 5 sit meaningfully lower on price for tasks that don't need the top score.

4. Know what US-only residency costs you

If your compliance requirements don't specifically require US-only inference, the 1.1x premium applies to every token category, input, output, and cache alike, for no performance benefit. Confirm it's actually a requirement before defaulting to it.

5. Tag and allocate at the call level

Finout's Virtual Tags map Claude spend to the team, feature, or customer driving it after the fact, so effort-level and caching decisions can be traced back to who's making them, not just what the aggregate bill looks like.

Turning Fable 5.1 Spend Into a Business Decision

The frontier tier just got more complicated to reason about, not less. Two vendors at identical per-token pricing, a model that costs more per task despite a real price cut, and five effort levels that all show up on the same line item. Squinting at a single sticker price won't tell you whether Fable 5.1 is the right call for a given workload.

Finout ingests Anthropic spend alongside OpenAI, AWS, Azure, GCP, and the rest of your cloud and AI bill, so you can see cost per task, per effort level, and per team without reconstructing it from raw invoices. See Finout's guide to AI cost visibility for the fuller picture, and Finout's Detection Agent for catching it automatically when an effort-level change or a cache-miss streak quietly reprices a workflow.