Anthropic shipped Claude Fable 5.1 and Claude Mythos 5.1 on 1 September. The announcement led with benchmarks: 52.6% on Terminal-Bench-Science 0.1, more than double Fable 5, and 55.8% on Terminal-Bench 4.0 against 42.0%.
The line that matters more to anyone running agents was further down. Input and output rates did not change. Cache reads dropped 75%, to $0.25 per million tokens. Anthropic put the practical effect at roughly 25% cheaper for typical workloads and up to 45% for highly agentic ones.
One commenter caught the asymmetry immediately, predicting the discourse cycle before it happened:
Days 1 to 3: "THIS IS AMAZING, it one-shot my entire refactor." Days 4 to 7: "THIS IS THE WORST MODEL EVER SHIPPED, Anthropic has ruined it." Meanwhile the line nobody screenshotted: cache reads rates dropped 75%, to $0.25 per million tokens. Up to 45% off a heavily agentic workload. That's the part that changes what you build.
Then somebody measured it
The next day a developer published a side by side through an agent loop on both models, same three prompts. The result deserves more attention than the launch post: they found Fable 5 at $7.65 and Fable 5.1 at $7.08, a 7.5% saving. Not 45%.
Their explanation is the whole lesson, and they were honest about it:
An agent loop keeps replaying a large mostly stable context on every step, so cache reads are what actually dominate the bill on any long run. New input and output are small compared to how many times the model rereads the growing context. Anthropic claims up to 45% cheaper on highly agentic workloads. We didn't hit that, because two of our three tasks converged in a few turns and cache didn't grow. The long one ate almost all of the delta. For chat-shaped usage the cut is pennies. For long autonomous loops it's [substantial].
Nobody is wrong here. Anthropic said up to 45% on highly agentic workloads. The tester ran a mixed set where two tasks finished quickly. The discount applied exactly as documented to the portion of spend it covers, which in their sample was small.
This is the most common way teams mislead themselves about model pricing: reading a headline elasticity and applying it to a workload with a different shape.
The number that predicts your saving
There is one ratio that tells you where in the 7% to 45% band you land:
cache read tokens / (fresh input + output tokens)
High ratio means a long horizon agent replaying a growing context on every step, and the cut lands close to the top of the range. Low ratio means short chat shaped turns, and it lands near the bottom. The model is identical in both cases. Your architecture decides which price you pay.
Two consequences follow, and the second is the interesting one.
First, per token rate cards are now actively misleading for agents. Fable 5.1 has the same input and output price as Fable 5 and a materially different effective price. Any comparison table that lists only input and output, including plenty of ours historically, cannot express that difference. We made the model selection version of this argument in 36x the price for 22% more quality: the headline price is the least informative part of the bill.
Second, cache pricing has become a competitive axis. Alibaba lists Qwen3.8-Flash cache hits at $0.016 per million against $0.15 fresh input. Anthropic has now moved its own cache read rate as the headline improvement of a flagship release. Vendors have worked out that agentic workloads are cache dominated and are competing there instead of on the sticker price. Expect the next several releases to do the same.
Where the saving quietly disappears
Two threads from the same 72 hours complicate the story, and both are worth knowing before you promise your finance team a discount.
A Max plan user asked whether anyone else was burning through usage faster with Fable 5.1 on the same workflow. A cheaper cache read does not help if a model is more verbose, thinks longer or takes more turns. Effort level and output length sit outside the discount entirely, and Anthropic's own note that 5.1 can match Fable 5 at lower effort settings is a hint about where the real saving lives: you have to turn the effort down to collect it.
And a cache discount is worth nothing if your cache is not being hit. Your hit rate depends on a byte stable prompt prefix, on not switching provider mid conversation, and on not flooding context with volatile tool output. A workload with a 20% hit rate collects 20% of the advertised benefit at best.
What to do this week
Measure your cache read share before you model any saving. Pull the ratio above from usage analytics and cost tracking. If cache reads are 15% of your token spend, a 75% cut on them is an 11% saving and no more.
Confirm caching is engaged. Automatic caching handles the common path, and manual caching plus Messages API caching give you explicit breakpoints where a provider supports them.
Re run your effort levels. The claim that 5.1 matches its predecessor at lower effort is testable in an afternoon and is probably worth more than the cache change. Use reasoning controls and compare cost per completed task, not cost per token.
Compare on blended cost, not on rate cards. When a vendor moves one component of the price, the only comparison that survives is total spend per unit of work on your own traffic. Keep fallback policies and managed policies in front of the choice so switching is a config change once the numbers come in.
Watch for the launch week regression. New flagship, new incident surface. Anthropic logged multiple status incidents across the same three days, including elevated errors on Fable 5.1 itself, so pair any migration with alerts and performance monitoring.
The takeaway
The most valuable artifact of this launch was not the benchmark table. It was a developer spending an afternoon running the same three prompts through both models and publishing a number that undercut the marketing claim by 6x, with a clear explanation of why both figures are honest.
Do that on your own traffic. The band between 7% and 45% is decided entirely by how your agents are built, and nobody outside your organisation can tell you where you sit.
Compare live prices and cache rates across providers in the models catalog and the cheapest rankings, or see how our pricing compares. For routing layers specifically we keep honest pages on OpenRouter, LiteLLM and Portkey.
Start routing on Requesty and find out which price you are paying.
Frequently asked questions
- What changed in Fable 5.1 pricing?
- Input and output rates are unchanged. Cache reads dropped 75%, to $0.25 per million tokens. Anthropic states this reduces practical cost by around 25% for typical workloads and up to 45% for highly agentic ones.
- Why did an independent test only measure 7.5% cheaper?
- Because the discount only applies to cache reads, and cache reads only dominate when a long context is replayed many times. In the published side by side, two of three tasks converged in a few turns so their cache never grew, and the single long task accounted for almost all of the saving.
- How do I know whether my workload gets 7% or 45%?
- Look at your ratio of cache read tokens to fresh input and output tokens. Long autonomous loops that reread a growing context on every step sit at the top of that range. Short chat shaped requests sit at the bottom, where the cut is worth pennies.
- Does a cache read discount change which model I should use?
- It can invert a comparison. Two models with identical headline input and output prices can differ severalfold in effective cost once cache read rates and your own hit rate are applied, so compare on blended cost per million tokens for your traffic rather than on the rate card.
- AUG '26
36x the price for 22% more quality: the Pareto data that should decide your model mix
Glean benchmarked 37 models and reasoning configurations across 1,000 enterprise tasks. The gap between the cheapest and the priciest frontier model was 36x on price and 22% on quality. Add an 11x provider spread and a 2x reasoning effort penalty on the same model, and single model deployments start to look like the most expensive decision in the stack.
- JUL '25
Cross-Provider Caching Deep Dive: Maximize Performance Across Your Stack
- JUL '26
$1.8k before anyone noticed: how to cap runaway agent spend
Retry storms, agent loops, and stolen keys are the three ways teams lose four figures of LLM budget in a day. A practical setup for hard caps, per key limits, alerts, and loop detection.