Claude Fable 5.1: the cache read price cut that changes routing arithmetic

Claude Fable 5.1 is out, and its cache read price cut is the part that reshapes routing

On 1 September 2026 Anthropic launched Claude Fable 5.1, the new flagship for coding and knowledge work, with cache read pricing 75 percent cheaper, at $0.25 per million tokens. The effect goes beyond a stronger model: the cost shape changed, and anyone routing long, agentic workloads needs to redo the math. The primary source is the Anthropic announcement.

This is not a cut across the whole price sheet. Fable 5.1 still costs $10 per million input tokens and $50 per million output, the same as Fable 5. What fell was the price of reusing already-processed context, the cache read token. Because agentic loads re-read the same context dozens of times, that is precisely the line that decides the final bill.

What happened

Fable 5.1 is the same model as Claude Mythos 5.1, with different safeguards: Mythos ships only through trusted access programs, for cybersecurity and life sciences work. Fable 5.1 is generally available today on AWS, Google Cloud, Microsoft Azure, and the Claude API under the identifier claude-fable-5-1.

The numbers Anthropic published on 1 September 2026:

• Cache read at $0.25 per million tokens, 75 percent below the prior price.
• Total cost about 25 percent lower than Fable 5 on typical workloads, measured at default effort over four weeks of real August 2026 usage.
• Up to about 45 percent lower on highly agentic workloads, where cache reads make up most of the cost.
• Indexed cost of Fable 5 = 100; on Fable 5.1, 75 typical and 55 agentic.
• Input and output pricing unchanged: $10 and $50 per million tokens.

On benchmarks the capability gain is larger than the price cut. On Terminal-Bench-Science 0.1 Fable 5.1 scores 52.6 percent, against 24.7 percent for Fable 5. On Terminal-Bench 4.0, 55.8 percent against 42.0 percent. On AutomationBench, 31.4 percent against 17.1 percent. On OSWorld 2.0 strict, 41.7 percent against 36.1 percent. On GDPval-AA v2, 1853 against 1723. On CursorBench 3.2.0, 73.4 percent against 70.5 percent. Fable 5.1 defaults to high effort in Claude Code and medium in Claude Cowork and on Claude.ai.

Why it matters

The cache read cut is the part of the announcement that runs through the bill of anyone routing LLMs. A frontier model carries two very different cost lines: the new input token, billed at $10 per million, and the cache read token, now billed at $0.25 per million. An agentic load reads the same context block hundreds of times across a dozen calls; the real cost of that load is dominated by the second line, not the first.

The position here is direct: the price cut does not make Fable 5.1 just another cheap model, it reshapes the shape of the bill. Routers that price every input token as a single line will charge agentic loads wrong, too high or too low, because they do not see the cache read line that now dominates. Anyone steering traffic by token volume, instead of by cost per completed task, misses most of the savings from this release.

For the CFO the bill shows up again at month end. For the CTO it shows up in the routing policy: the task that could not justify a long-lived agent now fits, because the cost of re-reading context fell to a quarter. For the CEO the signal is about pricing: frontier intelligence is no longer the ceiling it was, and choosing which model serves each task depends more on gateway design than on the price list.

What the shift in cost composition means in practice

What changes is where the bill concentrates. Before, on Fable 5, the cost of an agentic load spread between new input and context reuse. With Fable 5.1, cache reads now account for the largest share, but that share costs a quarter of the old price, and the result is a smaller total bill. It is a shift in the shape, not in the list.

The table below compares the prior regime with Fable 5.1 for anyone routing LLMs. The values come from the Anthropic primary source, published on 1 September 2026.

The most concrete reading is the viability window for agents. An agent that makes twenty calls to the model over the same repository pays twenty times the price of reused context. At $0.25 per million cache read tokens, that repetition is no longer the bottleneck it was. The same load that cost itself out of use now fits a long, autonomous analysis.

What to do now

Reviewing the routing policy is the answer, and it fits this week, while the new price is still fresh and the market has not yet repriced. The decisions below are ordered by implementation cost.

1. Classify the load by cache reuse. Before routing to Fable 5.1, split traffic into two buckets: single-read tasks and tasks that re-read context. The second bucket is where this release helps most.
2. Measure cost per completed task, with and without cache. The metric that answers is the total cost of a finished task, not the price of a token. Recalculate the break-even of each task type with cache reads at $0.25 per million.
3. Review the model-selection rule in the gateway. If the policy still picks a model by price per token, it misprices agentic loads. Switch to picking by estimated reuse.
4. Test Fable 5.1 side by side on a real segment. Track return per task and total cost before migrating the whole traffic share, because the capability gain does not come without validating your specific load.
5. Reopen tasks archived on cost. Long-duration workloads that left the budget at the old price deserve a fresh look, provided repeated context reading makes up most of the cost.

Frequently asked questions

Three questions close the frame. The values use the Anthropic primary source, read on 1 September 2026.

What changed in Claude Fable 5.1 pricing? Input and output pricing did not: $10 and $50 per million tokens. What fell 75 percent was the cache read price, now $0.25 per million tokens. On typical workloads total cost drops about 25 percent; on agentic loads, up to about 45 percent, per Anthropic.

Why does cache read matter for cost? Agentic loads reuse the same context dozens of times across several calls. The cache read token prices that reuse, and its bill tends to dominate the final cost of an agent. A price cut on that line is what moves the total cost of long autonomous work.

Does this change which model serves each task? It does, because comparing models on cost now depends on the context-reuse rate of the load. For anyone routing LLMs, choosing whether Fable 5.1 serves a task over a comparable should not lean only on the token price, and instead on cost per completed task, measured with and without cache.

Is Mythos 5.1 different from Fable 5.1? No. According to Anthropic they are the same model with different safeguards. Mythos 5.1 ships only through trusted access programs for cybersecurity and life sciences, while Fable 5.1 is generally available.

Is this a routine release? No. It is the first time a cache read price cut on a frontier model reshapes the form of the cost and, with it, moves the cost of an agentic load onto a far cheaper line. For anyone routing, it is a price-sheet change with policy effect.

References and further reading

• Anthropic announcement: Claude Fable 5.1 and Claude Mythos 5.1, 1 September 2026, primary source for capabilities, pricing and the cache read cut.
• Claude Fable 5.1 / Mythos 5.1 System Card, details on safety, alignment and evaluations.
• AlphaSignal (news@alphasignal.ai): "Anthropic Claude 5.1 25% cheaper", email of 1 September 2026, discovery channel for the event.
• Every / Vibe Check (hello@every.to): "Fable 5.1, Anthropic Is So Back (Again)", qualitative observations reported as per Every, not as Anthropic-confirmed attributes.

What to expect in the coming weeks

The cache read price is not a promotion with an expiry date; it is the new sheet. The test of merit will come from real workloads measured by cost per task, not from benchmark noise. If the 75 percent cache read cut holds across other model families, the routing decision gains a new dimension, because the line that separates one model from another is no longer the token price, and instead the gateway's ability to tell context reuse from fresh reads. The sheet changed the rule. For Nexforce the reading is direct: when a frontier model drops in price so unevenly between cost lines, the layer that sees that asymmetry and turns the sheet into a per-task cost decision becomes the control point. Keeping a layer that decides, per task, which model serves each load by measuring the real cost, rather than the token price, is exactly what the Nexforce Router does: one API, many models, and route selection based on the full cost of the task, instead of a single pricing line.