Fable 5.1 Cache Cut: Re-run Agent Cost Model

The cache cut is a decision trigger

Your agent bill changes only when the same context is reused. Anthropic cut cached input for Fable 5.1 to $0.25 per million tokens, a 75% reduction from Fable 5's $1.00 rate. Anthropic released Fable 5.1 and Mythos 5.1 as the same model with different safeguards; Fable 5.1 is generally available, Mythos 5.1 is limited to vetted cybersecurity and life-sciences users. Fable 5.1's standard input and output rates stay at $10 and $50 per million tokens, matching Fable 5.

Anthropic forecasts a roughly 25% reduction in effective Fable 5.1 spending for ordinary workloads and a maximum 45% reduction for heavily agentic workloads. Your workload decides where it lands.

The cost test comes first. The reported Terminal-Bench 4.0 score for Fable 5.1 was 55.8%, while Fable 5 scored 42%. The reported Terminal-Bench-Science 0.1 score for Fable 5.1 was 52.6%, versus 24.7% for Fable 5.

• Measure cache-hit share.
• Recompute cost per completed task at the new cache-read rate.
• Migrate only cache-heavy workloads; otherwise restructure prompts or stay.

Cache-hit share decides the migration

Pull recent run logs for the agent you are considering. Separate input tokens into cached and non-cached portions. Include system prompts, tool schemas, retrieved context, and conversation history. The output is a single percentage for each task type.

Track cache reads and cache writes separately. A read is cheap only when the same context is reused. A write is a setup cost. An agent that rebuilds context every run will not show the read cut in the bill.

A low percentage leaves the total nearly unchanged. A high percentage can change the budget. Cache-write rates for Fable 5.1 stayed at $12.50 per million tokens for five-minute caches and $20 per million tokens for one-hour caches. A workload that writes a lot and reads little still pays the write side.

The decision number is the hit share, not the rate.

Cost per completed task is the real number

Use the same task set as the baseline. Price non-cached input at the standard input rate, cached input at the new cache-read rate, and output at the standard output rate. Add cache writes when the run creates or refreshes cached content. Divide the total by completed tasks, not API calls.

Use the task distribution, not average token counts, when tasks vary widely. Long-running agent economics depend on retries, tool loops, and context growth. A failed restart can consume more cache writes and input tokens than a clean run.

Include failed tasks when they consume tokens. A completed task is the unit you pay for in production. An aborted run still spends tokens and can raise cost per completed task even when the per-call price falls.

The next step follows the number: lower cost gives the migration a cost case, flat cost moves the decision to quality, latency, or reliability, and higher cost means the cache cut is not doing the work.

Only cache-heavy workloads earn the move

Move workflows where cached input is a large share of spend and the recomputed cost per completed task is lower. Keep the rest on the current model until the prompt structure changes. The migration list should show a measured cost difference for each workflow.

A cache-heavy workload usually has stable prompt content, repeated tool schemas, long shared context, or many reads of the same retrieved material. A cache-light workload changes its prompt every run or sends small, unique inputs.

For cache-light workloads, restructure prompts before switching. Put stable content first, trim volatile context, and reduce repeated large payloads. Raise the cache-hit share or lower the non-cached input, then recompute the cost per completed task again.

For heavily agentic workloads, test the real task mix before moving. The cache cut pays when the agent reads the same large context many times.