Reprice long-running agents after Fable 5.1's 75% cache-read cut

If your agent re-sends the same context every turn, Fable 5.1 changes the bill. Anthropic made Claude Fable 5.1 generally available on September 1, 2026, with input at $10 per million tokens and output at $50 per million tokens, unchanged from Fable 5, while Claude Mythos 5.1 remained limited to vetted cybersecurity and life-sciences partners.

The cut changes cached context, not fresh tokens

Cache reads fell 75%, from $1.00 to $0.25 per million tokens. The cut applies to context an agent pulls back into later turns. Cached context can be at least half of token usage in some long-running tasks. Keep the cache-read line as a separate budget line.

Use the 50/50 rule before you touch the prompt

Long-running agents often re-send the same system prompt, tool results, file excerpts, and task state. The 50/50 cache-read rule: if cache reads are at least half of token usage, recompute the job's unit cost before changing model or prompt, then add output and safety-retry overhead. Keep the rule in the runbook next to the job definition.

Anthropic projected that the cache-price cut would trim Fable 5.1's effective cost by roughly 25% on typical workloads and up to 45% on heavily agentic ones. Treat the lower number as a baseline and the higher number as the case for loops that keep the same context alive. Agentic loops reuse context more often than one-shot requests. A chat-style request may pay mostly for fresh input and output. A multi-step agent may pay again and again for the same context. Jobs that run for many turns, call tools, and keep a growing task state are the ones to check first. Short prompts that finish in one pass are less likely to benefit. When a job is close to the cutoff, look at the trend over several runs, not a single run.

Recompute the job's unit cost and throughput

Treat the rule as a check, then verify with your own logs. Start with the token mix for the job you care about, not for the whole account. Recompute unit cost and throughput for that job. If a workload already runs on Fable 5, reprice it before deciding whether to move it to Fable 5.1. Build a before-and-after number you can defend in a budget review. Do not mix jobs with different cache behavior in one number. A research agent and a form-filler may look similar in token count and very different in cache-read share. Record the unit cost, throughput, and cache-read share for each job.

• Split recent token usage into cache reads, fresh input, and output.
• Calculate the cache-read share for the job, not for the whole account.
• Reprice the job with the Fable 5.1 rates.
• Compare the new unit cost with the old one before changing model or prompt.

Take the before-and-after number to decide which workloads move to Fable 5.1 and which stay on a cheaper model or a shorter prompt.

Add output and safety retries to the new unit cost

Add output tokens and safety retries to the new unit cost. Medical or biology requests get 85% fewer safety interventions, and some users can see about 60% fewer cybersecurity-related interventions per session. Track retries separately from clean runs. A job that looks cheap on the first pass can become expensive if it keeps hitting a blocked turn. Blocked turns can add output tokens and another pass through the loop. Workloads that clear the old budget can fail the new one once retries are included.

Set a budget threshold before shipping: keep the workload on Fable 5.1 when the recomputed unit cost, including output and safety retries, stays under it; move it when the cost clears it. Record the threshold and retry rate next to the job name.