Claude Fable 5.1 and the New Economics of Cache Hits

Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on September 1. The benchmarks will get attention, but the line I care about is the pricing: cache reads on Fable 5.1 now cost $0.25 per million tokens, 75% less than before, while input and output stay at $10 and $50 per million. Anthropic estimates typical workloads get about 25% cheaper than on Fable 5, and highly agentic work up to around 45% cheaper. For coding agents, that changes where the bill comes from.

What shipped#

• Fable 5.1 and Mythos 5.1 are the same model with different safeguard levels. Fable 5.1 is generally available; Mythos 5.1 is offered only through Anthropic's trusted access programs, currently to a set of US organizations.
• The API model ID is claude-fable-5-1, and the model is available on all platforms, including AWS, Google Cloud and Microsoft Azure.
• On Terminal-Bench 4.0, Fable 5.1 scores 55.8% against 42.0% for Fable 5. Mythos 5.1 reaches 60.9%.
• Effort defaults to High in Claude Code and Medium in Claude Cowork and on Claude.ai. Anthropic says Low or Medium effort gets results similar to or better than Fable 5's at much lower cost.

The ratio that changed#

Anthropic's chart note explains the 45% figure: its "highly agentic" workload is context-heavy, tool-heavy work where cache reads make up most of the cost. That is an agent loop. Every turn resends the tools, the system prompt and the conversation so far; only the tail is new.

The prompt caching docs list Fable 5.1 at $10 per million base input, $12.50 for 5-minute cache writes, $20 for 1-hour writes and $0.25 for cache hits. Fable 5 hits were $1. Quick arithmetic:

• An uncached input token now costs 40 times a cache hit. On Fable 5 it was 10 times.
• A 5-minute cache write costs 50 times a hit.
• Re-reading a 100k-token prefix costs $0.025 per turn as a hit and $1.00 as an uncached read.

So for agents, hit rate matters more than list price. A harness that quietly breaks its cache every few turns gets little of the saving and pays write premiums on top.

Structuring prompts for the cache#

The docs build the cache prefix in a fixed order: tools, then system, then messages. A change at one level invalidates that level and everything after it. For long-running agents:

• Freeze tool definitions per session. Changing tool names, descriptions or parameters invalidates the whole cache. If an MCP server's tool list shifts mid-run, you pay for it.
• Keep the system prompt static. No timestamps, request IDs or "current file" snippets at the top. Put volatile context in the latest message.
• Don't flip top-level thinking or effort settings mid-session. The docs list both among the things that invalidate cached message blocks. For cheap Low-effort sub-steps, use Fable 5.1's per-message effort beta, which keeps the cached prefix, or give that work its own session.
• Treat history as append-only. Rewriting or summarizing earlier turns moves the prefix. Relatedly, new API accounts created from launch day onward can no longer manually edit Claude's prior context while keeping its prior thinking transcript. Existing accounts are exempt for now; Anthropic says the change will reach all users with future model releases.
• Place breakpoints deliberately. You get up to 4 cache breakpoints, and a read looks back at most 20 blocks per breakpoint. If a growing conversation pushes a single trailing breakpoint 20 or more blocks past the last write, the lookback misses it; a second, earlier breakpoint avoids that.
• Pick the TTL for the pause, not the task. The default 5-minute cache is refreshed at no extra cost each time it is used. If your agent waits longer on CI or human review, a 1-hour write costs $20 per million instead of $12.50 but avoids a fresh write after every pause.

{
  "type": "text",
  "text": "<stable system prompt>",
  "cache_control": { "type": "ephemeral", "ttl": "1h" }
}

Measure it#

Every response reports cache_read_input_tokens, cache_creation_input_tokens and input_tokens. Log all three per turn and compute a hit ratio per session. A creation spike mid-session means the prefix changed or the cache expired. At these prices, hit ratio belongs on the same dashboard as latency and error rate.

The safeguard change#

Anthropic says Claude Code users can expect around 60% fewer cyber-safeguard interventions per session than under Fable 5's previous safeguards, and Fable 5.1 can now be used to discover software vulnerabilities but not to develop exploits. Penetration testing, exploit generation and binary-based vulnerability scanning still get redirected to Opus models. As someone working through CEH (AI edition) material, I read that as: defensive review is in scope, offensive tooling is not.

What I'm doing about it#

For teams running agents on Claude, this is my Monday list:

1. Pull a week of usage and compute the cache-read share of input tokens per workflow.
2. Diff the system prompt and tool list between turns of one real session. Anything that changes is a cache bug.
3. Try Low and Medium effort for routine sub-tasks, via per-message effort or separate sessions, before defaulting everything to High.

The 45% figure is Anthropic's estimate, not a guarantee. You only get close to it if your prefix holds still.