Claude Fable 5.1's Real Headline Is a 75% Cache Price Cut: Here's the Math for Your Agent

Anthropic shipped Claude Fable 5.1 and Claude Mythos 5.1 on September 1 with token prices frozen exactly where Fable 5 left them, $10 per million input, $50 per million output. That freeze means the entire economics of this release lives one row down on Anthropic's pricing page: prompt-cache reads fall from $1.00 to $0.25 per million tokens, a 75% cut on the line item that dominates agent bills.

According to Anthropic's announcement, both models are live on the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry, and Claude Platform on AWS. And one correction to coverage you may have seen elsewhere: the 1M-token context window and 128K maximum output are not new, both carry over unchanged from Fable 5, per the model overview. The price cut is the story.

What actually changed (and what didn't)

Unchanged from Fable 5: the 1M context window (default and maximum, at standard per-token pricing with no long-context premium), the 128K output ceiling, always-on adaptive thinking, $10/$50 token prices, cache-write prices ($12.50/MTok for the 5-minute tier, $20 for the 1-hour tier), the 512-token cache minimum, and batch pricing at $5/$25. Fable 5 itself stays available, at $1.00 cache reads, with no stated plan to lower it.

Changed: cache reads on the 5.1 models are priced at 0.025x base input, per the pricing page's footnote, where every other model uses the standard 0.1x multiplier. Alongside that come capability gains, three breaking API changes, and five additive ones, one of which is the lower cache-read price (more below).

On capability, according to Anthropic, Fable 5.1 scores 52.6% on Terminal-Bench-Science 0.1 against Fable 5's 24.7%, 55.8% against 42.0% on Terminal-Bench 4.0, and posts smaller gains on Humanity's Last Exam without tools (60.9% vs 57.8%), CursorBench 3.2.0 (73.4% vs 70.5%), and OSWorld 2.0 strict (41.7% vs 36.1%). These are vendor-reported figures, and Anthropic's comparisons to Opus 5 and OpenAI's GPT-5.6 Sol come from the announcement alone.

The one-line identity that decodes Anthropic's savings claims

The following is TechCurrent analysis, derived only from the verified prices above.

With token prices unchanged, your savings on Fable 5.1 equal 75% of whatever share of your bill was cache reads.

The identity is exact, and it turns Anthropic's marketing into data. The announcement estimates workloads will run "25% less" for typical use and "up to approximately 45%" cheaper for highly agentic work, without publishing the assumptions behind either number. Run the identity backwards and the assumptions fall out: a 25% saving implies cache reads were roughly a third of those bills; 45% implies roughly 60%. The reverse-engineering is our inference, but the arithmetic isn't, Anthropic has effectively told everyone what agent bills are made of.

The math for your agent loop

TechCurrent analysis: worked examples at the verified Claude API prices; partner-platform (Bedrock, Google Cloud) cache pricing wasn't separately verified.

Example A: a moderate coding-agent session. Assume a 20K-token stable prefix (system prompt plus tools), 50 requests, conversation history growing 3K tokens per request, ~1K output tokens per request, everything cache-friendly. The first request writes the prefix; the 49 that follow re-read it, plus the ever-growing history: 49 × 20K for the prefix plus 3K × (1+2+…+49) of re-read history ≈ 4.66M tokens of cache reads. On Fable 5 that's ~$4.66 in reads; on 5.1 it's ~$1.16. Cache writes (~170K tokens, ~$2.13) and output (~$2.50) don't move. Session cost: ~$9.28 → ~$5.79, about 38% cheaper.

Example B: the same shape at 100 requests. Cumulative reads reach ~16.83M tokens; the session goes from ~$25.83 to ~$13.21, about 49% cheaper.

Savings grow with session length because cache reads grow quadratically with turns, every turn re-reads all prior turns, while writes and output grow linearly. The longer your agent runs, the more cache-read-dominated its bill becomes. Honest caveat: Example B exceeds Anthropic's "up to ~45%" precisely because it's idealized, nearly pure cache-read. Real sessions carry uncached input, cache misses, and tool fees, web search alone is $10 per 1,000 searches on the pricing page.

When parking 1M tokens in context beats RAG

TechCurrent analysis. Take a 500K-token corpus, a monorepo, a document set. Writing it to the 5-minute cache costs $6.25 once. After that, every request re-reads the entire corpus for $0.125, down from $0.50 on Fable 5, versus $5.00 uncached. And the entry stays warm under steady traffic: according to the prompt-caching docs, "the cache is refreshed for no additional cost each time the cached content is used," so requests landing within five minutes of each other keep it alive indefinitely. At 2.5 cents per 100K cached tokens per request, whole-corpus-in-context starts competing with an embedding pipeline for bursty, session-shaped workloads where the model benefits from seeing everything.

Caveats, disclosed plainly: traffic with gaps longer than the TTL pays the write again (the 1-hour tier costs $20/MTok, $10 for this corpus, and needs fewer re-reads to pay off). First prefill of 500K tokens has real latency. RAG still wins for corpora past 1M tokens, sparse access patterns, low request rates, and anything that must persist across sessions.

The API now enforces the shape the pricing rewards

The what's-new page lists three breaking changes: forced tool use (`tool_choice` of `"any"` or `"tool"`) now returns a 400; thinking blocks are model-bound one-way (5.1 reads earlier models' blocks, not vice versa); and editing earlier turns (system prompt, tools, prior messages) invalidates a 5.1 thinking block. That last check is enforced with a 400 error for accounts created on or after August 31, 2026 (they can opt to have the block silently dropped instead); older accounts only have the mismatch recorded unless they opt in. Either way, the direction is set: conversations become effectively append-only.

And append-only (this is TechCurrent analysis, and the observation other coverage misses) is exactly the cache-stable request shape that earns the $0.25 rate. Mutate history and you invalidate the cache anyway; now the API can error instead of silently billing you at $10/MTok. The constraint and the discount are two faces of one design. (Mythos 5.1 skips the prefix-binding check.) The additive features, all currently in beta, point the same direction: per-message effort changes without cache invalidation, turn-scoped system messages, and progress updates between tool calls, plus provenance watermarking on output.

Two tiers, second generation

This is now a pattern, not a one-off: for the second consecutive release, Anthropic shipped one model behind two safeguard tiers, Fable 5.1 generally available with classifier-based safeguards, Mythos 5.1 offered only to approved participants in Project Glasswing. Alongside it, the announcement describes two formal access paths: a Cyber Verification Program giving verified security professionals access to certain Opus- and Sonnet-class models with reduced cyber safeguards for defensive work, and a Life Sciences Verification Program, developed in partnership with the US government, to enable access to Mythos 5.1's advanced biology capabilities.

The practical consequence (TechCurrent analysis): refusal handling is now baseline plumbing for anyone choosing this API. A Fable 5.1 refusal arrives as HTTP 200 with `stop_reason: "refusal"`; fallback configuration supports Opus 4.8 and Opus 5 targets, a `fallbacks: "default"` beta that routes by refusal category, fallback credit, and unbilled pre-output refusals. Verified security and bio teams get a formal access path instead of prompt-engineering around blocks. Anthropic's own tuning claims, attribute exactly, are that Claude Code users can expect "around 60% fewer interventions per session" from its cyber safeguards, and that its latest biology safeguards, applied to both Fable 5.1 and Fable 5, "fire 85% less often for benign requests." That's a vendor acknowledging the tax safeguards impose on legitimate users.

What to do this week

If you run agents on Fable 5, the upgrade is mostly a price cut, but grep your codebase for `tool_choice` with `"any"` or `"tool"`, and audit any client-side history-editing logic, before flipping the model ID. Those break. (Server-side compaction and context editing are explicitly exempt from the prefix check.) Migration is also the only path to the $0.25 rate; Fable 5 keeps $1.00 with no stated plan to change.

There's a second price-favorable move on the same pricing page: Sonnet 5's increase to $3/$15, previously scheduled for September 1, "will not occur", $2/$10 is now the standard price. Taken together, the signal is that Anthropic is competing on inference economics, which matters doubly for cost-sensitive teams on metered APIs, African startups among them: a 38–49% cut in agent running costs changes what's viable to ship.

The clock on this generation is already set: per the model page, Fable 5.1 won't be retired before September 1, 2027. The nearer question is whether the 0.025x cache multiplier stays a 5.1 exclusive or becomes the new standard, and every team that restructures its agents into append-only, cache-stable shapes this quarter is betting it's the latter.

“With token prices unchanged, your savings on Fable 5.1 equal 75% of whatever share of your bill was cache reads.”

Reporting by TechCurrent Staff · TechCurrent