Fable 5.1 vs Fable 5 vs Opus 5: It's All in the Cache

Claude Fable 5.1 landed on 1 September 2026 billed as the cheaper option. It is worth reading the price table before acting on that. The base rates did not move at all — input is still $10 per million tokens and output is still $50, exactly what Fable 5 costs. The entire reduction sits on one line: cache reads dropped from $1 to $0.25. That is a real 75% cut, and it is what Anthropic’s estimate of 25% for typical workloads and up to roughly 45% for highly agentic ones is built on. But it reaches your invoice only in proportion to how much of your input is actually served from cache.

The table everyone is skipping

Anthropic publishes five rates per model, and most summaries quote one. Here is the whole thing, in dollars per million tokens:

Read the two Fable rows side by side. Four of the five numbers are identical. Fable 5.1 is not a cheaper model — it is the same model at the same price, with a cheaper cache.

What that means for your bill

Cache savings compound with reuse, which is why one rate change lands so differently across workloads. Take one 1M-token prefix — a codebase, a document set, a long system prompt — write it to cache once, then read it back N times. The no-caching column is what the same N+1 full passes cost at $10 per million:

Assumes a single 5-minute cache write with no expiry between reads — reads refresh the entry for free, so this holds for a continuously active agent. If the prefix goes cold and has to be rewritten, add $12.50 per rewrite to both Fable columns; see the TTL section below.

At one read the two versions are within 6% of each other. At a hundred reads Fable 5.1 costs a third of what Fable 5 costs. Nothing about the model changed between those rows — only how many times you came back to the same cached context.

This is why Anthropic’s two figures sit so far apart. A workload that touches cached context a handful of times per session lands near 25%. A long-running agent that re-reads the same repository hundreds of times over an afternoon is what the “up to approximately 45%” end of that range describes. To know which applies to you, count cache hits, not requests.

Take a concrete case: 1M input tokens of which 90% are cache reads, plus 20K output.

• Fable 5 — $0.90 cache + $1.00 fresh input + $1.00 output = $2.90
• Fable 5.1 — $0.225 cache + $1.00 fresh input + $1.00 output = $2.225

A 23% saving, lining up with Anthropic’s estimated 25% for typical usage. Note where it comes from: the cache line fell by three quarters, but it was only about 31% of the bill to begin with.

The TTL choice costs more than the price cut

Anthropic sells two cache-write durations, and this is the part most cost models leave out. A 5-minute cache write is $12.50 per million tokens; a 1-hour write is $20.

The 5-minute clock is looser than it sounds. Anthropic’s documentation is explicit: “The cache is refreshed for no additional cost each time the cached content is used.” The lifetime is measured from the start of each request that writes or reads the entry. An agent hammering the same prefix continuously keeps a 5-minute entry alive on reads alone and pays for exactly one write.

The money leaks in the gaps. A planning turn, slow tool calls, a human reviewing a diff — any quiet stretch longer than five minutes lets the entry expire, and the next call pays the write again. An hour of that pattern costs at worst 12 × $12.50 = $150, against $20 for a single 1-hour write.

That worst case is a $130 gap per prefix per hour, larger than the entire cache-read reduction Fable 5.1 shipped — and it is decided by the shape of your traffic, not by which model you picked. A continuously busy agent may never reach it; one that idles between turns will. Work out which you are running before you bank the new pricing.

Fable 5.1 versus Opus 5: the trap in the numbers

One line does invert: Fable 5.1 reads cached tokens at $0.25, half of Opus 5’s $0.50 — even though Fable’s base input is twice Opus 5’s and its output is twice as expensive too.

It is tempting to conclude that cache-heavy agents should now move up to Fable. The algebra says otherwise. With input I, cache-read fraction f, and output O, Fable 5.1 is cheaper only when:

10(1−f)·I + 0.25f·I + 50·O  <  5(1−f)·I + 0.50f·I + 25·O
→  5I − 5.25f·I + 25O < 0
→  f > (5 + 25·(O/I)) / 5.25

Set output to zero — a workload producing nothing at all — and you still need f > 95.2%. Every token you actually generate pushes the threshold higher. There is no realistic mix in which Fable 5.1 undercuts Opus 5 on total cost.

The worked example above shows it plainly: Fable 5.1 costs $2.225 and Opus 5 costs $1.45 for identical traffic. One cheaper line does not carry a model that is 2x on everything else.

So the decision does not change, it just costs less on one axis: choose Fable 5.1 when you need its capability, not to undercut Opus 5 on cost. What the new cache pricing buys is that the premium for staying on a Mythos-level model through a long agentic run is smaller than it was on Fable 5.

An independent measurement points the same way

Artificial Analysis, which supported Anthropic with pre-release evaluation of the model, published cost-per-task figures that land on exactly this point. Their headline: Fable 5.1 “tops the Artificial Analysis Intelligence Index but costs 20% more per task than Fable 5 despite a 75% cache read price cut”.

The reason is the half of the equation this article has been circling: Fable 5.1 uses roughly 1.7x the output tokens of Fable 5 at max effort. Output is the $50 line, and no discount on the $0.25 line reaches it. A cheaper cache did not make it a cheaper model — measured end to end on the same task set, it got 20% more expensive than its own predecessor and 1.6x Opus 5.

It also scored highest of anything they have measured. That is the actual trade: you are buying the top of the index and paying more per task than before, not less.

What Anthropic says changed besides price

The Claude Fable product page is specific about behavior rather than benchmarks. It says Fable 5.1 “avoids easy-seeming shortcuts, fixes the root causes of problems rather than the symptoms, and keeps you updated as it works”. (The separate launch announcement words this differently, so it is worth naming which page a quote comes from.) The strongest concrete claim is not Anthropic’s own. Marquis Wong, Principal AI Engineer at IMC, is quoted on the same page: “Using Claude Code, it correctly identified the root cause of every broken build we tested, across all the effort levels.”

The first is the vendor describing its own model; the second is a customer describing its own tests. Neither is a benchmark, and both are worth reproducing on your own broken builds before you rewrite a routing policy around them. Anthropic’s announcement page does carry a full benchmark table — Terminal-Bench-Science, Terminal-Bench, OSWorld, HLE and more. This piece deliberately leaves those alone and sticks to the money, because the cost story is the one the launch coverage got wrong.

One correction, since it has spread widely: several write-ups presented “fewer false positives” as a Fable 5.1 feature. That change is dated 7 August 2026 and applies to Claude Fable 5’s biology safeguards, where Anthropic reports the update “reduced biology-related fallbacks by about 85% across our product surfaces”. It is an earlier, separate update with nothing to do with this release.

Specs, and a detail that gets overlooked

The model ID is claude-fable-5-1. Context and output limits are identical across all three, so the differentiators are price, behavior, and that last row — Fable 5.1 has the most recent reliable knowledge cutoff of the three models here, a month ahead of Opus 5 and five months ahead of Sonnet 5. For anything touching recent library versions or APIs, that gap does real work.

Mythos 5.1 appears in the same pricing table at identical rates. Anthropic’s announcement states the two “are the same underlying model” and that the gap between them “reflects the tasks on which our earlier, less precise cyber safeguards intervened”. Access runs through trusted-access programs, and the two are not at the same stage: the Life Sciences Verification Program is an invite-only beta that grants access now, while the Cyber Verification Program “will include access to Mythos models in the near future”. Availability is currently limited to a set of US organizations. Not a drop-in option for general use.

Picking one

Then go and check your cache TTL. On an hour-long agent that single setting swings more money than the price change this article is about.

Availability

As of 2 September 2026 the Ofox catalog carries anthropic/claude-fable-5, anthropic/claude-opus-5 and anthropic/claude-sonnet-5 at Anthropic’s list prices with no markup. Fable 5.1 is not in the catalog yet. If you want to run this comparison against real traffic today, Fable 5 and Opus 5 are both reachable on a single Ofox key through the OpenAI-compatible and native Anthropic protocols.

Prices and specifications verified against Anthropic’s pricing, model documentation, product page and launch announcement on 2 September 2026. Cost-per-task figures from Artificial Analysis, retrieved the same day.