Fable 5.1 vs Opus 5 Cost: When Cache Changes the Answer

Claude Opus 5 is the less expensive default. On Anthropic's standard API, it charges half as much as Fable 5.1 for fresh input, output, and prompt-cache writes. Fable 5.1 has one important advantage: a cache read costs $0.25 per million tokens, half of Opus 5's $0.50.

That exception is large enough to reverse the bill in a context-heavy agent, but only under demanding conditions. A short request, a low cache-hit rate, or a workload that Opus completes reliably will usually favor Opus. Fable can cost less per completed task when repeated context dominates the token mix or when it finishes with substantially fewer turns, retries, and tokens.

As of September 4, 2026, Anthropic's own model overview recommends starting with Opus 5 for most workloads. It positions Fable 5.1 for demanding reasoning and long-horizon agentic work, or for cases where Opus at higher effort still misses the evaluation target.

Start with the current Claude lineup, not a prestige ranking

The four headline models cover different operating constraints. All accept text and images, produce text, and support multilingual use, vision, and tools. Their standard direct-API prices and capacity are different:

Those are token prices for the standard first-party Claude API. They are not Claude Pro or Max subscription allowances, and they do not promise the same price on Amazon Bedrock, Google Cloud, Microsoft Foundry, or an independent gateway.

Haiku belongs on the shortlist when volume and latency matter and failures are easy to catch. Sonnet is a broad candidate for everyday coding, analysis, content processing, and tool use. Opus is a sensible capability baseline for complex engineering and expensive mistakes. Fable should enter after a measurable capability gap appears—not because the name sits at the top of the range.

List price and task cost answer different questions

The current Claude pricing table separates five token classes that matter to the Fable-versus-Opus decision:

Prompt caching does not turn every input token into a cheap read. Anthropic's prompt-caching guide explains that a stable prefix is written first and reused only while the prefix still matches and the entry remains valid. The default lifetime is five minutes; a one-hour entry costs more to write. Changes to tools, the system prompt, or earlier messages can reduce the hit rate that a spreadsheet assumed.

Opus 5 fast mode is another separate bill. This first-party research preview charges $10/$50 for lower latency. It should not replace the standard $5/$25 Opus rate in a normal comparison, and it is not available on the Batch API.

The exact cache break-even point

For an equal-token comparison, let each variable represent millions of tokens:

• I: fresh input;
• O: output;
• W5: five-minute cache writes;
• W1: one-hour cache writes;
• C: cache reads.

The standard token bills are:

Fable = 10I + 50O + 12.5W5 + 20W1 + 0.25C

Opus = 5I + 25O + 6.25W5 + 10W1 + 0.50C

Solving the difference gives one useful boundary:

Fable costs less only when C > 20I + 100O + 25W5 + 40W1

Suppose a task sends 100,000 fresh input tokens and produces 20,000 output tokens, with no new cache write. Fable needs more than four million cache-read tokens before its lower read rate offsets the higher input and output rates. At 1.5 million cached tokens, the arithmetic is about $2.38 for Fable and $1.75 for Opus. At six million cached tokens, it becomes about $3.50 versus $4.00.

A cache write raises the threshold further. The first turn can therefore favor Opus even when later turns favor Fable. The useful unit is not “one call” but the complete sequence required to obtain an accepted result.

This formula deliberately holds token quantities constant. Real models can produce different output lengths, take different tool paths, and finish after different numbers of turns. Use it to test a cost sheet, not to predict the winner.

The Devin result is a specific, explainable reversal

Cognition reports a lower Fable task cost in Fable 5.1 in Devin: $2.68 for Fable 5.1 versus $3.51 for Opus 5 on FrontierCode 1.1 Extended at medium thinking. It says Fable used 33% fewer tokens on those tasks. More than 95% of both models' tokens were cache reads; a typical task read roughly three million cached tokens with Fable and 4.5 million with Opus.

Those conditions activate both ways Fable can overcome its higher sticker price. Cached context dominates the bill, where Fable has the lower rate, and Fable uses fewer tokens in the reported harness. There is no contradiction between that result and Opus having cheaper fresh input and output.

There is also no universal guarantee. FrontierCode's tasks, Devin's tools and caching layout, the medium thinking setting, and Cognition's acceptance criteria are part of the result. A support classifier, a document pipeline, or a different repository can produce another token mix and another pass rate.

Anthropic's launch claim needs a separate comparison label. In its Fable 5.1 announcement, Anthropic estimates roughly 25% lower cost on typical workloads and up to about 45% on highly agentic work relative to Fable 5, based on four weeks of August 2026 usage at default effort. Those percentages are not discounts relative to Opus 5.

Measure the cost of an accepted task

Select a small, representative set of production work: common requests, difficult cases, and failures that matter. Give Fable 5.1 and Opus 5 the same input, tools, permissions, timeout, stop rule, and acceptance test. Comparing a polished Claude Code configuration with a bare Messages API call measures products and harnesses, not just models.

Capture enough information to explain the result:

Then calculate:

cost per accepted task = (tokens + tools + review + retries + rework) / accepted tasks

Opus is the better route when it clears the bar and Fable's higher fresh-input and output rates add cost without reducing failure or labor. Fable earns a narrow route when repeated context is genuinely cached and its capability reduces enough turns, tokens, or expert repair to lower the complete cost.

This also keeps Sonnet and Haiku in the decision. A routine workload should not become an Opus-versus-Fable contest if Sonnet passes at $2/$10, and a high-volume, automatically verified transformation may belong on Haiku. Test adjacent candidates rather than staging a four-model tournament for every endpoint.

Check the billing surface before changing production

The arithmetic above applies to standard token-billed requests on the direct Claude API. Batch processing, Opus fast mode, the documented US-only inference multiplier, partner-cloud regional pricing, gateway markups, and Claude subscriptions create different bills. Confirm the model ID and the actual usage fields on the platform that will receive production traffic.

Model switches can also change cost indirectly. Fable 5.1 uses adaptive thinking that is always on and has specific rules for tool choice and preserved thinking. Verify effort, max_tokens, cache hits, tool behavior, refusal handling, and rollback before expanding traffic. The dedicated Claude Fable 5.1 guide covers the model and migration contract in detail.

For most teams, the defensible first test is Opus 5—or Sonnet 5 when the failure budget permits it. Add Fable 5.1 only where measured failures or end-to-end cost justify the escalation.