🔍 Read the full analysis: Twice As Expensive AI Delivers 5.7% More Intelligence on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on office and shipping supplies

• Business-only prices and quantity discounts
• Tax-exempt purchasing
• Multiple users, one account, clear invoices

Create a free Business account →Free to register

As an affiliate, we earn on qualifying purchases.Hide

Anthropic’s Claude Fable 5 costs twice as much as Opus 4.8, while its score on Artificial Analysis’s Intelligence Index is 5.7% higher, according to a July 2026 analysis of public pricing and benchmark data. The analysis also finds that the public case for Fable 5 improving enterprise productivity rests on one unaudited customer account, not a controlled study comparing workers with and without the model.

Anthropic lists Fable 5 at $10 per million input tokens and $50 per million output tokens. Opus 4.8 is listed at $5 and $25, respectively; Sonnet 4.6 is listed at about $3 and $15. Those figures make Fable 5 the costliest model in Anthropic’s active lineup, according to the source analysis, which says its pricing is confirmed by Anthropic’s own materials.

Artificial Analysis reports an Intelligence Index score of 64.9 for Fable 5, compared with 61.4 for Opus 4.8. On GDPval-AA, a benchmark measuring knowledge-work performance, the reported Elo ratings are 1,932 and 1,890. The source analysis describes that as a 2.2% gap, while cautioning that Elo is a rating scale and the percentage framing does not directly convey head-to-head win probability.

The analysis estimates the cost of a full Intelligence Index run at about $9,940 for Fable 5 and $4,970 for Opus 4.8. It also notes that the models’ relative performance varies by task: Fable 5 reportedly records top results in five of the benchmark’s 10 sub-benchmarks. Buyers whose workloads align with those tasks could see a different value calculation from one based on the overall score alone.

The Premium Depends on the Workload

The figures put a measurable price beside a relatively small difference in one aggregate benchmark. For teams choosing between the two models, twice the list price may be difficult to justify if their work resembles the average tasks represented by the composite score. But a composite is not a direct estimate of a company’s output, and it cannot establish whether Fable 5 will save enough staff time to offset its higher token costs.

The task-level results may matter more to some buyers than the average. If a company relies on work where Fable 5 performs particularly well, a higher cost could be worthwhile; if its use cases do not benefit from those strengths, the premium may yield less value. The source analysis does not provide company-specific cost or productivity estimates, so organizations would need to test their own workloads and account for usage patterns before drawing a budget conclusion.

That distinction matters because benchmark performance and workplace productivity are different measures. A model can score well on a test without proving that it makes employees faster or produces better work in a particular organization. The available evidence described in the analysis does not close that gap.

Pricing Is Clearer Than Productivity

The source analysis, published by Thorsten Meyer on ThorstenMeyerAI.com in July 2026, separates public pricing and benchmark results from claims about productivity. It says Anthropic’s launch post, platform pricing documentation and product page confirm Fable 5’s rates, with secondary reporting providing additional corroboration. Its benchmark comparisons are attributed to Artificial Analysis, an independent model-evaluation service.

Fable 5’s headline rates are not the only available rates. The analysis says Anthropic offers prompt-cache reads at $1 per million tokens, a 90% discount from the standard input rate, and batch pricing of $5 per million input tokens and $25 per million output tokens. Using a stated mix of seven cache-hit units, two input units and one output unit, it calculates a blended cost of $7.70 per million tokens. That is an illustrative usage mix, not a universal rate; actual bills depend on how a customer uses the model.

The source also says Fable 5 and Mythos 5 use the same underlying weights and have the same pricing, with safeguards differing between the products. It places the general availability date at around July 1, after a June 9 launch. These details describe the product and its costs; they do not independently verify productivity claims.

“In a 50-million-line Ruby codebase, Fable 5 performed a codebase-wide migration in a day that “would otherwise have taken a whole team over two months by hand.””

— Anthropic’s launch post, quoting a Stripe customer account

No Controlled Productivity Results Cited

The source analysis reports zero controlled human-baseline productivity studies for Fable 5 and one enterprise proof point: the Stripe account in Anthropic’s launch materials. It says the story has no published methodology, audit or independent replication. The account confirms that a productivity claim was made; it does not establish a generally applicable or independently measured time saving.

The analysis does not provide enough detail to assess how the Stripe migration was scoped, what human review or other tools were involved, or whether the comparison was measured under equivalent conditions. It also does not report a controlled study establishing Fable 5’s effects on output quality, total task time or costs across a range of businesses. Those points remain unclear from the cited material.

Benchmark scores carry limits, too. The reported aggregate results do not show how Fable 5 performs on every company’s tasks, and the source notes stronger results in five of 10 sub-benchmarks without establishing that a buyer’s workload will match them. The figures should be read as benchmark comparisons, not forecasts of staff productivity or financial returns.

Buyers Need Workload-Level Tests

The next useful evidence would be controlled, independently reviewed productivity studies that compare relevant work with and without Fable 5, disclose methods and measure both quality and time. Replication across organizations and task types would help show whether the Stripe account is representative or an isolated example.

Until such results are available, organizations considering the model can compare its pricing with alternatives and run pilots on their own tasks. A useful evaluation would track token usage, completion time, human review and correction, and the quality of the final work. That would let buyers compare the cost of Fable 5 with Opus 4.8 using their own workload rather than relying on a single benchmark average.

The source analysis does not identify a scheduled productivity study or a new public results date. For now, the price and benchmark scores are published figures, while broader claims about enterprise gains remain unproven by controlled public evidence cited in the report.

Key Questions

How much does Claude Fable 5 cost?

Anthropic’s listed rates are $10 per million input tokens and $50 per million output tokens. The source analysis also reports discounted rates for prompt-cache reads and batch processing, so a customer’s actual cost depends on usage.

Is Fable 5 5.7% better than Opus 4.8 at every task?

No. The 5.7% figure refers to the difference between the models’ reported Intelligence Index scores, 64.9 for Fable 5 and 61.4 for Opus 4.8. It is not a claim that Fable 5 performs 5.7% better on every task or for every user.

What evidence supports the productivity claim?

The source analysis identifies one public example: a Stripe customer account relayed in Anthropic’s launch post about a code migration. It says the account has no published audit, methodology or independent replication, and is not a controlled productivity study.

Could Fable 5 still be worth the higher price?

Possibly, depending on a buyer’s tasks and usage. The report says Fable 5 leads in five of 10 sub-benchmarks, but it does not show whether a particular organization’s work benefits from those results. Testing the model on relevant tasks is needed to judge value for a specific use case.

Source: ThorstenMeyerAI.com