🔍 Read the full analysis: AI's Costly 5.7% Intelligence Gain, Explained on ThorstenMeyerAI.com
Get business pricing on monitors, keyboards and dev gear
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
Anthropic’s Claude Fable 5 is priced at $10 per million input tokens and $50 per million output tokens, twice Opus 4.8’s rates. Artificial Analysis reports a 5.7% lead on its Intelligence Index, but the source material identifies only one unaudited customer account—not a controlled study—as public evidence of productivity gains against a human baseline.
Anthropic’s Claude Fable 5 costs twice as much per token as Opus 4.8, while scoring 5.7% higher on Artificial Analysis’s Intelligence Index, according to pricing information and benchmark results cited in a July 2026 analysis. The comparison highlights a gap for businesses weighing the premium model: its prices are publicly documented, but the analysis identifies one unaudited customer account and no controlled human-baseline productivity study.
Anthropic lists Fable 5 at $10 per million input tokens and $50 per million output tokens, compared with $5 and $25 for Opus 4.8. The source analysis says those prices are confirmed by Anthropic’s launch materials and platform pricing documentation. Sonnet 4.6 is listed at approximately $3 per million input tokens and $15 per million output tokens. Fable 5 is the most expensive model in Anthropic’s active lineup, according to the analysis.
Artificial Analysis gives Fable 5 an Intelligence Index score of 64.9, against 61.4 for Opus 4.8, a reported 5.7% difference. On GDPval-AA, a knowledge-work benchmark reported as an Elo rating, Fable 5 scores 1,932 to Opus 4.8’s 1,890, a 42-point gap. The analysis estimates a full Intelligence Index run costs about $9,940 for Fable 5, compared with about $4,970 for Opus 4.8.
The index is an aggregate result, not a direct measure of how much faster or more productive employees will be. The analysis says Fable 5 set records in five of 10 sub-benchmarks, suggesting performance varies by task. It also reports that a 7:2:1 mix of cache-hit, input and output tokens would produce a blended rate of $7.70 per million tokens. That estimate depends on the stated usage mix; it is not the model’s list price.
AI MODEL ECONOMICS · JULY 2026 ANALYSIS
AI’s Costly 5.7% Intelligence Gain, Explained
Claude Fable 5 costs twice as much per token as Opus 4.8 and leads on an aggregate benchmark. The public evidence described here does not establish a comparable gain in workplace productivity.
01 / PRICING
What the premium costs
Listed prices make the token-cost gap clear. Effective bills still depend on input/output mix, caching, batch rates, and inference settings.
Claude Fable 5
Anthropic listed ratesClaude Opus 4.8
Half the Fable 5 rateClaude Sonnet 4.6
Lower-cost listed optionUsage changes the bill. The analysis estimates $7.70 per million tokens for a 7:2:1 mix of cache-hit, input, and output tokens. That is a scenario estimate, not Fable 5’s list price. It also cites a 90% prompt-caching discount, batch rates of $5 input / $25 output, and a 1.1× US-only inference multiplier.
02 / BENCHMARKS
A measurable lead, with limits
Artificial Analysis reports a modest aggregate Index lead and a 42-point GDPval-AA difference. Neither result is a direct measure of employee productivity.
Artificial Analysis Intelligence Index
Aggregate score; the relative difference is reported as 5.7%.
Estimated full Index run: about $9,940 for Fable 5 vs. $4,970 for Opus 4.8.
GDPval-AA knowledge-work benchmark
Reported as an Elo rating. The analysis says percentage framing can mislead; it estimates roughly a 56% head-to-head win rate.
Fable 5 reportedly set records in five of ten sub-benchmarks; results vary by task.
03 / EVIDENCE CHAIN
From benchmark to business claim
The evidence supports different kinds of conclusions at each step. Keep measured scores separate from unverified customer claims and company-specific outcomes.
Published prices
Anthropic pricing materials document per-token rates and options.
Third-party scores
Artificial Analysis reports benchmark results and comparisons.
Customer account
Anthropic’s launch post attributes a migration story to Stripe.
Workplace impact
Company-specific productivity and return remain to be measured.
“The migration would otherwise have taken a whole team over two months by hand.”
04 / PRODUCTIVITY CLAIM
What the public account does—and doesn’t—show
The cited example describes Fable 5 completing a codebase-wide migration in a 50-million-line Ruby codebase in one day. It is a reported account, not a controlled productivity study.
A specific customer story
The launch post attributes the migration and time comparison to Stripe. It establishes that the claim was made.
Independent verification
The analysis describes the account as unaudited, with no disclosed methodology or independent replication.
Human-baseline productivity
The source identifies no controlled human comparison for Fable 5 and no basis for translating a 5.7% Index lead into work completed.
Fit for another organization
Results depend on task mix, review and correction, token use, quality needs, and deployment costs.
Compare models on representative tasks. Track quality, completion time, token use, review effort, and deployment costs. Define the human baseline before the trial begins.
05 / PROCUREMENT SNAPSHOT
Evidence strength at a glance
A clear price can support cost planning; benchmark results can guide evaluation. Productivity claims need a stronger, disclosed study design.
| Question | Public evidence described | What it supports |
|---|---|---|
| Token prices | ✓ Documented | Estimate costs using your usage mix and applicable options. |
| Benchmark results | ✓ Reported | Compare performance on the cited test suites. |
| Stripe migration story | ~ One account | Consider as a possible use case, not a general result. |
| Controlled human baseline | ✗ Not identified | No demonstrated productivity uplift for comparable teams. |
| All ten sub-benchmark details | ? Not provided | Task-level stability and relevance remain difficult to judge. |
06 / DECISION GUIDE
Test the work you actually do
The 5.7% figure describes an Intelligence Index comparison. It is not a forecast that employees will finish 5.7% more work, and the benchmark gap alone does not settle the premium’s value.
Select real tasks
Use representative work and quality requirements from your team.
Set a baseline
Define the human comparison and evaluation method in advance.
Count the full cost
Include tokens, review, correction, and deployment effort.
Decide by task
Pay more where measured quality or time gains justify the rate.
How much does Claude Fable 5 cost?
The cited list rates are $10 per million input tokens and $50 per million output tokens. Discounts and usage patterns can change effective cost.
What does the 5.7% gain measure?
The reported difference on Artificial Analysis’s Intelligence Index: 64.9 for Fable 5 versus 61.4 for Opus 4.8.
Is there a controlled productivity study?
The analysis identifies no controlled human-baseline study and describes one unaudited enterprise account.
For organizations choosing a model for routine workloads, the comparison puts a concrete trade-off in view: double the listed price versus Opus 4.8 for a modest lead on one aggregate benchmark. That may be difficult to justify when a task does not benefit from Fable 5’s stronger results. On the other hand, a high-stakes or unusually demanding workload could make a larger task-specific improvement worth paying for.
The figures do not settle that business decision. An index score is not a forecast of labor savings, and the 5.7% benchmark difference should not be translated into a claim that workers will finish 5.7% more work. Buyers need to test their own tasks, token usage and quality requirements before estimating costs or returns. The analysis’s central distinction is between measured benchmark performance and demonstrated productivity in an organization.
That distinction matters for budgets and procurement. A documented rate allows buyers to calculate token costs, subject to their usage patterns and available discounts. A customer story can suggest a possible use case, but without a disclosed method or independent verification it cannot establish that comparable teams will see the same result.
As an affiliate, we earn on qualifying purchases.
Benchmarks Versus Customer Evidence
The analysis attributes the benchmark figures to Artificial Analysis, an independent model evaluation and comparison service. It reports that the aggregate Intelligence Index favors Fable 5, while the GDPval-AA difference is 42 Elo points. The source also cautions that expressing this Elo gap as a percentage can be misleading: Elo is a rating scale, and the analysis says the gap corresponds to an estimated head-to-head win rate of roughly 56%. That interpretation does not change the listed benchmark scores or establish workplace productivity.
The productivity example comes from Anthropic’s launch post, which attributes a result to Stripe: in a 50-million-line Ruby codebase, Fable 5 completed a codebase-wide migration in a day that, according to the account, would otherwise have taken a team more than two months by hand. This establishes that the claim was made, not that the work was independently audited or that the comparison was conducted under a published study design. The analysis says other coverage repeated the account, but repetition is not independent verification.
The source says Fable 5 and Mythos 5 use the same underlying weights and are priced identically, with differences in safeguards. It also cites a 90% prompt-caching discount, batch pricing of $5 per million input tokens and $25 per million output tokens, and a 1.1-times multiplier for US-only inference. These options can affect a buyer’s effective bill; they do not alter the benchmark comparison itself.
“The migration would otherwise have taken a whole team over two months by hand.”
— Anthropic’s launch post, quoting a Stripe account
As an affiliate, we earn on qualifying purchases.
Limits of the Productivity Case
The analysis says it found no controlled human-baseline productivity study for Fable 5 and only one enterprise proof point: the Stripe account repeated in Anthropic’s launch post. The source describes that account as unaudited and without disclosed methodology or independent replication. It does not establish whether the reported task, time comparison or labor estimate has been externally checked.
The public information also does not show how Fable 5 performs on a particular company’s work, how often the benchmark-leading sub-tasks arise in that company, or whether any improvement offsets higher token costs. The source material does not provide a full description of Artificial Analysis’s benchmark methodology, confidence intervals or the specific results for all 10 sub-benchmarks. Those details would help readers judge how stable and relevant the comparisons are.
Several price comparisons also depend on use. The blended $7.70 figure assumes a specified mix of cached and uncached tokens, while the headline prices apply per million tokens of each type. The source calls a comparison with GPT-5.5 directional because it relies on a single secondary source. Neither the blended estimate nor that cross-provider comparison should be treated as a universal cost for every customer.
AI productivity benchmarking tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Evidence Buyers Should Seek
Businesses considering Fable 5 can compare it with lower-priced models on representative tasks, tracking quality, completion time and token use rather than relying on a general benchmark score. Any internal trial should define the human baseline in advance and account for review, correction and deployment costs. Those results would be more useful for a purchasing decision than assuming the Stripe example applies across different teams and codebases.
The next evidence that could clarify the productivity case would include a disclosed study design, defined work tasks, a human comparison group, measured outcomes and independent review. The source material does not say that Anthropic or Stripe has announced such a study. For now, Fable 5’s pricing and cited benchmark scores are available for comparison, while its broader enterprise productivity impact remains unestablished by the evidence described in the analysis.
As an affiliate, we earn on qualifying purchases.
Key Questions
How much does Claude Fable 5 cost?
Anthropic’s listed rates cited in the analysis are $10 per million input tokens and $50 per million output tokens. Discounts and usage patterns can change a customer’s effective cost.
What does the 5.7% gain measure?
It is the difference reported on Artificial Analysis’s Intelligence Index: Fable 5 scores 64.9 and Opus 4.8 scores 61.4. It is a benchmark comparison, not evidence that workplace productivity rises by 5.7%.
Is there an independently verified productivity study?
The source analysis reports no controlled human-baseline study for Fable 5. It identifies a Stripe customer account in Anthropic’s launch post, but says it is unaudited and lacks independent replication.
Does Fable 5 outperform Opus 4.8 on every task?
No such conclusion is established by the aggregate score. The analysis reports that Fable 5 set records in five of 10 sub-benchmarks, indicating that results differ by task. Buyers would need to test their own workloads.
Source: ThorstenMeyerAI.com
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
