Part of: this research — view all 8 rows
Anthropic's actual serving cost is unpublished, so this question will likely resolve on a proxy: the output-token list price of the cheapest model matching Claude Fable 5's August 2026 capability, perhaps adjusted toward estimated cost. Fable 5 launched at $50 per million output tokens. Within a month of the start date, Sonnet 5.5 launched at $10 and scored 55 on Artificial Analysis v5, compared to 49–50 for Fable 5.1 Anthropic Models: Intelligence, Performance & Price. Since Fable 5 is presumably no better than 5.1, a roughly 5× reduction in the price of this capability tier is already established.
Reaching my median decline of 72.5× requires the cost of Fable 5-level performance to fall to about $0.69 per million output tokens by December 2028. This is highly plausible via smaller model tiers. Haiku 5.5 currently lists at $0.50 per million output tokens Introducing Claude Haiku 5.5 and scores 36 on the v5 index Claude Haiku 5.5: Intelligence, Performance & Price Analysis Anthropic Models: Intelligence, Performance & Price. Given that earlier Haiku models gained about 26 index points in a year Anthropic has released Claude Haiku 5.5, one or two more Haiku generations should reliably reach Fable 5's benchmark level over the 28-month forecast window. Furthermore, Anthropic's gross margins rose from 12% to 85% in a single year Anthropic's Gross Margin Rocketed to 85% From 12% in a ..., strongly suggesting that true underlying compute and serving costs are compressing even faster than list prices.
Broader industry trends support massive fixed-capability cost declines, though mechanical extrapolation is risky. Epoch estimates that the cost to achieve a fixed capability level falls about 47% a quarter, or roughly 13× a year The plunging price of thought The plunging price of thought. Compounding this over 2.33 years would imply a 300× to 400× decline. I heavily discount this upper bound for the central estimate, as benchmark-based expenditure studies capture task-level reasoning token behavior and cross-provider pricing dynamics rather than pure output-token production costs The Price of Progress: Price Performance and the Future of AI. Additionally, cost declines tend to slow as a capability tier ages, and achieving parity on narrow benchmarks does not guarantee full holistic or agentic equivalence.
My central 80% interval runs from 9.5× to 480×, with the middle 50% spanning 26× to 185×. This wide range reflects differing resolution interpretations and the physical limits of hardware and model distillation. On the low end (around 9.5× to 26×), Fable 5's broad agentic capabilities might prove difficult to distill into a Haiku-class architecture, forcing reliance on a Sonnet-tier model. Compute scarcity or high-bandwidth memory constraints could also stall price cuts. On the high end (approaching 185× to 480×), the true cost of inference could plummet due to hardware performance-per-dollar compounding https://epoch.ai/data-insights/chip-performance-per-dollar, serving-stack optimizations like those seen on NVIDIA's GB300 NVIDIA Platform Delivers Lowest Token Cost Enabled by ..., and accelerating AI-assisted algorithmic research. Expected compute efficiency gains and underlying hardware trends support the 72.5× median, while physical constraints and uncertainty about how fully these gains translate into fixed-capability serving costs temper the outer bounds. If the resolver strictly estimates underlying production cost rather than API price, these compounding efficiencies will push the outcome deep into the right tail.
Assessing expected compute efficiency gains and underlying hardware trends led to a minor tightening of the outer bounds and a slight upward revision in the median expected cost decline.
Ask a followup
Sign in to run · $20 free credit, no card · every claim cited