Fable 5.1 Cache Pricing Rewrites Agent Procurement Math
Anthropic cut Fable 5.1 cache reads seventy five percent while holding base token prices flat, shifting agent procurement math toward cache hit share instead of headline rates.
What changed
Anthropic shipped Claude Fable 5.1 on 1 September 2026 at unchanged base input and output rates of ten and fifty dollars per million tokens, but cut prompt cache reads from one dollar to twenty five cents per million tokens, a seventy five percent reduction versus Claude Fable 5. Official documentation states Fable 5.1 and Mythos 5.1 use a 0.025x multiplier on cache hits rather than the 0.1x standard applied to other Claude models.
OpenAI's GPT 6 Astra, released two days later, lists cache reads at one dollar per million tokens on the same ten dollar input base. Headline parity hides a four to one gap on the line item that dominates long agentic threads re reading stable system prompts and tool schemas.
Why it matters
Frontier model comparisons still lead with base input and output tables. Agent fleets re read massive cached prefixes every turn. Anthropic's pricing move targets exactly that line, which enterprise buyers had flagged as the pain point in multi hour coding and research sessions.
If your workload cache hit share exceeds roughly half of input tokens, Fable 5.1's card can undercut Astra on invoice totals even when intelligence scores sit within single digits on third party indexes. The strategic signal is that labs now compete on cache economics, not just benchmark rows.
Who is affected
FinOps owners modeling agent run costs. Platform teams standardizing on one frontier vendor for coding agents. Procurement comparing Anthropic, OpenAI, and Google offers during the September 2026 release cluster. Vendors deciding whether to match Anthropic's 0.025x multiplier or hold cache pricing as margin.
What to do next
Pull thirty days of agent logs, compute cache read share, and rerun total cost under Fable 5.1, Astra, and Gemini 3.8 Flash cards using your actual hit rate rather than list prices.
What to watch
Whether OpenAI cuts Astra cache reads in response. Google Gemini 3.8 Flash introductory cache pricing through December 2026. Customer announcements citing cache line items in enterprise renewals during Q4 budgeting.
Sources
- Primary. Anthropic documentation, Claude Fable 5.1 overview (September 2026). Cache read rate, unchanged base pricing, release date.
- Secondary. TokenCost analysis, Claude Fable 5.1 Pricing: The 75% Cache Read Cut (September 2026). Comparative multipliers across Claude models and savings math.