Claude Fable 5.1 and the Permanent $2/$10 Price: Recalculating Production Agent Costs

Two announcements, one week apart

On August 15, 2026, Anthropic quietly updated its pricing page: the price increase to $3/$15 per million tokens that had been scheduled for September 1 would not happen. The $2/$10 per MTok rate for its flagship tier stays in place as the standard price, not a limited-time promotion.

Two weeks later, on September 1, Anthropic shipped Claude Fable 5.1 (the generally available flagship, tuned for agentic and tool-using workloads) alongside Claude Mythos 5.1, a restricted-access variant for vetted cybersecurity and life-sciences organizations. Fable 5.1 pairs with meaningfully cheaper cache reads than its predecessor.

Individually, either announcement is routine vendor news. Together, they change the cost model for anyone running long-lived agent sessions — which is most of webhani's client base at this point.

Why cache-read pricing matters more than the headline rate

Agentic coding tools re-send large chunks of context on every turn: system prompts, tool definitions, file contents, conversation history. Without prompt caching, you pay full input-token price for that context on every single call. With caching, only the first call pays full price; subsequent calls within the cache window pay a fraction for cache reads.

This is why the "sticker price" per MTok tells you less than it used to. A model with a higher list price but cheaper cache reads can be less expensive in practice for agentic workloads that reuse context heavily — which describes almost every Claude Code session, CI-integrated agent, or long-running background job.

A simplified way to model this for a real workload:

session_cost = (first_call_input_tokens * input_price)
             + (cached_reads * cache_read_price)
             + (output_tokens * output_price)

For a typical multi-turn coding session — say a 20-turn agent run where 90% of the input context is stable (system prompt, tool schemas, file tree) and only 10% changes turn to turn — the cache_read_price term dominates total spend far more than the input_price term. That's the number Fable 5.1's cheaper cache reads actually move.

A concrete comparison

Take a hypothetical but representative session: 20 turns, ~8K stable tokens cached per turn, ~800 new input tokens per turn, ~600 output tokens per turn.

Stable context reused per turn: 8,000 tokens (cache read)
New input per turn:              800 tokens (full price)
Output per turn:                 600 tokens (output price)
Turns:                           20

Before the cache-read price drop, the cached-context portion of a 20-turn session could easily represent 40-50% of total spend, even though it's the "discounted" line item — simply because it's read 20 times per session. Any reduction in cache-read price compounds across every turn, every session, every day the agent runs. This is the lever that actually moves your monthly bill for agentic workloads, not the headline input/output rate.

What we're telling clients

Stop comparing models by list price alone. For agentic workflows, ask vendors (or check your own usage logs) what fraction of your token spend is cache reads versus fresh input. If it's over 30%, cache-read pricing is your primary cost lever, not the number in the marketing table.
The permanent $2/$10 rate removes a planning risk, not a cost. Teams that had budgeted for a September price hike can redirect that buffer. But don't treat "no increase" as "cheaper" — it's a removed uncertainty, which is still useful for annual cost forecasting.
Re-run your cost model after any model swap. Migrating from Fable 5.0 to Fable 5.1 (or between any two model versions) changes the cache-read coefficient in the formula above. We've seen teams keep the same prompt structure and assume costs scale linearly with list price — they don't, once caching is in the mix.
Architect for cache hits, not just model choice. Ordering your prompt so static content (system instructions, tool definitions, file context) comes first and volatile content (the latest user turn) comes last maximizes what a caching-aware model can actually reuse. This is a prompt-engineering decision independent of which model you're on, and it has more leverage than most people assume.

Takeaway

The interesting story here isn't "prices didn't go up" — it's that the actual unit of cost for agentic AI has shifted from per-call input pricing to cache-read economics, and most cost dashboards still aren't built to show that breakdown clearly. If you're running agents in production and haven't separated cache-read spend from fresh-input spend in your monitoring, that's the gap worth closing before your next model migration.

References: Anthropic pricing page (accessed September 2026); Claude Code Changelog — September 2026; Anthropic Claude News — September 2026