The discount is a cache hit

01Tuesday's USE_IT, Thursday's bill

Tuesday's issue led with Claude Fable 5.1 as the default Fable model in Claude Code, on the API, and in GitHub Copilot.¹² The numbers that sold the ship were plain: 1M context, $10/$50 per million tokens, cache reads $0.25 per million.³⁴ Anthropic's own product page estimated that cut would reduce typical workload cost by about 25 percent and highly agentic workloads by up to about 45 percent.⁵ The desk printed USE_IT.

Wednesday Codex 0.153.0 and Claude Code 2.1.259 took the lede — managed MCP, headless deny, multi-session race fixes — while Fable heat was still "yesterday's story, measured as burn and retention."⁶ Thursday the lede flipped again: Claude Code 2.1.260 fixed Fable 5.1 cache and effort mid-session.⁷ The release notes say the quiet part: prompt caching on Fable 5.1 had not covered the context attached after tool results, so that context was re-sent as uncached input on every tool-call turn; /effort on Fable 5.1 no longer invalidates the prompt cache mid-session; /cost and the status line now name a likely cause when the cache missed.⁸

Three weekday issues. One model. The question the daily format could not hold is what happens when the advertised discount and the client that ships the model disagree for 54 hours.

02What $0.25 actually buys

Base input and output did not move. Claude Fable 5.1 is $10 per million input tokens and $50 per million output tokens — the same sticker as Claude Fable 5.⁴⁵ The only price that changed is the cache-read line: $0.25 per million tokens, which is 0.025× base input, against the 0.1× multiplier every other Claude model still uses.⁴ Five-minute cache writes stay $12.50; one-hour writes stay $20.⁴

That is a conditional discount. Anthropic's estimate of roughly 25 percent on typical work and up to about 45 percent on highly agentic work assumes cache hits on a frozen prefix that the agent re-reads every turn.⁵ The pricing page is blunt about the arithmetic: a cache hit on Fable 5.1 costs 2.5 percent of the standard input price; on every other model it costs 10 percent.⁴ Miss the cache and you pay $10 again. There is no third rate for "almost cached."

The API surface that makes the hit reachable shipped the same day. Per-message effort changes are in beta on Fable 5.1: a mid-conversation role: "system" message carrying output_config.effort is supposed to change effort for later turns while preserving the prompt cache, under the mid-conversation-output-config-2026-07-01 header.³ Turn-scoped system messages with clear_at: "next_user_message" are the other beta path that keeps per-turn reminders from invalidating the cache.³ The platform notes and the Claude Code client are two different layers. The notes describe what the wire can do. The client decides what the wire actually sees.

Miss the cache and you pay $10 again. There is no third rate for almost cached.

03The 54-hour gap on latest

Claude Code v2.1.257 published 1 Sep 2026 at 17:53 UTC. The first bullet: Claude Fable 5.1 (claude-fable-5-1) is the default Fable model — 1M context, $10/$50, $0.25 cache reads.⁹ v2.1.258 the same evening fixed macOS 12 launch and remote session failures after a re-sent permission approval; it did not touch caching.¹⁰ v2.1.259 on 2 Sep added managedMcpServers and --permission-prompts none, fixed concurrent sessions clobbering ~/.claude.json, and expanded Bash Read deny coverage — still no Fable cache line.¹¹

v2.1.260 published 3 Sep 2026 at 23:48 UTC.⁸ Among the fixes: prompt caching on Claude Fable 5.1 had not covered the context attached after tool results, so it was re-sent as uncached input on every tool-call turn; /effort on Fable 5.1 no longer invalidates the prompt cache mid-session; model: fable agents honour [1m] on ANTHROPIC_DEFAULT_FABLE_MODEL instead of silently running a 200K window; the /model picker finally shows Fable 5.1 for orgs that have it.⁸ The same cut adds a likely cause for prompt-cache misses to /cost and the status line's prompt_cache field.⁸

From 2.1.257's publish to 2.1.260's publish is about 54 hours. Anyone who took the Tuesday USE_IT pin — Claude Code ≥2.1.257, try Fable 5.1 for long agent runs — spent those hours on a default that billed tool-turn context at full input.¹⁸ We did not meter a production fleet across that window. The release notes name the failure mode; they do not publish a token delta. The claim is the mechanism, not a dollar figure for your shop.

04What uncached tool context costs on paper

Take the shape Anthropic is selling: a long agent that re-reads a large frozen prefix on every tool turn.⁵ Let the cached prefix be 150,000 tokens and the session run 40 tool-call turns. Cache reads alone are 6.0 million tokens. At $0.25 that line is $1.50. At the full $10 input rate the same tokens are $60.00. The gap on that one line is $58.50 before uncached input, output, or cache writes enter the bill.

That arithmetic is a worked example from the published rates, not a field measurement of any team's 1–3 Sep traffic.⁴ Real sessions mix uncached user turns, growing history, cache writes at $12.50, and output at $50. Anthropic's "up to approximately 45 percent" already assumes hits dominate.⁵ Strip the hits on the tool-result path and the agentic end of that range does not apply to the turns that define the product.

Claude Code is the path most of this desk's readers use for Fable. The API caller who already structured prompts for automatic caching and never hit the broken client path is outside this claim. The reader who ran /effort mid-session on 2.1.257–259 to save money may have paid to invalidate the prefix the API beta was designed to preserve.³⁸

At $0.25, 6M cache-read tokens are $1.50. At full input they are $60.00.

05Stable never saw the model

npm dist-tags through the week kept stable on 2.1.236 while latest climbed 2.1.252 → 2.1.258 → 2.1.259 → 2.1.260.^12131415 As of this Saturday morning the same registry still shows stable at 2.1.236; latest has moved on past 2.1.260.¹⁶ Fable 5.1 as default Fable, the cache break, and the cache fix all live on latest only.

Last Saturday's long read already argued that stable is a delayed path, not a safer one: the money fixes of that week shipped on latest days before stable moved.¹⁷ This week is the sequel with a different price tag. The careful channel did not absorb a broken cache implementation. It never received the model. Homebrew claude-code, the stable apt/dnf/apk repos, and autoUpdatesChannel: "stable" without a minimumVersion floor stayed on a build that cannot select Fable 5.1 as the default Fable model because that default did not exist on 2.1.236.¹⁸⁹

So the exposure split is inverted from last week. Latest took the new model and the broken bill path. Stable took neither. Teams that chase USE_IT pins on day one are the ones who met the 54-hour gap. Teams that wait for stable are still waiting for Fable 5.1 on the slow channel entirely.

06The rest of the surface was not free either

GitHub Copilot made Fable 5.1 GA for Pro+/Max/Business/Enterprise on the same calendar day, with the policy off by default for Business/Enterprise admins.² Unlike other Claude models in Copilot, Fable 5.1 retains prompts and outputs for safety classifiers by default; eligible enterprises can get zero-data-retention routing through year-end while Enterprise Frontier Safeguards rolls out.²⁵ Enabling the policy is an acknowledgment of retention, not only a model picker flip.²

Cline 4.1.17 moved 57 provider defaults the same window, most consequentially Anthropic resolving to Fable 5.1 instead of Opus 5, and fixed a Hub memory balloon that could hit tens of gigabytes on long tasks.¹⁹ An unpinned Anthropic default in Cline became a Fable default whether or not the operator had read the cache economics.

None of that is the Claude Code client bug. It is the surrounding pressure: the model arrives as a default in more than one harness on day one, the discount is explained as a cache-hit story, and the harness that most coding agents actually run had not finished the hit path.

07What Tuesday's verdict should have said

Issue 022 printed: pin Claude Code ≥2.1.258; try Fable 5.1 for long agent runs; on Copilot, read the retention policy before enabling org-wide.¹ The retention half held. The long-agent half assumed the client delivered the rate card.

A tighter pin would have been: try Fable 5.1 on the API or on a harness whose cache behaviour you can meter; on Claude Code, wait for a cut that names tool-result caching on Fable, then pin. That cut was 2.1.260, two mornings later.⁸⁷ The desk did not have those release notes on Tuesday. It had the model card and the Copilot changelog. The missing half was the client.

The correction is operational, not theatrical. USE_IT on a model whose economics are cache-hit contingent is incomplete until the path that builds the prompt is known to hit. Rate cards are not harnesses. Harnesses are not rate cards. This week spent 54 hours proving the difference on the default Fable pin.