Fable Cache Becomes Cheaper After 24 Reuses When Gemini Misses

Fable’s cache can become cheaper than Gemini 3.8 Flash for a repeated prompt, but only after many successful Fable hits and frequent Gemini misses. With equal billed prefix-token counts, 24 Fable reuses beat Gemini with no Gemini cache hits; the one-hour Fable option requires 39. This isolates repeated input, not the cost of a complete AI workload.

The comparison starts with two very different pricing structures. Through December 31, 2026, Fable 5.1 charges $12.50 per million tokens to create a five-minute cache, $20 for a one-hour cache and $0.25 for each cache hit. Gemini 3.8 Flash charges $0.75 per million standard input tokens and $0.075 for cached context.

Fable therefore begins at a steep disadvantage, then benefits from a low read price. If every Fable reuse hits while every Gemini reuse is billed as standard input, one five-minute Fable write plus 24 reads costs less than Gemini’s initial request and 24 reuses. The higher-priced one-hour write reaches that point at 39 reuses.

Prompt length does not change those thresholds once both services can cache the same prefix: the calculation is per million repeated-prefix tokens, so equal token counts cancel out. That simplification excludes output and any changing prompt suffix, which are separate billable components.

The zero-hit case is deliberately harsh on Gemini. Its documented cached-context rate is already lower than Fable’s $0.25 read price. The pivotal question is whether a request receives that rate. Gemini enables implicit caching by default on newer models and returns cache-hit tokens in response usage metadata, but Google does not guarantee a hit for similar requests.

At a Gemini cache-hit fraction of 74.074% or more of reuse tokens, Fable never becomes cheaper under the model’s single-successful-write assumption. Gemini’s expected marginal reuse price then reaches $0.25 per million tokens or less, matching or beating Fable’s read price while Gemini retains the cheaper initial input.

Below that boundary, there is no single crossover point. For a Gemini hit fraction h below 74.074%, the minimum Fable reuse count is floor((W-0.75)/(0.5-0.675h))+1. W is 12.50 for the five-minute Fable write and 20 for the one-hour version. This is a price equation, not a forecast of Gemini’s cache performance in a particular application.

Three constraints behind the crossover

• Eligibility matters first. Fable can cache repeated prefixes from 512 tokens, while Gemini 3.8 Flash implicit caching requires at least 4,096 tokens. Between those limits, Gemini cannot record an implicit-cache hit.
• Fable needs continuity. A five-minute cache hit refreshes the entry without another write, but a request after expiry requires another costly write. If every request arrives after expiry, Gemini is cheaper even with no Gemini cache hits.
• The prefix must remain stable. Fable caches the complete prompt prefix through its breakpoint, so a change at or before that point creates a different cache entry.

Fable’s default cache lifetime is five minutes, measured from the start of the request that writes or reads the entry. Generation time counts against that window. A long-running response can leave little time for the follow-up request needed to preserve the low-read sequence. The one-hour setting extends the window, but its $20 write charge explains the later 39-reuse crossover.

Gemini’s documented 24-hour implicit-cache retention is an upper retention limit, not a hit commitment. Google advises sending similar prefixes within a short interval. That makes Gemini’s economics potentially better for distributed reuse, but the rate card alone cannot convert the longer retention period into guaranteed savings.

The analysis cannot identify the cheaper application overall. Fable output costs $50 per million tokens, versus $3.75 for Gemini 3.8 Flash. A workload that generates substantial text can erase any savings from its repeated prompt prefix, even if Fable wins the narrow cache calculation.

The practical test is a token and timing audit. Anthropic’s Usage and Cost API separates uncached input, cached input, cache creation and output tokens at the organization level. Gemini reports cached tokens in response metadata. Those measurements, plus the size of changing suffixes and the interval between requests, determine whether either provider’s cache economics hold in production.