Cache Discounts: Which Provider Saves You the Most?

The read discount flattened, then started moving again

For a year, "who discounts cached reads hardest" was a real question. By mid-2026 it had stopped being one: Anthropic, OpenAI and Google all converged on 0.10x list input, and this post used to end there.

It has reopened, on one family. Here is where the four stand today:

• -Anthropic (varies by model): 90-98% off cached reads - 90% on most models, but 98% on Claude Fable 5.1
• -OpenAI (all models): 90% off cached reads
• -Google (all models): 90% off cached reads
• -xAI (varies by model): 75-85% off cached reads

So the brand-level answer is now wrong in both directions. xAI is shallower and inconsistent between models - Grok 4.5 and Grok 4.6 carry the same $2.00 list input, but a cached token costs $0.30/MTok against $0.50/MTok. And Anthropic is now DEEPER on one model than anyone else is anywhere: a cached token on Claude Fable 5.1 costs $0.25/MTok against $1.00/MTok on Claude Fable 5, which lists at the same input price.

The lever is not flat, and it is not a brand-level property. Check the row.

Where providers still differ: the write fee

Writing a prefix into cache is the part nobody quotes, and it is now the only structural difference in cache pricing:

• -Anthropic: 1.25x input on every model (25% surcharge to write) - the write fee did NOT move when the read discount did
• -OpenAI: 1.25x on the GPT 5.6 family; GPT 5.5 and GPT 5.4 charge nothing to write
• -Google: no write fee
• -xAI: no write fee

This matters exactly in proportion to how often your prefix changes. A stable system prompt is written once and read forever, and the fee rounds to nothing. A prefix that churns - retrieved documents, a rolling conversation window, per-user context - pays it on every miss.

What that looks like on a real workload

Take Gemini 3.1 Pro against Claude Sonnet 5. They now list at the same $2.00/MTok input and discount reads identically, so a cached token costs $0.20/MTok on both. There is no crossover left on the read side at any hit rate.

The write fee is the whole difference. At an 80% hit rate, with the other 20% being fresh writes:

• -Gemini 3.1 Pro: $0.56/MTok of prompt
• -Claude Sonnet 5: $0.66/MTok of prompt

Claude Sonnet 5 is 18% dearer on the prompt side - and 17% cheaper on output, at $10.00/MTok against Gemini 3.1 Pro's $12.00/MTok.

So the comparison is not about caching at all. It is about your output ratio. Once output exceeds roughly 5% of your prompt volume, Claude Sonnet 5 is cheaper overall - and almost every real workload clears that bar comfortably.

Flagship tier: same read price, different write bill

Claude Opus 4.8 and GPT 5.5 both list at $5.00/MTok input, so both read from cache at $0.50/MTok. Writes split them: $6.25/MTok on Claude Opus 4.8 versus $5.00/MTok on GPT 5.5, which has no write fee at all.

At the same 80% hit rate that is $1.65/MTok of prompt on Claude Opus 4.8 against $1.40/MTok on GPT 5.5. Output pushes back the other way: Claude Opus 4.8 bills $25.00/MTok against GPT 5.5's $30.00/MTok, 17% in its favor.

Worth noting inside OpenAI's own range: GPT 5.6 Sol lists at the same $4.00 / $20.00 as GPT 5.5 but adds the 1.25x write fee. For a cache-heavy workload the older model is the cheaper one.

Four rules for cache optimization

1. Do not shop providers on read discount. Three of four are identical. The decision is list price and output price. 2. Check the write fee if your prefix churns. It is the only cache figure that still varies, and it lands entirely on your misses. 3. On Grok, read the model row. xAI is the one family where the discount moves between models, and the newer model is not the cheaper one to cache. 4. Measure your hit rate before any of this matters. Below 50% hits you are optimizing the wrong number - fix prompt ordering first.

One caveat on xAI: our weekly benchmark source does not cover Grok yet, so we track its prices but publish no confidence scores and make no quality recommendation on it.

Model your cache savings at /teardown, or start tracking cache hit rates with a free dashboard.