Sference served 92.6% of input tokens from cache in July 2026, the highest among major providers. The median major provider served 72.3%.

Key findings

1. 01Sference leads at 92.6%; Mistral is lowest at 13.7%.
2. 024 of 16 major providers served at least 80% of input tokens from cache.
3. 03xAI moved the most since June 2026: cache hit rate went from 36.9% to 74.2% (+37.3 pp).

Cache hit rate by month

What changed since June 2026

• xAI36.9% → 74.2% (+37.3 pp)
• Mistral32.4% → 13.7% (-18.7 pp)
• DeepInfra38.3% → 55.2% (+16.8 pp)
• OpenAI45.5% → 29.7% (-15.8 pp)
• Novita23.1% → 38.7% (+15.6 pp)

What to do about it

• Keep the system prompt and tool definitions at the start of the request and unchanged between turns.
• On providers that need explicit cache markers, set them. Without them the hit rate stays near zero.

Data

16 rows published. Rows below the floors are withheld.

Questions this answers

Which LLM provider has the highest prompt cache hit rate?

Sference served 92.6% of input tokens from cache in July 2026, the highest among major providers. The median major provider served 72.3%.

What is a typical prompt caching hit rate in production?

Sference served 92.6% of input tokens from cache in July 2026, the highest among major providers. The median major provider served 72.3%.

How much of my input is served from cache on Anthropic, OpenAI or Bedrock?

Sference served 92.6% of input tokens from cache in July 2026, the highest among major providers. The median major provider served 72.3%.

Method and definitions

• Cache hit rate is cached input tokens divided by input tokens, the same definition /rankings uses.
• Providers that report no cached tokens are left out of the chart. Only providers with at least 0.5% of published traffic are shown.
• A row is published only when it has at least 20 requests, at least 5 distinct organizations, and no single organization above 80% of its traffic. These are the same floors /rankings uses. Rows below the floors are withheld, not estimated.
• Only shares, rates, medians and organization bands are published. Absolute request, token and spend totals stay private.

Cache hit rate:

Cached input tokens / input tokens, summed over the month.

Per-model numbers with the same definitions are on /rankings, refreshed weekly.

Caveats

• Hit rate depends on workload. Long agent sessions with a stable prefix cache far better than one-off requests.
• Providers report cached tokens differently. Some only count explicit cache reads.

Revision history and other months

• October 2026first published later
• September 2026first published later
• August 2026first published later
• July 2026 (this page)first published 2026-11-11
• June 2026first published 2026-11-11
• May 2026first published 2026-11-11

Related

• Provider error ratesHow often do LLM providers return errors, and what kind?
• Time to first token by providerWhich LLM provider returns the first token fastest?
• Prompt-cache hit rate per provider, April 2026April 2026 note

Cite and reuse

Free to reuse under CC BY 4.0. Credit "Requesty Data" and link to this page. Cite the month URL so the numbers you quote stay fixed.

Requesty (2026). Prompt cache hit rate by provider, July 2026. Requesty Data. https://www.requesty.ai/data/provider-cache-hit-rate/2026-07

@misc{requesty_provider_cache_hit_rate_2026_07,
  author       = {{Requesty}},
  title        = {Prompt cache hit rate by provider, July 2026},
  year         = {2026},
  howpublished = {\url{https://www.requesty.ai/data/provider-cache-hit-rate/2026-07}},
  note         = {Requesty Data}
}

period,provider,cache_hit_rate_fraction,request_share_fraction,sample_band
2026-07,Sference,0.9262,0.017,100-499
2026-07,DeepSeek,0.8845,0.093,500+
2026-07,Fireworks,0.8663,0.0357,500+
…