[ runaii cloud — cache economics, measured ]
The cheapest measured effective input price in the market — on our own fleet, on real agentic traffic. Not a promo price: a hit rate. The method, the comparison table, and a recipe to run it on your traffic are all on this page.
| Provider | Cache hit | Effective input / 1M | Note |
|---|---|---|---|
| runaii (measured) | 95.8% | $0.023 | this page's numbers |
| Z.ai | 94.6% | $0.025 | market leader |
| Phala | 94.0% | $0.328 | — |
| Together | 91.0% | $0.362 | — |
| AkashML | 60.0% | $0.608 | — |
| Reka AI | 32.5% | $0.866 | 37× our effective input |
Cache hit rate ≈ market position. The leader (Z.ai) wins on hit rate, not list price — and a provider with a 32.5% hit rate pays 37× more per effective input token than we measured.
Warm chat turn
182-200 ms
multi-turn conversation, full prefix cached
TTFT, 1M context
0.28-0.93 s
warm resend of a ~1M-token context
788K-token warm resend
2.1 s
radix + host-tier KV, no re-prefill
697K-token cold ingest
11,945 tok/s
first touch — the miss path, priced honestly
Cold (miss) and warm (hit) paths are both priced and both published — a benchmark that only shows the warm path is an ad, not a measurement.
# 1) Cold call — the first touch is a MISS (nothing cached yet).
# Use a long, stable prefix (agents resend the whole history every turn).
curl -s https://api.runaii.cloud/v1/chat/completions \
-H "Authorization: Bearer $RUNAII_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "deepseek-v4-flash",
"messages": [{"role":"user","content":"<your ~100K-token agent history>\nSummarize."}],
"max_tokens": 32
}' | jq '.usage'
# -> prompt_tokens: N, prompt_tokens_details.cached_tokens: ~0
# 2) Warm call — SAME prefix, one more turn. This is the HIT.
curl -s https://api.runaii.cloud/v1/chat/completions \
-H "Authorization: Bearer $RUNAII_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "deepseek-v4-flash",
"messages": [{"role":"user","content":"<the SAME ~100K-token history>\nSummarize again."}],
"max_tokens": 32
}' | jq '.usage.prompt_tokens_details'
# -> cached_tokens: ≈ N (the prefix came from cache: no re-prefill)
# 3) Your effective input price (deepseek-v4-flash: $0.22 input / $0.022 cached per 1M):
echo "effective = (cached*0.022 + (prompt-cached)*0.22) / prompt $/M"The cached_tokens field is the engine’s own radix-cache accounting on the response — not a marketing number. Your hit rate depends on your prefix reuse: agentic multi-turn workloads with stable histories score highest; single-shot novel prompts score near zero. That’s the honest shape of the metric.
cached_tokens field you read in the recipe above.$5 of credit, no card. Point your existing OpenAI client at one URL and watch cached_tokens climb turn over turn.