September 19, 2026
A promise of 99% cache signals DeepSeek reselling: nobody else offers that
With DeepSeek, a cache hit costs 50 times less than a miss, which is why providers put cache percentages in their marketing.
dax
@thdxr
by the way, another kind of scam: I see inference providers claiming “99% cache hit rates” the only provider in the world delivering that much right now is deepseek. so they’re just wrapping deepseek and then they’ll switch to another provider and claim the same thing
· 311.4K views
The “99% cache hit rate” figure is no longer boasting. It reveals whose model a provider is really running.
Only DeepSeek currently delivers that kind of rate. Dax Raad, creator of the OpenCode agent, draws this conclusion: a provider promising 99% is actually running DeepSeek under its own label, and does not change the promise on its pricing page after switching to another model.
Why cache percentages have become a commodity
Where such a gap comes from. With DeepSeek, a million tokens from cache cost $0.006 during peak hours versus $0.30 on a miss (DeepSeek pricing, 19.09.2026). The hit rate moves the bill more than the listed price.
Cache got more expensive along with the model. On 16.08.2026, DeepSeek raised prices fourfold and said it wanted to allocate resources more sensibly and shift load to off-peak hours (Fortune, 13.08.2026).
Raad has numbers of his own. On 07.08.2026, Raad wrote that more requests to DeepSeek had passed through OpenCode over two days than to anyone else, and from different customers, not just his agent.
The public OpenCode dashboard shows the same. For deepseek-v4.1-flash, 98% of input tokens come from cache while it ranks first in usage for the week, and only DeepSeek models stay above 97% on this list: everyone else is at 84–92% (window 26.07–19.09.2026).
A month earlier, price was the subject of deception. In August, Raad wrote that dozens of providers claimed to host DeepSeek at the old price, and not one of them did (16.08.2026).
How to calculate your cache rate
DeepSeek returns the `prompt_cache_hit_tokens` and `prompt_cache_miss_tokens` fields in its API response, so the real rate can be calculated from your own logs without taking the seller’s word for it. Cache is enabled by default and works on a best-effort basis; no one guarantees 100% hits.
On OpenRouter, the `cache_discount` field inside `prompt_tokens_details` shows the savings per request. DeepSeek cache reads are billed at 0.1x the input price, while writes are billed normally. A single request’s breakdown opens with the Detail button in the Activity tab.
Provider substitution is prevented by strict pinning. In an OpenRouter request, set `"provider": {"order": ["<slug>"], "allow_fallbacks": false}` or `"only": ["<slug>"]`, and failover will not move traffic to someone else’s pricing.
The metric survives only until a second vendor also learns to deliver 97% from cache.
Source
