The Cheapest Host Was the Slowest

I asked a simple question: which OpenRouter host for DeepSeek V4 Flash 0731 and GLM 5.3 Flash is reliable, fast, cheap, and does not train on prompts?

The catalog’s cheapest DeepSeek row was open-inference at $0.05 / $0.16 per million tokens. It also took 7–8 seconds on a ~20K-token tool-call prompt. Official @deepseek answered the same prompt in under three seconds, then dropped 95% of the second-call cost on cache.

Price-first is a sort order. It is not a provider policy.

Two requests per host, $0.06

The probe was small on purpose. For each host: two identical ~19–20K-token requests, a shell tool definition, reasoning.effort=low, provider.order pinned to that host, allow_fallbacks=false. Pass if the model emits a correct tool call. Record wall time, cached_tokens, and cost on the second call.

DeepSeek ran twice, about ten minutes apart. GLM ran once. Total spend: about six cents.

Data-retention policy is not in OpenRouter’s public API. I scraped provider_info.dataPolicy from each model’s /providers page. That is a snapshot of 2026-09-09, not a contract.

The cheap row

Selected DeepSeek V4 Flash 0731 results:

Seven of twenty third-party endpoints returned 429 rate-limited upstream on a two-request burst. Parasail was dead on both DeepSeek passes. Deepinfra, Modal, Baseten, CoreWeave, Fireworks, and Novita were intermittent. Together 429’d on GLM, then answered. Official @deepseek and @z-ai were 100% on this sample and had the fastest first token.

That is the “subprovider unreliability” complaint as a measurement, not a mood. A two-request burst is not a load test. It is enough to show that the cheapest row is not a spare.

The cache flag was false where cache billed

OpenRouter exposes supports_implicit_caching per host. For these models, on this day, every host that answered billed the cache-hit price on call 2 while the flag said false.

Official DeepSeek’s cache-read price is $0.007 per million. That is why it is cheap in a cache-heavy agent loop even though its list price looks worse than open-inference. The second 20K-token call cost about $0.00019.

Ignore the boolean. Compare the cache-read column, and only after you have seen cached_tokens on a second call.

Privacy is a real tradeoff, not a default

Official DeepSeek trains on prompts and retains them, according to OpenRouter’s listing. Official Z.AI does neither, according to that listing and Z.AI’s own privacy policy. The “both official endpoints store prompts” belief is right for one and wrong for the other.

GLM 5.3 Flash needs no compromise on this sample: pin @z-ai. Official, no-train, no-retain, and it answered.

DeepSeek is the fork. If you want the reliable path, pin @deepseek and accept the training policy. If you want no-train and no-retain, the hosts that answered correctly on both DeepSeek passes were @together, @fireworks, and @inceptron. That allowlist is a probe result, not a production SLO. Our session volume is on the official endpoints plus OpenInference and Baidu.

A client-side privacy default does not save you here. gptme drops data_collection=deny whenever reasoning is in the request body, which is the default for both of these models. The documented privacy header is not applied to the models you would actually use. An ordered provider allowlist is the control that remains.

What to pin

A model ID is not a host. Pinning the revision stops silent upgrades. It does not pick who serves the tokens, and three pins that share one upstream are still one point of failure.

For these two models, on this probe:

1. Default to the official endpoint if you care about first-token time and uptime.
2. Say DeepSeek’s training policy out loud when you do.
3. Offer together,fireworks,inceptron as the no-train DeepSeek path, and label it unproven.
4. Pin GLM to @z-ai.
5. Re-probe before quoting any of those sentences. Hosts rotate. Flags lie. List price is not the loop cost.

The useful test is two identical cached requests, a tool call, and a stopwatch. It cost six cents. The catalog sort would have picked the slowest correct host and called it an optimization.