Research
DeepSeek API pricing, and the same model on 35 other hosts
DeepSeek V4.1 Flash costs $0.15 per million input tokens and $0.60 per million output tokens on DeepSeek's own API at off-peak hours, and twice that at peak hours. On OpenRouter, 35 other providers sell the same model, some for a fraction of DeepSeek's rate and some for more, and in our tests a cheap host matched the fastest one while one 4-bit host fell behind the rest on quality.
That makes DeepSeek's rate one of 36 prices for the same model, and several hosts label their copy as 4-bit. We run this model for live caption translation in Verli's multilingual rooms, so we wanted to know whether a cheap host or a 4-bit host costs you quality or speed. We timed eleven hosts twice, 20 minutes apart, and compared four of them on quality. One host's speed changed almost fourfold between the two runs. DeepSeek's own prices come first.
DeepSeek's own prices
DeepSeek charges twice as much at peak hours as off-peak, and cached input costs one fiftieth of uncached input.
| DeepSeek API, per 1M tokens | Off-peak | Peak |
|---|---|---|
| V4.1 Flash, input (cache miss) | $0.15 | $0.30 |
| V4.1 Flash, input (cache hit) | $0.003 | $0.006 |
| V4.1 Flash, output | $0.60 | $1.20 |
| V4-Pro, input (cache miss) | $0.66 | $1.32 |
| V4-Pro, output | $1.98 | $3.96 |
We read these prices on DeepSeek's pricing page in November 2026. The API name deepseek-flash means version DeepSeek-V4.1-Flash-1017.
Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, except Chinese public holidays. Every other hour is off-peak, including weekends.
Peak hours this week in UTC
- Monday 01:00 to 04:00 and 06:00 to 10:00
- Tuesday 01:00 to 04:00 and 06:00 to 10:00
- Wednesday 01:00 to 04:00 and 06:00 to 10:00
- Thursday 01:00 to 04:00 and 06:00 to 10:00
- Friday 01:00 to 04:00 and 06:00 to 10:00
Saturday and Sunday are off-peak all day.
Peak hours from DeepSeek's pricing page, read in November 2026. Your time zone comes from your browser. Chinese public holidays are off-peak all day, and the grid does not mark them.
DeepSeek's pricing page also lists a context length of 1 million tokens, a maximum output of 384,000 tokens and JSON output.
The same model on other hosts
On OpenRouter, standard-priority input prices for the same model ran from $0.035 to $0.45 per million tokens in November 2026, a factor of 13. OpenRouter's list of endpoints for DeepSeek V4.1 Flash 1017 had 45 endpoints from 36 providers, including DeepSeek. Other companies can serve the model because DeepSeek publishes its weights under the MIT license.
A low-priority "flex" endpoint charged $0.025. Output prices ran from $0.15 to $1.80. Nine endpoints charged $0.30 and $1.20 at every hour of the day, the same prices DeepSeek charges at peak hours. The chart below places all 45 endpoints by their two prices.
$0.30 input and $1.20 output per million tokens
9 endpoints: baseten, fireworks (timed), mistral/fp8, near-ai, nebius/fp8, parasail/fp8, siliconflow/fp8, together (timed), venice/fp8
Hover over or tap a dot to see its endpoints.
Prices from OpenRouter's endpoint list for DeepSeek V4.1 Flash 1017, read in November 2026, including flex and priority endpoints. Endpoints with the same prices share one dot. Both axes use a log scale.
We tested eleven endpoints across that price range. Ten returned translations we could time. Each received the same request, to translate one unfinished sentence from a sermon into Spanish, Korean and Chinese, with the previous sentence as context. We sent each endpoint eight calls, then eight more 20 minutes later.
We also compared four of those endpoints on FLORES-200, a set of sentences with professional translations. Scores come from COMET, a model trained to predict human ratings. OpenInference and Sail Research ran in November 2026. The Wafer and Morph scores come from our earlier benchmark runs.
| Endpoint | Input | Output | Time, run 1 / 2 | COMET | Room cost vs Wafer |
|---|---|---|---|---|---|
 together | $0.30 | $1.20 | 0.63 / | not tested | 3.1x |
 wafer/ | $0.038 | $0.50 | 0.63 / | 0.896 | 1.0x |
 open-inference/ | $0.035 | $0.50 | 0.79 / | 0.879 | 1.0x |
 morph/ | $0.05 | $0.45 | 1.26 / | 0.896 | 1.0x |
 relace/ | $0.035 | $0.50 | 1.72 / | not tested | 1.0x |
 inference-net/ | $0.08 | $0.55 | 1.85 / | not tested | 1.2x |
 fireworks | $0.30 | $1.20 | 1.49 / | not tested | 3.1x |
 sail-research/ | $0.101 | $0.15 | 2.02 / | 0.896 | 0.6x |
 deepinfra/ | $0.098 | $0.294 | 1.89 / | not tested | 0.9x |
 streamlake/ | $0.15 | $0.60 | 2.14 / | not tested | 1.6x |
 deepseek, off-peak | $0.15 | $0.60 | not tested | not tested | 1.6x |
| deepseek, peak | $0.30 | $1.20 | not tested | not tested | 3.1x |
We took prices per million tokens from OpenRouter in November 2026. Times are medians of eight calls through OpenRouter, without streaming.
The times include OpenRouter's routing and the network, and each endpoint received the same request eight times in a row. The times compare one host with another. They do not show how long a caption takes to appear.
The room cost column projects what translation would cost for one hour of one of our rooms on each host, as a multiple of Wafer's cost. We took the input and output tokens our room used on Wafer in replays of seven recorded talks and priced them at each host's rates. We priced every input token as a cache miss because our records do not say how many were hits. Our actual Wafer bill for those replays came in a little under the projection.
On those projections, Sail Research would be the cheapest host for our room, at about 0.6 times Wafer's cost, because our captions produce a lot of output and Sail charges the lowest output price. For uncached input, Wafer, OpenInference and Relace charge about a quarter of DeepSeek's off-peak price. If your input repeats, compare cache prices too. DeepSeek's off-peak cache hit, at $0.003, is cheaper than the $0.005 those hosts charge.
Is a 4-bit host worse?
A 4-bit label alone did not predict quality in our test. On FLORES, Wafer and Sail Research, which OpenRouter labels as 4-bit, scored the same 0.896 COMET as Morph, which it labels fp8, while OpenInference, with an fp4 label, scored 0.879.
OpenRouter labels each endpoint's number format, and it labels four of the ten we timed as 4-bit. A smaller format can mean a host runs a compressed copy of the model, and compression can lower quality.
According to vLLM's serving recipe for DeepSeek V4.1 Flash, DeepSeek's own released weights already store the expert layers in MXFP4, a 4-bit format. Other parts use 8-bit. The embeddings, the output layer and the norms use higher precision. So a 4-bit label alone does not mean a host compressed the model. We did not check which weights any host actually serves.
We did not compute confidence intervals for those FLORES scores. On live speech, two AI reviewers, Claude Opus 5.5 and GPT-6.1 Sol, scored 220 passages from seven talks. Wafer scored 97.2, and Morph scored 97.5. The 95% range for the difference ran from -0.1 to +0.8. Both hosts made 20 serious errors per 100 passages. We call that a tie.
OpenInference's 0.879 is below every other DeepSeek host, and it ranks below eleven of the twelve LLM setups in our LLM translation test. It also rate-limited us. When we sent its 700 FLORES calls in a burst, OpenInference returned 311 with a rate-limit error. We sent those calls again slowly.
Our results do not show why OpenInference scored lower. They show that two endpoints with the same fp4 label and the same model scored differently. So test the host you plan to use, since the label alone does not tell you how well it translates.
Speed changes by the hour
A host's speed changed from one run to the next, and price did not predict it. Morph's median time rose from 1.26 seconds to 4.93 seconds between our two runs, 20 minutes apart. Two hosts failed once each. DeepInfra returned a rate-limit error, and Fireworks returned "no healthy upstream".
Wafer stayed at 0.63 seconds in both runs that evening. Earlier that day, a much smaller Wafer request took 3.3 to 3.6 seconds. The day before, our three-language test took 0.96 seconds on Wafer.
Together and Wafer tied in the first run. Together was the fastest host in the second run, at 0.56 seconds. Together costs about three times as much per room hour as Wafer.
Measured in November 2026 through OpenRouter. Each time is the median of eight calls without streaming. Room cost is the projected cost per room hour as a multiple of Wafer's.
For live workloads, test a host more than once. Keep a fallback even when today's results are fast. Verli's rooms send each call to Wafer first, then Morph, then a different model, Gemma 4 31B, each with an 8-second limit.
How to pick a host
Start with what your workload costs most. For long uncached inputs with short answers, compare input prices. The cheapest standard endpoints charge about a quarter of DeepSeek's off-peak rate. For long answers, compare output prices. Sail Research charges a quarter of DeepSeek's off-peak output rate. For anything a person waits on, time the host yourself with your own request, more than once, at the hours you will run it.
Then run a quality check on the host you pick, even a small one. Each host's 700-call FLORES run cost us under six cents. That run found the one host that translated worse.
Pin your endpoint so OpenRouter does not route the call elsewhere. Keep your own fallback list. OpenRouter adds a 5.5% fee when you buy credits on its Standard plan. OpenRouter pricing vs going direct covers that fee and when going direct pays off.