DeepSeek V4 Pro: DeepSeek's Own API vs Together, Tested
By Promptster Team · 2026-11-09
Should you call DeepSeek V4 Pro through DeepSeek's own API or through a reseller like Together? On price lists, the answer looks obvious. DeepSeek charges $1.32 per million input tokens and $3.96 per million output at peak. Together charges $2.10 and $4.40.
On our 20 graded calls, the reseller was cheaper and faster: $0.0391 on Together against $0.0498 on DeepSeek's API, with a median call of 2.1 s against 4.1 s. Both passed 20/20. The higher rate card produced the lower bill because the same prompts produced about a third fewer output tokens on Together. One caveat applies throughout: DeepSeek's off-peak pricing is not modelled in these figures.
How we tested
Four script-graded tasks with traps, and no LLM judge:
- json-order: strict JSON from an order email that includes a decoy PO, a cancelled line and a conditional discount.
- stock-floor: a ten-step inventory calculation that rounds down at each step (answer: 587).
- seven-word-taglines: three lowercase lines of exactly seven words.
- launch-notes: pick the final approved date and budget out of superseded notes, replying with only
Nov 21, $37500.
We ran each task five times per configuration (n=5) with temperature 0.7 and max_tokens 4000, on 2026-10-23, through Promptster's own provider handlers. We called deepseek-v4-pro on DeepSeek's API and deepseek-ai/DeepSeek-V4-Pro-0813 on Together. Costs are cost_usd at list rates from shared/pricing.ts. DeepSeek's rows there are peak-hour rates, and the file notes that off-peak is half. We did not model off-peak pricing, and we did not check which window these calls fell in. Latency is wall-clock time from one machine.
The results
| DeepSeek API | Together | |
|---|---|---|
| Model id | deepseek-v4-pro |
deepseek-ai/DeepSeek-V4-Pro-0813 |
| List price (in/out per 1M) | $1.32 / $3.96 (peak) | $2.10 / $4.40 |
| Passed | 20/20 | 20/20 |
| Cost, all 20 calls | $0.0498 | $0.0391 |
| Median latency, all 20 calls | 4.1 s | 2.1 s |
| Median: json-order / stock-floor / taglines / launch-notes | 3.6 / 5.7 / 5.3 / 3.1 s | 2.0 / 3.6 / 2.2 / 1.7 s |
| Fastest / slowest single call | 2.7 s / 8.7 s | 1.3 s / 5.9 s |
| Input tokens, all 20 calls | 3,930 | 2,350 |
| Output tokens, all 20 calls | 11,260 | 7,758 |
| of which reasoning | 10,181 | 6,565 |
Neither configuration missed a call, timed out or hit a rate limit. Every response was correct, and every diagnostic passed: no code fences, a clean Answer: line, and no explanation after "reply with only". On accuracy, this is a tie.
Why the cheaper rate card cost more
We checked the per-run token counts before explaining the difference. Two things stand out.
Input: exactly 79 more prompt tokens per call on DeepSeek's API, on every task. The prompts were identical. DeepSeek counted 224, 182, 151 and 229 prompt tokens on the four tasks, and Together counted 145, 103, 72 and 150. A fixed offset like that most likely comes from each host's chat template, not from the text we sent. At these rates it adds little to the bill.
Output: DeepSeek's API wrote 45% more. Almost all of the gap was reasoning: 10,181 reasoning tokens against 6,565. The biggest difference was on the tagline task, where DeepSeek's API spent 3,108 output tokens across five runs and Together spent 1,473. One DeepSeek run alone used 1,041 output tokens, 1,014 of them reasoning, to write three seven-word lines.
| Task | DeepSeek API: output tokens (5 runs) | Together: output tokens (5 runs) | DeepSeek API cost | Together cost |
|---|---|---|---|---|
| json-order | 2,083 | 1,545 | $0.0097 | $0.0083 |
| stock-floor | 4,297 | 3,245 | $0.0182 | $0.0154 |
| seven-word-taglines | 3,108 | 1,473 | $0.0133 | $0.0072 |
| launch-notes | 1,772 | 1,495 | $0.0085 | $0.0082 |
Output is most of the bill on both hosts. Together's output rate is 11% higher, but it produced about a third fewer output tokens, which more than covers the difference. Together came out cheaper on every task, by anything from a few hundredths of a cent on launch-notes to almost half on the taglines.
Is it the same model?
We can't confirm it. Together's id carries a date suffix (-0813), and DeepSeek's deepseek-v4-pro alias doesn't say which snapshot it serves. The responses were very similar. Both hosts wrote similar day-by-day working on stock-floor and ended with Answer: 587. On taglines, DeepSeek's API wrote "refill often and carry your favorite drink" and Together wrote "refill often and carry your own style". The reasoning length was consistently different, though. That could come from a different snapshot, different serving defaults, or both. Our data shows the effect but not its cause.
What the caveats change
Off-peak. If your traffic falls in DeepSeek's off-peak window, both of its rates halve. On these token counts, that would very likely bring the first-party API below Together's bill. We did not measure or model this, so treat it as something to check against your own traffic pattern, not as a result.
Cache. DeepSeek's API reported 128 cached prompt tokens on 16 of the 20 calls. Together reported cached tokens on 3. cost_usd prices cached input at the full rate for both, because shared/pricing.ts has no cached rate for either. Input is a small share of the cost here, so a cache discount would not change the order.
Latency. Together's median was faster on every task. Its slowest call, 5.9 s, came in under DeepSeek's slowest, 8.7 s. Both figures come from one machine on one afternoon.
How this fits with earlier posts
In June we benchmarked DeepSeek V4 Pro on its official API against Qwen3 235B on Together (that comparison). In May we tested whether V4 was frontier-class at all. This is the first time we have run the same DeepSeek model on both hosts. It is also another example of a pattern from our host comparisons: what you pay depends on tokens multiplied by rate, and the token count can differ between hosts even for the same model family.
What we'd do
Don't pick a host from the rate card. Here the host with the higher rate cost 21% less, because the bill is driven by output tokens. Run your own prompts on both and compare usage.completion_tokens, not the listed prices.
If you run on a schedule, check off-peak first. A nightly batch that falls inside DeepSeek's off-peak window may well flip this result. An interactive workload during peak hours probably won't.
For latency-sensitive calls, Together was the better host in this run, with a faster median on all four tasks and a lower worst case.
For the broader open-weight vs closed comparison, see the original open-source vs closed-source benchmark.
Run this comparison yourself — no account, no API key
- Run the seven-word tagline task: this is the task where the two hosts' output lengths differed most.
To run DeepSeek V4 Pro on both hosts side by side and compare token counts and cost on your own prompts, use Promptster.
Run 2026-10-23 through Promptster's provider handlers: deepseek-v4-pro on DeepSeek's API and deepseek-ai/DeepSeek-V4-Pro-0813 on Together. Four script-graded tasks, five runs each, temperature 0.7, max_tokens 4000. Costs are cost_usd from list rates in shared/pricing.ts: DeepSeek at its peak-hour rate (off-peak pricing is not modelled), with no cache discount on either host. Latency is wall-clock from one machine. Raw responses and usage objects for every call are in content/blog/data/q4b-2026-battery/.