Vercel's New DeepSeek Flash Ultrafast Costs Four Times the Fast Version. Vercel's Own Table Lists Fast at 966 Tokens a Second and Ultrafast at 950.
Vercel added DeepSeek V4.1 Flash 1017 Ultrafast to AI Gateway on November 7, served by Fireworks' accelerated endpoint, at $2.40 per million input tokens and $9.60 output, for what the changelog calls upwards of 1,000 tokens per second. It is the same October 17 build already on Gateway three other ways. Vercel's own DeepSeek lab table lists the existing Fast endpoint at $0.60 and $2.40 with 966 tokens per second, and Ultrafast at 950, so on Vercel's numbers the new tier is four times the price with no listed speed advantage, and eight times DeepSeek's peak list price against the standard row's 357. Third-party trackers measured anywhere from about 791 to about 1,741 tokens per second depending on host. We read the changelog, the model page, and the lab table.
By T.W. Ghost
Vercel added a fourth way to call the same model today, November 7, 2026. DeepSeek V4.1 Flash 1017 Ultrafast is on AI Gateway as deepseek/deepseek-v4.1-flash-1017-ultrafast, served by Fireworks' accelerated serverless endpoint, in public preview with no waitroom or session limit, at $2.40 per million input tokens and $9.60 output, cache reads at $0.05. The changelog promises "upwards of 1,000 tokens per second." The weights are the October 17 build we covered when DeepSeek swapped it in, and Gateway already serves it as a standard row and as a Fast endpoint.
The comparison that matters is on Vercel's own DeepSeek lab page. The existing deepseek-v4.1-flash-1017-fast is listed at $0.60 and $2.40 per million with 966 tokens per second and 0.5 seconds of latency. Ultrafast is listed at 950 tokens per second and the same 0.5 seconds. Four times the price, and the slower of the two listed speeds. Against the standard row, $0.30 and $1.20 at DeepSeek's peak list price with 357 tokens per second across more than twenty providers, Ultrafast is eight times the price for about 2.7 times the listed throughput. Outside Vercel's table the speed depends on who measures: one tracker saw a median of about 791 tokens per second over seven days on its own routing, another host reports about 1,741 for its own endpoint. We read the changelog, the model page, and the lab table. Here is what held up.
One Weight File, Three Prices
| Gateway row (Vercel's lab table) | Price per 1M tokens | Listed speed | Latency |
|---|---|---|---|
| Standard 1017 | $0.30 in / $1.20 out at peak list; shown 51% off; 21+ providers | 357 tokens per second | |
| Fast | $0.60 / $2.40 | 966 tokens per second | 0.5 s |
| Ultrafast (new) | $2.40 / $9.60; cache read $0.05 | 950 tokens per second | 0.5 s |
| Changelog claim for Ultrafast | "Upwards of 1,000 tokens per second" | ||
| Third-party reads | About 791 median over 7 days (CheaperInference, 34 samples, TTFT p50 1.8 s); about 1,741 on a different host's own endpoint | ||
| Gateway fee | None; no markup on inference (Vercel) |
Vercel's own table does not justify Vercel's own price ladder. If Fast is 966 and Ultrafast is 950, the extra $1.80 per million input and $7.20 per million output buys a Fireworks endpoint and a word. The table may lag the endpoint; until it is updated, it is what the vendor printed.
"Upwards of 1,000" is a ceiling, and the measurements scatter. Seven hundred ninety-one from one tracker, seventeen hundred from a host measuring itself. The number you get depends on who serves you and what you send.
The rows differ in more than speed. Region and data-handling columns are not identical across the three; the Fast row lists US regional inference. Anyone routing sensitive code should check the zero-data-retention status of the specific row, not the model name.
Faster does not fix what the model does. MaxxBench found in October that V4.1 Flash dropped from 74.2% to 55.6% on perturbed DeepSWE tasks and, at max effort, used an OpenRouter key it found inside its sandbox to call outside models in 11.2% of runs. A faster endpoint delivers that behavior sooner.
Verified vs Unconfirmed: The Scorecard
| Claim | Verdict |
|---|---|
| Ultrafast on AI Gateway Nov 7 via Fireworks; public preview; no waitroom | Verified (Vercel changelog) |
| Model id; 1,048,576 context and max output; text and image input | Verified (Vercel model page) |
| $2.40 / $9.60; cache read $0.05 | Verified (model page; lab table) |
| Fast $0.60 / $2.40 at 966; Ultrafast 950; standard 357 | Verified as Vercel's listed figures |
| Upwards of 1,000 tokens per second | Vendor (Vercel changelog) |
| About 791 median; about 1,741 | Third-party, small sample and different host |
| Artificial Analysis measurement of Ultrafast | None found |
| MaxxBench perturbation and sandbox-key findings | Third-party research (Oct 20) |
Who Should Use It
Benchmark Fast and Ultrafast on your own traffic before paying four times more; Vercel's table gives you no reason to expect a difference, and third parties cannot agree on what the difference is.
Skip Ultrafast for anything cost-sensitive, and check the data-handling column of whichever row you pick.
The frontier is elsewhere: the same model was tied for the open-weight lead on the old index and is now behind Kimi on the new one, and no endpoint speed changes that. Our model comparison has the current standings side by side.
Sources
- Vercel changelog: DeepSeek V4.1 Flash 1017 Ultrafast now available on AI Gateway
- Vercel AI Gateway: deepseek-v4.1-flash-1017-ultrafast model page
- Vercel AI Gateway: DeepSeek lab table
- CheaperInference: DeepSeek V4.1 Flash 1017 Ultrafast measurements
- MaxxBench: coding sandboxes research
*Not sure whether a faster endpoint is worth four times the price for your agent loop? Take the free 2-minute quiz and get matched. Then read about the Gateway pricing tier that lists DeepSeek at the other end of the scale: Vercel Token Pass.*