Vercel adds DeepSeek Ultrafast and paid web retrieval tools
Vercel announced DeepSeek V4.1 Flash 1017 Ultrafast on AI Gateway on November 7, 2026, giving developers another way to buy accelerated inference through Fireworks. A separate November 6 update adds DeepSeek Search and Fetch tools. Together, the changes let an application select a faster model route and retrieve web sources through the same gateway.
The tradeoff is price. Fireworks lists Ultrafast at eight times its standard Flash 1017 token rates. Developers should compare the cost and completion time of real tasks before choosing the faster route. The underlying model remains the October 17 checkpoint.
Faster serving for an existing model
Vercel’s November 7 announcement describes a public preview powered by Fireworks and claims generation above 1,000 tokens per second. That is a vendor performance claim, not a result measured by ByteForward. The announcement says there is no waiting room or session duration restriction.
The Gateway model reference identifies the route as deepseek/deepseek-v4.1-flash-1017-ultrafast. It lists text and image inputs, reasoning controls, tool calling, structured outputs and prefix caching. Its million-token context is shared by the prompt and response, rather than a separate million-token allowance for each.
Fireworks introduced its own accelerated preview on November 3, according to its changelog. November 7 marks the Vercel distribution announcement. ByteForward’s original DeepSeek Flash 1017 guide covers the earlier model update and direct DeepSeek API. Those releases should not be counted again as new weights.
The speed option carries a substantial token premium
Vercel lists $2.40 per million input tokens and $9.60 per million output tokens. The Fireworks pricing reference gives an exact cached-input rate of $0.048 per million, compared with the rounded $0.05 displayed in Vercel’s catalog.
Fireworks’ standard Flash 1017 rates are $0.30 for input, $0.006 for cached input and $1.20 for output. Each Ultrafast rate is eight times its corresponding standard rate. This comparison is between Fireworks serving options, not every provider in Gateway or DeepSeek’s separate direct API.
For illustration, a job using 100,000 uncached input tokens and 10,000 output tokens costs $0.336 at the listed Ultrafast rates, versus $0.042 at the listed standard Fireworks rates. That arithmetic excludes retrieval calls, retries and any other charges. It assumes identical token usage, which an actual workflow must measure.
Search and Fetch are separate tools
The November 6 tool release lets any compatible tool-calling model use DeepSeek retrieval through one Gateway key, without a separate DeepSeek account. Search discovers pages. Fetch retrieves page content as Markdown. Choosing Ultrafast does not by itself require either tool.
The bundled helpers require AI SDK 6.0.321 or later, or 7.0.160 or later. Each tool costs $0.75 per 1,000 billable requests, with model tokens charged separately. A research agent therefore has at least two costs to track: inference and retrieval.
Vercel’s web-search documentation limits each search query to 128 Unicode characters and allows one to 50 results, with ten as the default. Requesting more results does not change the tool’s per-request price, but sending larger results back to the model adds token consumption.
A successful nonempty search is billable even when results are later filtered. Fetch is billable when it returns nonempty content. Failures, timeouts and empty results incur no tool charge. Fetch still depends on the target website’s policies and availability. Applications need a useful response when a source cannot be retrieved.
Test complete tasks, including failures
The Fireworks serving guide describes Ultrafast as a preview whose behavior may change quickly. Priority tier is unavailable. Its rate-limit guide also gives Ultrafast lower adaptive limits than standard serving. Direct Fireworks account ceilings should not be treated as guaranteed capacity for a Gateway customer.
A useful evaluation records time to first output, total task time, successful completion, token usage, retrieval calls and retries. Faster generation may help a long coding response while making less difference to a task waiting on external tools. Keep the same prompts and acceptance checks when comparing routes.
Search-provider fallback also needs testing. Vercel describes provider choice, order and retry limits as model-directed and best effort. Verify what actually happened instead of assuming a failed search always triggers the intended replacement. Preserve source URLs and inspect whether retrieved pages support the final answer.
ByteForward has not benchmarked this preview. The confirmed change is a new Gateway access route and separate retrieval integration. The practical decision is whether the time saved on a particular workload justifies the higher inference bill.
Featured image: AI-generated ByteForward editorial illustration using the official Vercel logo. The gateway and document ribbon are conceptual, not a product interface. No endorsement is implied. Vercel, the Vercel design, Next.js and related marks, designs and logos are trademarks or registered trademarks of Vercel, Inc. or its affiliates in the US and other countries.
