DeepSeek V4 Flash Is 90% Off Through Novita on Vercel AI Gateway: The Cost Math

DeepSeek V4 Flash is 90% off on Vercel AI Gateway when you route through Novita, for Pro customers, through August 11, 2026. The official changelog, dated August 4, landed with no fanfare: pin Novita first with the order option and the effective rate drops from $0.14 to $0.014 per million input tokens, and from $0.28 to $0.028 per million output, against DeepSeek's official list prices verified today.

That is a tenth of the list price on the model this site's own agent stack runs on. V4 Flash is the 284B-parameter MoE (13B active per token) with a 1M-token context window that DeepSeek made the official API release on July 31 as build DeepSeek-V4-Flash-0731 - we covered that release and the OpenCode setup in our 0731 guide. The discount is a limited-time promo, not a price cut, which changes the calculus: it is a window to re-benchmark how much agent work you route to the cheap tier.

The before/after price sheet#

List prices verified against the DeepSeek pricing page on August 5, 2026. The discounted rate matches what Vercel's AI Gateway model page currently displays for the model (Input 0.014/token, Output 0.028/token, in their per-token notation). The changelog does not state whether the discount applies to cached input, so the math below counts cache at the standard rate. The discount's center of gravity is fresh input and output - exactly where agent workloads spend most tokens anyway.

How the deal works#

Three conditions matter before you build around this:

1. Pro plan required. The discount is only for Vercel Pro customers through August 11. Hobby accounts pay the standard rate.
2. Provider pinning via order. Set the model to deepseek/deepseek-v4-flash or deepseek/deepseek-v4-flash-0731 and put Novita first in the gateway order option. The changelog's example:

import { streamText } from 'ai';

const result = streamText({
  model: 'deepseek/deepseek-v4-flash', // or 'deepseek/deepseek-v4-flash-0731'
  prompt: 'Fix the failing tests in this repo.',
  providerOptions: {
    gateway: {
      order: ['novita'],
    },
  },
});

3. Fallback is automatic and priced at standard rates. If Novita cannot serve a request, the gateway falls back to other providers at the standard rate. After August 11 the model stays available at standard rates with no markup.

The cost-per-task math#

Take the workload V4 Flash is built for: a background agent inner loop - classification, extraction, retrieval synthesis - running 40 turns, call it 2M cumulative input tokens (1.5M fresh after cache reuse settles) and 60K output tokens per task.

At DeepSeek list price: 1.5M input at $0.14 ($0.21) + 0.5M cached at $0.0028 ($0.0014) + 60K output at $0.28 ($0.0168) = about $0.23 per task.

At the Novita rate through August 11: 1.5M input at $0.014 ($0.021) + 0.5M cached at $0.0028 ($0.0014) + 60K output at $0.028 ($0.0017) = about $0.024 per task.

Run that loop 10,000 times a month and the bill drops from roughly $2,300 to $240. This is the strongest counterexample yet to the AI affordability crisis worry that agent costs only ratchet upward: at these rates the question stops being "can we afford the fan-out" and becomes "why are we still paying for the frontier tier on this workload."

The honest caveat: list price is not cost per task. V4 Flash is 79-class on agent benchmarks per DeepSeek's 0731 release notes, not 80-plus-class like V4 Pro, and if the cheap model needs more retries or more turns on your workload, the 10x sticker advantage shrinks. The deal is worth exactly what your own evals say - but a two-week window at 90% off is the cheapest possible time to run them. Our V4 economics post has the full Flash-vs-Pro routing framework.

What it pressures#

A 90% off window on the cheapest competent agent model does two things to the market.

First, it moves the default routing conversation. The standard cost-control pattern is router-heavy: easy calls to a cheap model, hard ones escalate. During this window, the "easy" lane gets so cheap that the escalation threshold moves up - more workloads qualify for the cheap lane, which is the argument behind our LLM router comparison and the AI Gateway guide.

Second, it pressures the other cheap-access routes. GLM-5.2 and Kimi K3 both compete for the "free and cheap access" lane, and our GLM-5.2 access post and Kimi K3 access post track those routes. A two-week 90% off on V4 Flash is a promotional shot across that lane: anyone on the budget tier should re-measure V4 Flash against their current cheap route before August 11, not after.

And it is worth stating what it is not. This is a promo, not a price war - the DeepSeek list price is unchanged, and the fallback rate after the window is the standard one. Compare that with OpenAI's GPT-5.6 Luna 80% cut, which was permanent. The durable signal is distribution, not price: Vercel making its gateway the place where cheap models get cheaper, Novita buying volume through the fallback lanes. If you already pin providers, this is a no-op config change worth $2,000 a month on a real workload; if not, it is a cheap reason to learn the pattern before the window closes.

Sources#

• DeepSeek V4 Flash is 90% off through Novita on AI Gateway - Vercel changelog - the announcement, dated August 4, 2026, including the AI SDK example and fallback terms
• DeepSeek V4 Flash 0731 - Vercel AI Gateway model page - current displayed per-token rates
• DeepSeek pricing - official API docs - list prices verified August 5, 2026
• DeepSeek API updates - DeepSeek-V4-Flash-0731 release notes - the July 31 official release and benchmark claims

Continue Reading#

• DeepSeek V4 Flash 0731: The Official Release, Benchmarks, and How to Run It in OpenCode - what the July 31 release changed and the OpenCode setup
• DeepSeek V4 Economics: The Cost-Quality Frontier for Agentic Coding - Flash vs Pro routing, with the full pricing and benchmark tables
• Vercel AI Gateway Guide - how the gateway, provider pinning, and fallback lanes work
• GLM-5.2: Free and Cheap Access Routes - the competing budget-tier lane
• OpenAI Cuts GPT-5.6 Luna 80%: The Cost-Per-Task Math - the permanent price cut that set this promo in context