DeepSeek V4 Flash 0731 Intelligence, Performance and Price Analysis

TL;DR Highlight

DeepSeek V4 Flash 0731 scores 34 on the Intelligence Index at $0.28 per 1M output tokens

Who Should Read

Agent and coding-tool developers looking for a cheaper model to cut LLM API costs

Core Mechanics

Intelligence Index score of 34
- It beats the median of 19 for open-weight models of similar size.
- It is listed at #14 of 12134 overall.
Price is far below the median
- Input is $0.14 per 1M tokens versus a $0.43 median.
- Output is $0.28 per 1M tokens versus a $1.68 median.
Speed is 145.9 tokens per second
- That is faster than the 77.6 t/s median for similar open-weight models.
- Time to first token is 1.28s versus a 1.71s median.
It is verbose
- It generated 240M output tokens on the Intelligence Index versus a 140M median.
- The average cost per evaluation task is $0.17.
MoE with a 1M-token context window
- It has 284B total and 13B active parameters per token.

Community Reactions

Low price holds up in daily use
- One user codes all day with tools like reasonix or pi and pays only pennies.
- Another says a session stays under $1.
Token efficiency is a concern
- One comment says it uses about 3.6x more tokens than Gemini Flash 3.6 for the same work.
- Another questions the chart showing Kimi K3 using fewer tokens, which contradicts their experience.
Harness quality matters as much as the model
- One user improved the prompt and routing of a codebase Q&A agent and it matched closed-source SOTA on 50+ questions.
- Initially it gave bogus answers when there was no real reason behind a design.
Results vary by reasoning mode
- One user got a disappointing result in default reasoning mode and a much better one on high.
- Another asks whether the DSML tool call leaking seen in Flash and Pro was fixed.

How to Apply

Calculate your cost against the current model
- In 30 minutes, pull your monthly input and output token volume and compute it at $0.14 input and $0.28 output.
- Also run a scenario with output tokens 1.5x higher since this model is verbose.
Run the same prompts across reasoning modes
- Run 20-50 identical questions in default and high modes.
- Compare accuracy, output tokens, and response time in a table.
Add a guard for tool-call parsing
- Check logs for DSML-style tool-call tags leaking into text responses.
- If they leak, strip them in the parser and retry.
Improve your agent harness first
- Build a 50-question benchmark from internal tasks such as codebase Q&A.
- Tune prompts and routing, then compare models.

Try It

copy & run

# Paste into Claude or ChatGPT (cost-switch calculation prompt based on the article's data)

I run a service on an LLM API. Calculate my monthly cost if I switch to DeepSeek V4 Flash 0731 under these conditions.
- Current model: {{model name}}, input ${{input price}}/1M, output ${{output price}}/1M
- Monthly input tokens: {{input tokens}}M, monthly output tokens: {{output tokens}}M
- DeepSeek V4 Flash 0731: input $0.14/1M, output $0.28/1M, 80% cache discount
- This model is verbose: it used 240M tokens on the Intelligence Index versus a 140M median. Calculate with output tokens at 1.0x, 1.5x, and 1.7x.

Put it in a table and suggest 5 test items I should run first to validate quality.

Next step: ask it to compare accuracy and output tokens between default and high reasoning modes on 20 of your real prompts.

Terminology

Intelligence IndexA composite score from Artificial Analysis that combines multiple benchmarks. Think of it as a model's overall report card.

MoE (Mixture of Experts)An architecture that routes each token to a subset of expert blocks. Here only 13B of 284B parameters are computed per token.

Active parametersThe parameters that actually take part in generating one token. Like counting only the staff assigned to a job.

Cache discountA price cut when you resend an already-processed prompt. This model gives an 80% discount.

HarnessThe agent framework wrapping a model, handling tool calls and prompts. The same engine performs differently in different chassis.