DeepSeek V4 Flash 0731 Pricing and Benchmark Analysis

DeepSeek has moved DeepSeek-V4-Flash-0731 into public beta without creating a new API route or raising its regular token prices. The existing deepseek-v4-flash name now serves the re-post-trained checkpoint. DeepSeek also published the model weights under the MIT license. Its architecture and size are unchanged; V4 Pro and DeepSeek’s app and web models were not upgraded.

For buyers, the launch matters because it improves the low end of the coding-agent cost curve. See the live DeepSeek pricing page, full API price comparison, and token cost calculator for current values.

DeepSeek V4 Flash 0731 pricing

DeepSeek lists a 1 million-token context window, maximum output of 384,000 tokens, thinking and non-thinking modes, and concurrency of 2,500. It supports Chat Completions, the Anthropic-compatible interface, tools, JSON, and the Responses API.

The live table and chart above compare Flash with V4 Pro and nearby intelligence leaders using today’s canonical data. DeepSeek did not change Flash’s regular token rates on July 31, but token price alone does not prove the same cost per completed task.

DeepSeek’s pricing page now warns that it plans a significant overall API price increase, but it does not disclose the new rates, affected billing items, or effective date. AI Pricing Guru has therefore kept the canonical dataset on the current published rate card. See our rapid-response price increase analysis for the confirmed details and buyer checklist.

What DeepSeek reports improved

DeepSeek’s changelog calls out agent and coding results rather than a broad general-knowledge scorecard:

Benchmark	DeepSeek-reported 0731 score	Caveat
Terminal Bench 2.1	82.7	Public benchmark; DeepSeek harness minimal mode, max effort
NL2Repo	54.2	Vendor-reported run
Cybergym	76.7	Vendor-reported run
DeepSWE	54.4	Vendor-reported run
Toolathlon Verified	70.3	Vendor-reported run
Agent Last Exam	25.2	Vendor-reported run
Automation Bench Public	25.1	Vendor-reported run
DSBench FullStack	68.7	Internal DeepSeek test set
DSBench Hard	59.6	Internal DeepSeek test set

DeepSeek says the code-agent benchmarks used its unreleased minimal harness at maximum effort, top_p=0.95, and temperature 1.0. A tuned private harness may not transfer to another coding agent, lower effort setting, or shorter output budget.

DeepSeek says 0731 is adapted for Codex-style Responses API integrations. The stable model name avoids migration work but also upgrades behavior in place, so regression-test prompts calibrated for the preview.

The open weights create a second deployment path, but DeepSeek’s official vLLM recipe requires a four-GPU GB300 node. Most application teams should compare API economics before self-hosting.

Independent price-performance result

Artificial Analysis independently tested the reasoning, maximum-effort configuration. Its Intelligence Index v4.1 gives DeepSeek V4 Flash 0731 a score of 49.93, up from 40.28 for the earlier V4 Flash snapshot. That is a gain of 9.65 points, or about 24% relative.

The score is nearly level with Gemini 3.6 Flash at 50.07 and higher-effort GPT-5.6 Luna variants. Artificial Analysis ranks 0731 first for weighted cost per Intelligence Index task on its comparison page.

The catch is token use. Artificial Analysis reports 210 million output tokens for the Intelligence Index evaluation, compared with a 62 million median across models on the page. The model is cheap per token, but a long reasoning trace can consume part of that advantage. Procurement should track cost per accepted task, latency, and output volume instead of assuming the lowest rate card always creates the lowest bill.

Artificial Analysis also reports an Omniscience Index of -15.72. That metric rewards correct knowledge and penalizes hallucinated answers without penalizing refusals. It is a warning against reading the coding-agent gains as universal factual reliability.

Labs status: included, but the 0731 refresh is blocked

DeepSeek V4 Flash is already part of the AI Pricing Guru Cost-per-Task Labs leaderboard. The last accepted run scored the previous snapshot at 48 correct answers out of 49 with zero API errors. Because it predates July 31, the public label now says pre-0731.

We attempted a fresh 49-task run at 11:38 UTC on July 31. All requests returned 402 Insufficient credits, so the publication guard rejected the incomplete run.

A valid refresh requires funded credit and confirmation that deepseek/deepseek-v4-flash serves 0731. Until then, labeling the old 48/49 result as 0731 would be misleading.

Buyer verdict

DeepSeek V4 Flash 0731 is a high-priority test for coding agents, repository work, terminal automation, and tool-heavy workflows where cheap tokens allow more verification and retries. The in-place route upgrade makes adoption easy, and the unchanged rate keeps the economic case intact. Compare current alternatives on the Google AI pricing page, OpenAI pricing page, and our DeepSeek vs ChatGPT guide.

It is not an automatic replacement for premium models. The independent run shows high verbosity, the official benchmark configuration depends on a DeepSeek harness that is not yet public, and two headline scores use internal test sets. Teams should compare accepted changes, review findings, rollbacks, latency, and total token volume on their own repositories.

For third-party hosting, compare Novita’s listed DeepSeek routes, but confirm the served snapshot before treating any generic V4 Flash listing as 0731.

Affiliate disclosure: AI Pricing Guru may earn a commission from the sponsored Novita link at no extra cost to you. It does not affect the pricing or benchmark analysis.

The practical move is to canary 0731 on a fixed set of real tasks, cap its reasoning budget, and measure cost per accepted result. If its agent gains hold under your harness, the unchanged rate can buy materially more useful work than the preview. If verbosity or factual reliability dominates, route only the well-scoped coding and tool tasks where Flash earns its cost advantage. For the same-day counterargument to generic prompt routing, read why Manifest deprecated its LLM router.

Sources: DeepSeek’s official July 31 changelog, models and pricing documentation, Codex integration guide, official Hugging Face model card, Artificial Analysis’ corrected DeepSeek V4 Flash 0731 page, and the Hacker News discussion.