DeepSeek V4 Flash 0731: The Budget Tier Just Overtook Pro Preview on Agent Benchmarks

DeepSeek shipped a quiet update on July 31 that changes the agent economics picture: V4 Flash got re-post-trained into an agent specialist, and the numbers it publishes are well past what V4 Pro Preview managed in the spring. The pricing model is unchanged - $0.14 per million input tokens, $0.28 output - which makes this the first time a budget-tier open-weights model is posting frontier-class agent benchmark scores at commodity prices. It also picked up native Responses API support with first-party Codex integration, so there is a real workflow to try today.

What Actually Changed#

The update is a post-training run, not a new architecture. From the official change log: DeepSeek-V4-Flash-0731 "keeps the same model architecture and size as DeepSeek-V4-Flash-Preview, and was only re-post-trained." The API is now officially in public beta, you call it with the same deepseek-v4-flash model name, and the pricing page already lists the model version as DeepSeek-V4-Flash-0731.

What changed is where the model is aimed. The agent benchmarks DeepSeek published are the headline:

DeepSeek's own framing is that these results "far exceed V4-Pro-Preview." Read the fine print: the public-benchmark runs used the DeepSeek Harness minimal mode (the harness itself is unreleased, "to be released soon"), at max effort, with topp=0.95 and temperature=1.0. DSBench-FullStack and DSBench-Hard are internal test sets, so the two most developer-facing numbers, Terminal Bench and DeepSWE, are the ones worth benchmarking against other models yourself. These are vendor-reported scores with the config documented, which is more than most releases ship with, but they are not third-party verified.

The other two changes matter as much as the scores:

• Responses API, native. V4 Flash is the only DeepSeek model that supports the Responses API format today. The pricing page confirms Pro does not support it yet, with support expected in early August.
• Codex adaptation. DeepSeek publishes a Codex integration guide with a one-click setup script, and the integration works across Codex CLI, the ChatGPT desktop app, and the VS Code extension. Currently only deepseek-v4-flash supports Codex; deepseek-v4-pro is expected to follow in early August.

Why This Matters for Developers#

Put the two facts together and the story is simple: the cheapest tier of the cheapest frontier family is now a credible agent runtime, and it plugs into the most popular open coding agent with no shim.

When V4 launched in April, our developer guide pinned Flash at 54.7 SWE-bench Verified, roughly R1-class, and recommended it for "bounded, low-stakes decisions" in agent inner loops - classifiers, extraction, high-volume work where a cheap model will not embarrass you. The 0731 update keeps those economics and pushes the model into the driver's seat. Terminal Bench 2.1 at 82.7 and DeepSWE at 54.4 are hard long-horizon benchmarks - real terminal sessions and issue-resolving loops, not multiple-choice. If the numbers hold up in independent testing, the "cheap model for inner loops, expensive model for hard tasks" split that the cost analysis walked through starts bending: the inner loop model can now handle the whole loop.

Three implications worth thinking about this week:

1. Agent cost-per-solve just dropped again. The economics post computed Flash's standing rates at $0.14 input / $0.28 output with $0.0028 cache hits. A coding agent that burns a few hundred thousand tokens per task now pays pennies for DeepSWE-class scores. The cost ceiling on "let the agent retry" gets much higher when retries are this cheap, and higher retry budgets are how agent success rates go up. Pair that with the peak/off-peak pricing notice on the pricing page (2x during 9:00-12:00 and 14:00-18:00 Beijing time once it launches) and batching agent work into off-peak hours becomes real money.

2. Codex support is the distribution play. The setup script is one line: bash <(curl -fsSL https://cdn.deepseek.com/api-docs/codex-deepseek-setup-en.sh) (PowerShell irm ... | iex on Windows). It backs up ~/.codex/config.toml, writes a ~/.codex/models.json catalog declaring context window and reasoning-effort levels, and adds a [model_providers.deepseek] section, validating syntax before writing. Your MCP servers and project trust settings are preserved. That means a Codex user can A/B the same task against DeepSeek V4 Flash and OpenAI models in the same client, the fastest way to see whether the benchmark spread shows up in your work. Our agentic dev stack breakdown covers where DeepSeek fits in the broader tooling picture.

3. The Pro release is now the wild card. The change log says the official V4 Pro release "will follow soon," and the Codex docs expect Pro Responses API support in early August. If Flash posts these scores after a post-training pass, the same treatment on the 1.6T-parameter Pro base is worth watching. The open-weights showdown looked different when Flash was the cheap option and Pro the aspiration; a post-trained Pro changes that math again.

How to Try It Today#

Path 1: Codex. Install the DeepSeek provider with the setup script above, launch Codex, and pick deepseek-v4-flash from the model menu. Works in Codex CLI, the desktop app, and the IDE extension with one configuration.

Path 2: The Responses API. DeepSeek's Responses API is the OpenAI-style /v1/responses dialect. A minimal call looks like:

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["DEEPSEEK_API_KEY"],
    base_url="https://api.deepseek.com",
)

response = client.responses.create(
    model="deepseek-v4-flash",
    input="Explain the tradeoff between tool retries and token cost in a coding agent, in 3 sentences.",
)
print(response.output_text)

Path 3: Any OpenAI-compatible agent. If your agent talks to https://api.deepseek.com with the OpenAI Chat Completions dialect, nothing changes - the same deepseek-v4-flash name now resolves to the 0731 checkpoint. For teams still on the legacy aliases, the deepseek-chat migration guide covers the cutover.

The Honest Caveats#

Three things keep this from being an unqualified win. First, the benchmarks are vendor-reported and two of the nine are internal sets; the public ones were run with an unreleased harness configuration, so expect independent verification to move the numbers. Second, this is still a 158B-parameter model - not the deepest reasoner in the world - and hard multi-hour agent sessions are exactly where the spend guardrail conversation applies. Third, only Flash moved; if your pipeline routes on model tier, the Pro numbers in your comparison tables are stale until the GA release lands. Re-run your own evals on the 0731 checkpoint before rewiring anything in production - baseline receipts still apply.

FAQ#

What is DeepSeek V4 Flash 0731?#

It is a re-post-trained version of DeepSeek V4 Flash, released July 31, 2026, with the same architecture and size as the previous V4-Flash-Preview. The update focuses on agent capabilities: higher agent-benchmark scores, native Responses API support, and first-party Codex integration. The API is now officially in public beta.

How much does DeepSeek V4 Flash 0731 cost?#

Unchanged from the April pricing: $0.14 per million input tokens, $0.28 per million output tokens, and $0.0028 on cache hits. DeepSeek has announced a future peak/off-peak policy that doubles prices during Beijing-time peak hours.

How do I use DeepSeek V4 Flash with Codex?#

Run the one-click setup script from the Codex integration guide (one bash line on macOS/Linux, a PowerShell equivalent on Windows), then select deepseek-v4-flash in Codex CLI, the desktop app, or the VS Code extension. It backs up your existing config first.

What agent benchmarks did DeepSeek publish for V4 Flash 0731?#

Terminal Bench 2.1 at 82.7, Cybergym at 76.7, Toolathlon verified at 70.3, DeepSWE at 54.4, NL2Repo at 54.2, Agent Last Exam at 25.2, Automation Bench (Public) at 25.1, plus two internal sets (DSBench-FullStack 68.7, DSBench-Hard 59.6). These are vendor-reported, run with DeepSeek's unreleased harness in minimal mode.

Did the V4 Pro API change with this update?#

No. The update only covers the V4 Flash API. The V4 Pro API and the APP/WEB models are unchanged, and DeepSeek says the official V4 Pro release will follow soon, with Responses API support expected in early August.

Sources#

• DeepSeek change log, 2026-07-31: DeepSeek-V4-Flash Update
• DeepSeek models and pricing, fetched 2026-07-31: api-docs.deepseek.com/quick_start/pricing
• DeepSeek Codex integration guide: api-docs.deepseek.com/quick_start/agent_integrations/codex
• HN front page, July 31 2026: DeepSeek-V4-Flash Update and DeepSeek V4 Flash 0731 analysis (both charted on the front page the day of release)

Continue Reading#

• DeepSeek V4 Economics: Cost, Quality, and the Frontier Agentic Coding Case - the full pricing and benchmark picture, updated for the new Flash numbers
• DeepSeek V4 on a Budget: Coding Agent Cost Benchmarks - what cheap agent loops actually cost in practice
• DeepSeek V4: The Developer's Guide to Flash and Pro - architecture, SDK setup, and thinking mode
• The Agentic Dev Stack in 2026 - where DeepSeek fits among the coding agents and runtimes
• DeepSeek-chat to V4 Migration Guide - cutting over legacy model aliases