DeepSeek V4-Flash: Flash Model Outperforms Pro on Agents

What Changed

DeepSeek put the flash model on top. The July 31 update changed only the agent-focused post-training while keeping the architecture from the previous preview version. Across nine benchmarks, the cheaper model now beats the V4-Pro-Preview. Cost collapsed. Capability rose. The model that costs less than half is also the most capable for agentic workloads.

What Happened

On July 31, 2026, DeepSeek released the public version of V4-Flash-0731, replacing the previous preview. The update was announced as a post-training change only: the company's official changelog states that the architecture and model size are identical to V4-Flash-Preview, the previous version in the same Flash line. No structural changes. Only the calibration shifted. The benchmark comparisons in the changelog are against V4-Pro-Preview; for readability, this article uses "V4-Pro" as shorthand for that reference model.

V4-Flash delivered nine agent benchmark results that DeepSeek describes as significantly better than V4-Pro-Preview. The numbers, taken from the July 31, 2026 changelog in the official API documentation:

• Terminal Bench 2.1: 82.7
• NL2Repo: 54.2
• Cybergym: 76.7

These three benchmarks cover the core of agent-environment interaction: terminal, repository, and offensive security. They measure real execution, not text comprehension.

• DeepSWE: 54.4
• Toolathlon verified: 70.3
• Agent Last Exam: 25.2

Here the focus shifts to end-to-end software engineering and tool use. DeepSWE is particularly relevant for companies running code agents in production.

• Automation Bench (Public): 25.1
• DSBench-FullStack (internal): 68.7
• DSBench-Hard (internal): 59.6

The last two are internal DeepSeek benchmarks, not externally auditable, but consistent with the pattern across the seven public benchmarks. None of these nine exercises is academic: each one reflects a capability that a real agent consumes daily.

The tests were run using the DeepSeek Harness in minimal mode, maximum effort level, top-p 0.95, and temperature 1.0. The company did not publish individual comparative values for V4-Pro-Preview on each benchmark, but the changelog wording is unambiguous: the aggregate result of V4-Flash surpasses V4-Pro-Preview across all nine metrics.

Beyond the numbers, V4-Flash debuted two capabilities absent from V4-Pro. The first is native support for the Responses API format, which enables direct integration with agent tools. The second is compatibility with Codex, the OpenAI coding assistant, including CLI, VS Code extension, and desktop application. V4-Pro will only receive Codex support in early August 2026, according to DeepSeek.

Why It Matters

Pricing tells the other half of the story. V4-Flash costs USD 0.14 per million input tokens with cache miss and USD 0.28 per million output tokens. V4-Pro costs USD 0.435 for input and USD 0.87 for output. The flash model is 3.1 times cheaper on both axes.

For a typical agentic workflow, where the model is called repeatedly in a loop, every call accumulates input and output tokens. The cost difference is not linear: it compounds. An agent that consumes 10 million input tokens and 5 million output tokens per day spends USD 2.80 on V4-Flash and USD 8.70 on V4-Pro. The daily savings of USD 5.90 becomes USD 2,153.50 per year from a single workflow.

What makes this number significant is the inversion. Until now, the market standard dictated that flash models sacrificed capability for speed and cost. They were the economic route in a routing policy, triggered for simple tasks: classification, extraction, unstructured text generation. When the task required multi-step reasoning, tool calls, or code execution, the router climbed to the Pro tier.

V4-Flash breaks that logic. The cheaper model is also the more capable one for the use case that consumes the most tokens: agentic tasks. A router configured with the classic rule "flash for simple tasks, Pro for complex ones" is now sending the most demanding work to the more expensive and potentially less competent model.

The competition data reinforces the asymmetry. V4-Flash offers 2,500 concurrent calls versus 500 for V4-Pro. For a company scaling agents horizontally, the five-times margin in concurrency means the flash model accepts five times more parallel work before hitting the rate limit ceiling. Combined with the 3.1× lower cost, the throughput per dollar is over fifteen times higher on parallelizable workloads.

DeepSeek also signaled a peak pricing model that will take effect soon: during the 9 a.m. to noon and 2 p.m. to 6 p.m. Beijing time (UTC+8) windows, prices double. For an enterprise router operating 24 hours, this additional layer of complexity turns model selection into a time-window problem: the cheapest route off-peak may not be the same as during Chinese business hours, and a routing system without hourly pricing support will lose money every day.

What Changes in Practice

The V4-Flash update reorders the routing hierarchy for agentic workloads: the cheapest model becomes the first choice for tool-based tasks, and V4-Pro retreats to specialist status in categories where benchmarks are still silent. The table below details each dimension of the shift.

The column that matters most for routing decisions is the first one. V4-Flash stopped being the contingency model and became the preference model for agentic workloads. The only meaningful degradation is the cost uncertainty introduced by the peak pricing model, which turns today's fixed price into a clock-dependent variable. It is a variable that doubles.

What to Do Now

Four decisions that the V4-Flash update places on the table of anyone operating model routing. None of them is optional. Inaction here is not neutral; it is a bet that V4-Pro will remain superior on agentic tasks, and the nine available benchmarks say otherwise.

1. Flip the routing rule for agentic workloads. If the current policy sends agent tasks to V4-Pro and simple tasks to V4-Flash, the order is backwards. Route agentic workloads to V4-Flash and reserve V4-Pro for situations where the agent benchmark is not the main discriminator. The most obvious candidate is long-form text generation with high stylistic requirements, but until DeepSeek publishes comparisons in that category, the available evidence points to V4-Flash as the correct route for any workload involving tools.

2. Recalculate the annual cost of agentic workflows. Multiply your agents' daily token volume by the price difference between V4-Flash and V4-Pro. For a single workflow of 10M input tokens and 5M output tokens per day, the switch saves over USD 2,000 per year. For a company with dozens of agents in production, the savings buy infrastructure. Include the five-times concurrency effect in the calculation: the 500 concurrent call ceiling on V4-Pro may force purchasing additional capacity from another provider, while V4-Flash with 2,500 calls absorbs the peak.

3. Prepare the routing layer for hourly pricing. DeepSeek's peak/off-peak announcement turns cost per token into a function of Beijing time. A router that does not check the window before selecting the model will pay double without knowing it. If the Nexforce Router is in the stack, hourly pricing rules are a configuration parameter, not a code change.

4. Test V4-Flash via Codex before the procurement decision. The Codex CLI and VS Code integration let you run V4-Flash as a coding backend without writing a single API call. DeepSeek's setup script automates the configuration. Testing real performance in an agentic development workflow produces more useful data than comparing benchmark scores.

FAQ

The nine published benchmarks cover exclusively agentic tasks: code, automation, and tools. DeepSeek did not publish comparisons for long-form text generation, pure mathematical reasoning, or prose quality. These gaps are what the questions below address, and the honest answer to several of them is "not yet known."

Is V4-Flash better than V4-Pro at everything?

No. The nine published benchmarks cover exclusively agent tasks: code, automation, tool use, and full-stack software engineering. For tasks like long-form text generation, pure mathematical reasoning, or prose quality, DeepSeek has not published comparisons, and V4-Pro may retain an advantage. V4-Flash is superior on agentic workloads, not universally.

Does V4-Flash have the same architecture as V4-Pro?

The official July 31, 2026 changelog states that V4-Flash-0731 keeps the same architecture and size as V4-Flash-Preview, the previous version in the Flash line. DeepSeek has not published information about the architectural relationship between V4-Flash and V4-Pro. What is known: V4-Flash received post-training focused on agent capabilities, and the result of that tuning surpasses V4-Pro-Preview on the nine benchmarks.

Why does the flash model have integrations the Pro model lacks?

DeepSeek prioritized V4-Flash for the Codex and Responses API launch because the cost and latency profile of a flash model aligns better with agentic workloads, which make frequent looped calls. V4-Pro will receive Codex support in early August 2026, per the official documentation.

What changes with peak pricing?

Prices double during the 9 a.m. to noon and 2 p.m. to 6 p.m. Beijing time (UTC+8) windows. For a company in New York (UTC-4), the price peak falls between 1 a.m. and 4 a.m. and between 6 a.m. and 10 a.m., partially overlapping the morning workday. For a company in London, the peak covers most of the business day.

Is V4-Flash the right model to replace V4-Pro in my application?

It depends on the workload. If the application uses agents, calls tools, or executes code, the answer is probably yes, and the recommendation is to test via Codex before changing the route. If the application depends on writing quality, long-form formatting, or pure mathematical reasoning without tools, V4-Pro remains the safer choice until DeepSeek publishes benchmarks in those categories.

References and Further Reading

Primary sources for the event:

• DeepSeek API Docs: Change Log, 2026-07-31. Official changelog with the nine benchmarks and technical notes from the update.
• DeepSeek API Docs: Integrate with Codex. Official documentation for Codex integration, including the auto-install setup script and connection parameters.
• DeepSeek API Docs: Models & Pricing. Current pricing table and the peak/off-peak announcement with applicable time windows.

For the buyer who decides on cost, not technical scores:

• LLM Benchmarks for CFOs: What Matters Beyond the Technical Score. How to read model benchmarks through a buyer's lens, separating what moves the budget from what moves the paper.
• LLM Cost Comparison in 2026: Smart Routing with Nexforce Router. Cost as a function of the route, not the model: the premise V4-Flash just validated with nine benchmarks.

For the routing architecture that turns this inversion into real savings:

• Model Router: The Middleware Missing From Your AI Stack. How a routing layer reverses single-vendor dependency and turns benchmark events into automatic route decisions.
• Unified API for Multimodal Models: LLM Gateway in 2026. Gateway architecture as a cost and capability control surface, with hourly pricing as a configuration parameter.

The Road Ahead

DeepSeek turned a golden rule of the LLM market into an exception. The flash model beat Pro-Preview on nine agent benchmarks. And not by a small margin: the advantage came with fivefold concurrency, support for tools the more expensive model still lacks, and a price 3.1× lower. Anyone operating model routing who does not reposition V4-Flash in the hierarchy is paying more for less capability, and the equation only worsens when peak pricing takes effect. For the Nexforce Router buyer, this inversion reinforces a truth the product has exploited since launch: the model decision is a route decision. It changes with the benchmark, the workload, and the clock, and anyone who outsources that logic to a fixed pricing table buys the wrong answer every day.