DeepSeek has finished assembling its newest AI model family, and the real reward for developers is a sharper choice, not another benchmark scoreboard. The Chinese startup now sells DeepSeek V4-Pro as its heavyweight and V4-Flash as the quicker, cheaper option. DeepSeek shipped the general release of V4-Pro on August 13, about two weeks after it pushed a refreshed Flash model into a public API beta on July 31. That pairing completes the DeepSeek V4 lineup for building AI agents.
Between them, the two tiers pose a single hard question for agent builders. How much added power is worth a far higher cost?
That question stings, because the performance gap looks far smaller than the price gap.
Pricing splits the DeepSeek V4 rate card three ways
DeepSeek reworked its prices on Aug. 16, and the new plan charges by the clock. Peak hours cost more, and off-peak hours cost exactly half.
Inside peak windows, V4-Flash runs $0.44 per million uncached input tokens and $1.32 per million output tokens. V4-Pro climbs to $1.32 for that same input and $3.96 for output. So Pro costs three times as much as Flash for uncached input and output within one rate period.
Those peak windows fall from 01:00 to 04:00 UTC and again from 06:00 to 10:00 UTC each day. Every other hour earns the cheaper off-peak rate.
That design rewards planning. Teams can push flexible background jobs into the low-cost hours and trim the bill. Customer-facing agents rarely get that break because users expect an answer immediately.
Both models still share a broad base. Each one reads up to 1 million tokens in a single exchange, and each writes up to 384,000 tokens back. Both handle tool calls, support OpenAI’s Responses API, and are compatible with Anthropic setups. Developers also pick low, high, or maximum reasoning effort, depending on how hard the job runs.
Flash lands within a point of Pro across DeepSeek V4 tests
V4-Pro stays the far bigger model. DeepSeek lists it at 1.6 trillion total parameters, with 49 billion active per request. Flash holds 284 billion parameters and fires 13 billion at a time. Both use a mixture-of-experts design, which wakes only a slice of the network for each token.
Yet size alone does not buy a matching lead. The independent firm Artificial Analysis scored V4-Pro at 53 on its Intelligence Index and V4-Flash at 52, testing both at maximum reasoning effort. That index blends nine evaluations spanning agent work, coding, math, knowledge, and long documents.
DeepSeek frames Flash as the model that nearly keeps pace with Pro. The company says Flash can even match its flagship on some, in its phrasing:
“simple Agent tasks”
Still, DeepSeek never spells out which jobs count as simple.
Flash also answers faster. Artificial Analysis clocked it at 107.2 output tokens per second against 80.2 for Pro, and Flash began replying in 1.19 seconds versus 1.74 for its sibling. That head start compounds when an agent calls a model dozens of times to close one task.
Cost per finished task reshapes the DeepSeek V4 math
Token prices tell only half the story. Artificial Analysis also measured what each model spent to complete one standardized task. Pro cost $0.25 per task, while Flash cost $0.11. That gap, roughly 2.3 times, runs tighter than the flat three-times spread, because it folds in how many tokens each model burned.
One point cuts against Flash there. It produced 210 million output tokens across the suite, against 130 million for Pro, and Artificial Analysis does not explain the extra load. Even so, Flash’s lower rates kept its average task cheaper.
A lone benchmark point, though, cannot settle a purchase. A failed task can trigger retries, human review, and fresh token spend. A pricier model that dodges those corrections can still win on the final invoice. So cost per successful task, not the sticker rate, becomes the number that matters most for DeepSeek V4 buyers.
DeepSeek V4 ships with an open agent harness
DeepSeek did not stop at the models. It also open-sourced DeepSeek Harness, a framework that runs an agent’s tools, workflow, sessions, sandbox, and interface. The project rides on the Cordis meta-framework and follows one rule:
“Everything is a plugin.”
Because every part plugs in the same way, developers can swap one piece, keep another, and add their own tools without adopting the whole stack. DeepSeek released the code under the MIT license, and teams can start a local interface with the command npx @deepseek-ai/dsh web. The repository drew tens of thousands of GitHub stars within days, a sign of strong early interest in the DeepSeek V4 ecosystem.
That reach carries a warning. DeepSeek calls Harness v0.1 a developer preview and expects updates that break compatibility. It has not shown proof of production reliability, stable interfaces, or enterprise support, so the tool fits experiments more than a long-term rollout for now.
The DeepSeek V4 battle turns on economics
For firms running hundreds or thousands of automated jobs, small per-task gaps pile up fast. Flash wins when it clears the quality bar at a lower price and higher speed. Pro earns its premium when harder work demands its extra reach, or when it prevents enough failures to pay for itself.
The smarter DeepSeek V4 buy, then, may hinge less on a leaderboard and more on which model finishes real work cleanly at the lowest total cost. Teams should test full jobs from their own operations, and they should count retries, review time, and speed next to the price.
That test may decide the DeepSeek V4 question for good, as agents move from demos into daily business work.
Which model would you pick for your AI agents, the stronger V4-Pro or the leaner V4-Flash?
Please share your views in the comments.