DeepSeek V4 Pro: the price that changes the routing break-even

A frontier model at general availability, at $0.87 output, changes the math before it changes any benchmark

DeepSeek moved V4 Pro from preview to general availability on August 12 and 13, 2026, in the DeepSeek-V4-Pro-0813 build, per Reuters. The model ships with a published price of $0.435 per million input tokens and $0.87 per million output tokens, a 1,048,576-token context window, and a stated focus on agent pipelines.

The detail the headlines miss: cache hits at $0.003625 per million change the arithmetic for anyone reusing context between calls.

What happened

DeepSeek had been running V4 Pro in preview since April. On Thursday, August 13, the company formalized the model as its general-availability flagship. Press corroboration is broad: beyond Reuters, Unite.AI confirmed the preview exit in the same window, with Yahoo Tech, Tech Times, and Cherry Creek News repeating the date.

What entered production in the 0813 build carries four marks. First, the price: $0.435 per million input tokens on cache miss, $0.003625 per million on cache hit, and $0.87 per million output. Second, the 1,048,576-token context window. Third, tool calling, which lets the model invoke tools inside an agent flow. Fourth, the stated emphasis on tool use, code execution, and multi-step workflows.

The company reports benchmark gains of up to 49.9 percentage points on some tasks. No independent evaluator had replicated those numbers as of this edition's cutoff, so they stay labeled as a company claim, not a validation. The same caution applies to a self-declared ranking that places V4 Pro below a frontier model referred to internally as "Fable 5" on agentic coding. With no verifiable source to reproduce that comparison, it sits outside the cost analysis that follows.

Why it matters

The number that matters is not the benchmark.

It is the cost per accepted task.

A team running agents pays for output. Every generated token is billed, and a multi-step flow multiplies the generation before the task is counted as resolved. In that regime, a frontier model at $0.87 output with a 1M context window shifts the break-even for anyone currently paying a single expensive provider for every task.

The argument rests on a market fact already covered here: cost per AI task has fallen as price competition between frontier providers accelerated through 2026. DeepSeek V4 Pro is the most recent event in that competition, not the first, and it is the output price that makes it relevant to the buyer, not its position on any benchmark table. The price fight stopped being a footnote and became the providers' sales pitch itself, which changes the nature of the buyer's choice: a model's advantage is no longer something you can hold onto, it gets renegotiated at every release.

There is a structural irony in this. The faster providers cut prices, the less sense it makes to buy loyalty to any one of them. The portfolio becomes the asset, and the route becomes the decision. Whoever keeps procurement pinned to a single name is paying to avoid the problem of choosing, and that invisible cost usually exceeds the benchmark gap between rivals.

The break-even appears when the account moves from the isolated 1M and lands on the task. Think of an agent that answers support emails: each ticket fires a call, the model reads the history, generates a reply, and the task only "counts" if the reply is accepted. At $0.87 output, the marginal cost drops to a fraction of what the pricier frontier providers charge for the same generation. If one reply in ten is rejected and needs regeneration, the effective price rises, but the route still comes out cheaper than keeping the expensive model on a hundred percent of the volume.

The direct consequence: the buying question stops being "which model has the best score?" and becomes "which route delivers the right capability for a given task, at the lowest cost per accepted task, with fallback when the cheapest route fails?" A benchmark in isolation cannot answer that. Only operations can.

What changes in practice

Three things change for anyone building a model portfolio. The output price dropped aggressively relative to the preview and to frontier providers betting on high prices. The 1M context enables tasks that once demanded manual chunk stitching. And the agent focus pushes V4 Pro toward exactly the kind of load that generates the most output tokens.

The change cheapest to ignore is the cache hit. An agent that repeats the same context base between calls pays almost nothing for the reused parts: the cache-hit price is $0.003625 per million, roughly 120 times below cache miss. Whoever designs to reuse the prompt prefix captures the cost drop; whoever does not keeps paying the cache-miss price every round, which cancels much of the model's advantage.

The order matters. The bucket of simple, repetitive tasks is the first to pay for the test, because that is where the acceptance rate tends to be high and the output price dominates the math. Long-reasoning workloads, with many steps and retraction, go last, because there the cost of error outweighs the token price.

What to do now

Measure first. Before changing any route, calculate the cost per accepted task you have today, because that number, not the isolated per-million price, decides whether V4 Pro enters the portfolio. From that baseline, the five steps below organize the test without risking delivery.

1. Slice the output price by cost per accepted task, not by the isolated 1M. A cheap model that gets it wrong and forces regeneration can cost more than an expensive one that solves it in one pass. Measure acceptance rate, not just the bill.

2. Split tasks by complexity. Simple, repetitive loads can route to V4 Pro; loads that demand a pricier frontier model stay where that model actually adds value, not everywhere.

3. Turn on fallback as a requirement, not an improvement. When the cheapest route fails to resolve the task, fallback to the reference route has to fire without reintegrating the application.

4. Enforce limits per key, per agent, per project. The savings from a cheap route vanish if a single agent loops without a ceiling.

5. Audit calls per task. Without a log of which model served each call and at what cost, the routing decision stays guessed, not measured.

The sequence matters because the steps chain together: without a measured acceptance rate there is no way to know which task is "simple"; without fallback, the "simple" task that turns complex breaks delivery; without a per-agent limit, one agent's loop eats the budget. The full flow is what turns a new price into a cost advantage instead of a catalog bet.

FAQ

Is DeepSeek V4 Pro available in production?

Yes. The company formalized general availability on August 12 and 13, 2026, in the DeepSeek-V4-Pro-0813 build, after a preview since April. The operational difference is that the published price and capability now stand as a stable offer, though DeepSeek's own list price can change over time; through Nexforce Router, the provider price is passed through as the final price, with no markup on cost per token.

What is the published price of DeepSeek V4 Pro?

$0.435 per million input tokens on cache miss, $0.003625 per million on cache hit, and $0.87 per million output tokens. These are company-published prices, not costs verified in production.

Is the 1M-token context real?

The company announces a 1,048,576-token context window with tool calling support. That enables long-context agent pipelines, provided the application actually uses caching so it does not pay the full price on every call.

Have the company's benchmarks been independently verified?

Not yet. DeepSeek reports gains of up to 49.9 percentage points, but no independent evaluator had replicated those results as of this edition's cutoff. Treat them as a company claim until further notice.

Should I switch from my current model to V4 Pro?

This is not a switching question. It is a routing question. V4 Pro breaks the cost break-even for lower-complexity agentic tasks, but only a test with fallback, a measured acceptance rate, and per-agent limits shows whether the route actually resolves the task.

References and further reading

• Reuters: DeepSeek releases official V4 Pro model as it steps up expansion
• Unite.AI: DeepSeek Ships V4 Pro as Its Flagship Model Leaves Preview
• AI model costs in 2026: the routing argument
• Model Router: how to prove the real economics of AI in production
• DeepSeek V4-Flash: the Flash model beats the Pro on agents
• Grok 4.6: the real test begins after the benchmark

The Nexforce Router at the landing point

The Router does not improve V4 Pro.

The Router passes the provider token price through as the final price, with no markup on cost per token. DeepSeek's own published list price can change, and that change comes from DeepSeek, not the Router. What it does is turn a catalog decision into an operational one: compare models side by side on cost, performance, latency, and context, apply fallback when the cheap route fails, enforce limits per key, agent, and project, and audit every call without reintegrating the application at each swap.

For the buyer paying a single frontier provider per task today, this week's economic event opens a real route to cut cost by routing lower-complexity tasks to a 1M-context model at $0.87 output, keeping the pricier model only where it adds value. The ability to compute that point per task, with evidence rather than a benchmark table, is exactly what the Nexforce Router puts into practice.