DeepSeek V4-Pro Pricing: 5 Essential 2026 Changes

DeepSeek V4-Pro pricing changed the day the model left preview: the flagship’s flat $0.87-per-million-token output rate is being replaced by peak and off-peak billing, with peak-hour output climbing as high as $3.96 per million tokens starting August 16, 2026. The GA release, internally labeled build 0813, ended a nearly four-month preview period and shipped across DeepSeek’s app, web, and API on August 13, 2026.

This guide breaks down the new pricing tiers, what actually changed under the hood, how the model performs on coding benchmarks, and whether the switch makes sense for your workload. If you’re comparing options across the frontier-model field, our AI coding assistants comparison is a useful companion read. Everything below is sourced from DeepSeek’s own release notes and independent reporting published within days of the GA launch, so the numbers reflect the current, not preview-era, state of the API.

What Is DeepSeek V4-Pro and Why Did the Pricing Change?

DeepSeek V4-Pro is DeepSeek’s flagship large language model, and its pricing changed because the company moved the model from a nearly four-month preview into general availability on August 12–13, 2026. According to DeepSeek’s own API changelog, the GA build (designated 0813) keeps the preview’s one-million-token context window and 384,000-token maximum output unchanged in name, but the underlying agent performance jumped enough that DeepSeek repriced output tokens rather than leave preview-era rates in place indefinitely.

Tech outlet The Decoder reported that DeepSeek paired the GA launch with two other moves: raising API prices and open-sourcing its in-house agent software, DeepSeek Harness. Read together, the timing suggests DeepSeek is pricing V4-Pro closer to what its new agentic capabilities are actually worth, rather than treating the preview discount as permanent.

How Much Does DeepSeek V4-Pro Cost Now?

DeepSeek V4-Pro costs $0.435 per million input tokens and $0.87 per million output tokens under the 0813 build’s initial published rate, but that flat output price stops applying once the new schedule takes effect at 16:00 UTC on August 16, 2026. From that point, output pricing splits by time of day: peak-hour output rises to $3.96 per million tokens, while off-peak output holds at half the peak rate.

Peak hours are defined as 01:00–04:00 UTC and 06:00–10:00 UTC; every other hour of the day counts as off-peak. That’s roughly seven peak hours and seventeen off-peak hours daily, so a workload that can tolerate scheduling around those windows keeps most of its costs close to the old flat rate, while one that runs constantly during peak hours will see a meaningfully higher bill.

DeepSeek V4-Pro Pricing: Before and After the August 2026 Change

What Is Peak and Off-Peak Billing on DeepSeek V4-Pro?

Peak and off-peak billing on DeepSeek V4-Pro means the API charges different output-token rates depending on the UTC hour a request completes, instead of one flat rate around the clock. It’s a demand-shaping mechanism: DeepSeek’s infrastructure sees heavier load during specific UTC windows, and time-based pricing nudges cost-sensitive customers toward off-peak hours without hard rate limits.

For developers, the practical takeaway is that batch jobs, evaluation runs, and anything without a hard latency requirement should be scheduled inside the 17 off-peak hours whenever possible. Interactive, user-facing traffic that has to run whenever a real person is using it will simply absorb the peak-hour rate during those windows, which is why understanding your own traffic’s time distribution matters more with this pricing model than with a flat-rate API.

What New Features Shipped With the DeepSeek V4-Pro GA Release?

The DeepSeek V4-Pro GA release shipped native support for OpenAI’s Responses API format, a one-click Codex setup, and explicit low, high, and max reasoning-effort controls, according to DeepSeek’s own August 13, 2026 announcement. Native Responses API support matters most for teams that already built tooling around OpenAI’s API shape, since it removes the adapter layer they previously needed to point that tooling at DeepSeek instead.

The one-click Codex setup lowers the barrier for developers who want to try DeepSeek V4-Pro inside a Codex-style coding workflow without hand-rolling configuration files. Reasoning-effort controls let a single API call trade off latency and cost against answer depth — low effort for quick lookups, max effort for multi-step agentic tasks — a pattern also used by other frontier providers, including in our GPT-5.6 guide and Gemini 3.7 Flash guide.

For teams already integrated against OpenAI’s Responses API, migration effort is close to zero: pointing an existing client at DeepSeek’s endpoint with an updated base URL and key is often the whole change, since the request and response shapes match. Teams using DeepSeek’s older chat-completions-style integration aren’t forced to migrate — that format still works — but new projects gain the most from starting on the Responses API shape given where DeepSeek is clearly investing its compatibility effort.

How Does DeepSeek V4-Pro Perform on Coding Benchmarks?

DeepSeek V4-Pro’s Terminal-Bench 2.1 score jumped from 72.1 in the April 2026 preview to 87.9 in the August GA release, a 15.8-point gain that puts it within striking distance of the field’s top agentic coders. On the DeepSWE benchmark, the jump was even sharper: from 12.8 to 62.7, according to MindStudio’s hands-on benchmark review published after the GA launch.

That Terminal-Bench 2.1 score of 87.9 sits close to competing frontier models scoring in the high-80s on the same benchmark, and multiple agent-focused benchmarks reportedly show V4-Pro edging past Claude Opus 4.8 on specific tasks. Benchmark scores vary by task category and test harness, so treat any single number as directional rather than a guarantee your exact workload will see the same gain — pair it with your own evaluation set before switching a production pipeline over.

What Do Terminal-Bench 2.1 and DeepSWE Actually Measure?

Terminal-Bench 2.1 measures how well a model completes realistic, multi-step tasks inside a command-line interface — things like debugging a failing build, chaining shell commands, or navigating an unfamiliar codebase without a graphical editor. It’s designed to stress agentic behavior rather than single-turn question answering, which is why a 15.8-point jump is a meaningful signal about DeepSeek V4-Pro’s agent capability specifically, not just its raw language modeling quality.

DeepSWE focuses more narrowly on software-engineering tasks that mirror day-to-day developer work: implementing a feature from a spec, fixing a specific bug, or refactoring code while preserving existing behavior. The jump from 12.8 to 62.7 is unusually large for a single release cycle, which is exactly why it’s worth verifying against your own tasks rather than assuming it transfers directly to your codebase’s language, framework, or coding conventions.

What Is DeepSeek Harness and Why Did DeepSeek Open-Source It?

DeepSeek Harness is the agent software DeepSeek uses internally to run coding-agent workloads, and the company open-sourced version 0.1 as an MIT-licensed Developer Preview alongside the V4-Pro GA launch. It’s positioned as an alternative to closed agent frameworks like OpenAI’s Codex CLI and Anthropic’s Claude Code, giving developers a transparent reference implementation instead of a black-box wrapper.

DeepSeek runs its own official benchmarks using what it calls “minimal mode” — a stripped-down configuration limited to shell access and a file editor, with no extra tool scaffolding. That choice matters for interpreting the benchmark numbers above: a score achieved in minimal mode tends to reflect the model’s raw reasoning more than a heavily tool-augmented agent stack would, which is a fairer comparison point against other base models.

Because it’s tagged as a Developer Preview rather than a stable 1.0 release, expect DeepSeek Harness’s interfaces and configuration format to keep shifting for a few months. Teams evaluating it for internal tooling should pin a specific commit or release tag rather than tracking the main branch, the same caution that applies to any early-stage open-source agent framework before its API surface settles.

DeepSeek V4-Pro Pricing: 5 Essential Changes at a Glance

1. Flat $0.87/M output pricing ends August 16, 2026, replaced by time-of-day billing.
2. Peak-hour output jumps to $3.96/M during two UTC windows totaling seven hours a day.
3. Off-peak output settles at $1.98/M — still more than double the original preview rate.
4. Input pricing stays unchanged at $0.435/M regardless of time of day.
5. The price increase arrives bundled with real capability gains: native Responses API support, reasoning-effort controls, and a large Terminal-Bench 2.1 jump.

How Does DeepSeek V4-Pro Pricing Compare to Other Frontier Models?

DeepSeek V4-Pro’s new off-peak output rate of $1.98 per million tokens still undercuts many closed frontier models on a pure per-token basis, even after the increase, while its peak-hour rate of $3.96 per million tokens moves it into the same neighborhood as several premium API tiers. Input pricing at $0.435 per million tokens remains aggressive either way, since input tokens typically dominate cost in retrieval-heavy or long-context applications.

Where DeepSeek V4-Pro sits relative to Claude, Gemini, and GPT-family pricing depends heavily on your traffic’s time distribution and whether you’re paying for output-heavy agentic tasks or input-heavy retrieval tasks. For a detailed breakdown of one major competitor’s tiers, see our Claude Sonnet 5 pricing guide, which uses the same per-million-token framework to make direct comparison easier.

A concrete example makes the comparison easier to reason about. A workload generating 10 million output tokens a month, run entirely during off-peak hours, would cost roughly $19.80 in output charges under the new DeepSeek V4-Pro rate — up from about $8.70 under the original flat rate, but still a fraction of what many premium closed-model tiers charge for the same volume. Run that identical workload entirely during peak hours instead, and the output bill nearly doubles again to around $39.60, which is the gap that makes scheduling matter.

What Does Time-Based AI Pricing Mean for the Industry?

DeepSeek V4-Pro’s peak and off-peak split is one of the more explicit examples of a frontier lab pricing compute the way a utility company prices electricity, and it’s a pattern worth watching rather than a one-off. As agentic workloads that consume far more output tokens than simple chat become common, providers have a stronger incentive to shape demand across the day instead of over-provisioning for a global peak that only lasts a few hours.

For budgeting purposes, this means the era of quoting a single per-token price for an AI vendor is becoming less reliable. Teams evaluating any frontier model going forward should ask not just “what’s the rate” but “what’s the rate at the hours my traffic actually runs,” since the answer can now differ by more than 4x on the same model depending on when a request happens to complete.

Should You Switch to DeepSeek V4-Pro for Your Project?

Switch to DeepSeek V4-Pro if your workload is output-light, tolerant of scheduling around UTC off-peak hours, or specifically needs its strong Terminal-Bench 2.1 and agentic coding performance; hold off if you run constant, latency-sensitive, output-heavy traffic that would land mostly in the new peak-hour window. Batch processing, offline evaluation, and asynchronous coding-agent tasks are the clearest fits, since they can be scheduled deliberately.

Real-time chat applications, customer-facing assistants, and anything that can’t predict or control when its traffic lands are the workloads most exposed to the new peak-hour multiplier. For those cases, it’s worth benchmarking DeepSeek V4-Pro against your own evaluation set and comparing the effective blended rate — factoring in your actual peak-versus-off-peak traffic split — against alternatives before committing a production pipeline to it.

A practical migration checklist before switching a production workload: pull your last 30 days of API logs and bucket them by UTC hour to estimate your real peak-versus-off-peak split; run your own held-out evaluation set through V4-Pro rather than relying on published benchmarks alone; confirm your client library or framework supports the Responses API format DeepSeek now prioritizes; and set a budget alert at the new blended rate before rolling out to full production traffic, not after the first invoice arrives.

What Are the Risks of Relying on Vendor-Published Benchmarks Alone?

The risk of relying only on vendor-published benchmarks is that a lab naturally showcases the tasks and configurations where its own model performs best, and “minimal mode” testing — while fairer than a heavily tool-augmented setup — still reflects DeepSeek’s own harness choices rather than your production environment. A 15.8-point Terminal-Bench gain is a strong signal that something real improved, but it doesn’t tell you how the model handles your specific codebase’s conventions, your internal tooling, or edge cases your team has hit before with other providers.

The safest path is to treat published benchmarks as a filter for which models are worth testing, not as the final decision. Run a small, representative slice of your actual workload — real tickets, real bug reports, real feature specs — through DeepSeek V4-Pro and whatever model you’re currently using, side by side, before committing budget or engineering time to a full migration.

Frequently Asked Questions

When exactly does DeepSeek V4-Pro’s new pricing take effect?

The new peak and off-peak pricing schedule takes effect at 16:00 UTC on August 16, 2026. Before that time, the API bills at the flat 0813-build rate of $0.435 per million input tokens and $0.87 per million output tokens.

Does DeepSeek V4-Pro’s input pricing change with the new schedule?

No. Input token pricing stays at $0.435 per million tokens around the clock. Only output token pricing splits into peak and off-peak tiers under the new schedule.

What is the context window and maximum output for DeepSeek V4-Pro?

DeepSeek V4-Pro’s GA build keeps a one-million-token context window and a maximum output of 384,000 tokens, unchanged from the preview version that ran for nearly four months before graduating to general availability.

Is DeepSeek Harness required to use DeepSeek V4-Pro?

No. DeepSeek Harness is an optional, open-source agent framework DeepSeek released alongside V4-Pro. You can call the V4-Pro API directly, through the Responses API format, or through any other agent framework you already use.

How much more expensive is peak-hour DeepSeek V4-Pro output versus the original rate?

Peak-hour output at $3.96 per million tokens is roughly 4.5 times the original flat rate of $0.87 per million tokens. Off-peak output at $1.98 per million tokens is about 2.3 times the original rate.

Are the DeepSeek V4-Pro benchmark gains verified independently?

The Terminal-Bench 2.1 and DeepSWE figures cited here come from DeepSeek’s own release materials and MindStudio’s independent hands-on benchmark review; as with any vendor-adjacent benchmark, run your own evaluation set before making a production decision.

DeepSeek V4-Pro pricing is more complex than it was a week ago, but it’s not necessarily worse for every workload. Off-peak output at $1.98 per million tokens remains competitive, peak-hour output at $3.96 per million tokens is a real cost to plan around, and the accompanying capability jump — native Responses API support, reasoning-effort controls, and a 15.8-point Terminal-Bench gain — is substantial enough that the repricing reads as a capability tax rather than a pure markup.

Sources: DeepSeek API Docs, “DeepSeek-V4-Pro GA Release,” August 2026; The Decoder, “DeepSeek ships improved V4 Pro, open-sources its agent software, and raises API prices,” August 2026.