DeepSeek V4 Pro 0813 in 2026: The Rate Card Expires This Week

Why Does the DeepSeek Rate Card Change This Week in 2026?

Because DeepSeek is retiring flat per-token pricing in favour of time-of-day billing. Its API pricing page states that the new prices take effect at 16:00 UTC on 16 August 2026, that pricing moves to peak and off-peak billing, and that off-peak rates are half the peak rates. Peak hours in 2026 are 01:00 to 04:00 and 06:00 to 10:00 UTC. Every other hour is off-peak.

Much of the coverage in 2026 reports this as a 17 August change. That is not a contradiction. DeepSeek is based in China, and 16:00 UTC on 16 August is midnight on 17 August in Beijing. The version of the date that matters is the one in your own timezone, because that is when your invoice starts behaving differently.

The deadline in your timezone

The cutover is a single global instant, so it lands on a different clock face depending on where your finance team sits. For teams in India it arrives on the evening of 16 August 2026, not the 17th. Working backwards from the published UTC timestamp is the only reliable way to know how long you actually have.

What the new rate card says

The new card splits every existing line into a peak rate and an off-peak rate, with off-peak fixed at exactly half of peak. The comparison below sets DeepSeek's published pre-change rates against the published peak and off-peak rates for V4 Pro. The multiplier column is straightforward division of the new figure by the old one, and it is the number most teams have not yet worked out.

This is where the widely repeated 12x headline comes from, and it deserves precision. Cache-hit input rising from 0.003625 to 0.044 US dollars is roughly 12.1 times, and it is the largest single multiplier on the card. It is not the multiplier on your bill unless your traffic is almost entirely cache hits, which is unusual. Quoting 12x as the overall increase overstates what most teams will pay.

What Does the Rate Change Do to a Real Workload in 2026?

It roughly doubles to roughly quadruples the monthly bill, depending on cache behaviour and when your jobs run. The illustration below uses DeepSeek's published rates and a set of volumes we invented for the purpose. The volumes are not a forecast, a survey result or a client figure. They exist only to show how the multipliers combine, and you should substitute your own usage before drawing any conclusion.

Assume a mid-sized content and support workload of 40 million cache-miss input tokens, 60 million cache-hit input tokens and 12 million output tokens in a month. That shape favours a heavy system prompt reused across many calls, which is typical of marketing automation in 2026.

The interesting row for Indian teams is the middle one, and it is not a coincidence. Peak hours of 01:00 to 04:00 and 06:00 to 10:00 UTC translate to 06:30 to 09:30 and 11:30 to 15:30 IST. A conventional 09:00 to 18:00 Indian working day therefore sits inside a peak window for four and a half of its nine hours. Interactive workloads driven by staff at their desks in 2026 will land close to a fifty-fifty split by default.

That points at the cheapest available mitigation. Batch work with no human waiting on it, such as overnight content generation, bulk classification and reporting, can be scheduled into off-peak hours at half the peak rate without changing application logic. Interactive traffic cannot be moved, so the saving is partial.

What Is DeepSeek V4 Pro 0813 in 2026?

DeepSeek V4 Pro 0813 is the general availability build of DeepSeek's flagship mixture-of-experts model, rolled out on 13 August 2026 across the DeepSeek app, web and API. It supersedes the earlier preview version, adds substantially stronger agent behaviour according to the vendor, introduces selectable thinking effort levels and ships native support for the OpenAI Responses API format.

The published specification is unusually generous at the price. A one million token context window and a 384,000 token output ceiling put it in the same structural class as far more expensive models in 2026, which is precisely why cost-led buyers picked it up so quickly.

One detail matters for anyone comparing tiers. DeepSeek also sells V4 Flash, whose 0731 build carries the same one million token context and 384,000 token output ceiling at a fraction of the price. The tier decision in 2026 is therefore not about context length, which is the assumption many teams start from.

Why Should You Not Read the Headline Benchmarks as a Buying Signal in 2026?

Because almost every headline number is DeepSeek comparing the new build against its own previous build. That is a legitimate way to report progress on a point release, and it is what most vendors do in 2026. It answers the question of whether the engineering team improved something. It does not answer the question a buyer is actually asking, which is whether this model beats the specific alternative under consideration.

Put plainly: if last year you ran a hundred metres in twenty seconds and this year you run it in twelve, that is a real and impressive improvement. It tells you nothing about whether you would win the race you have entered. The self-comparison is a measure of the vendor's progress, not of your options.

The figures below are the vendor-reported preview-to-GA gains circulating in 2026. DeepSeek published them as an image in its announcement rather than as text, so these come from secondary coverage transcribing that image rather than from a table we could read directly. Treat them as vendor-reported and directional.

Two of those rows deserve a flag. DSBench-Hard and DSBench-FullStack are described in the surrounding 2026 coverage as DeepSeek's own internal test sets rather than public third-party benchmarks. A score on a test the vendor wrote, scored and reported is the weakest form of evidence available, even when the vendor is acting in complete good faith.

This pattern is not unique to one company. The recent Gemini 3.7 Flash release in 2026 published its gains the same way, and the cross-vendor benchmark comparisons we looked at for GPT-5.6 show how quickly the picture changes once numbers are measured on a common harness.

What Did the Independent Evaluation Actually Find in 2026?

Independent measurement exists, and it tells a more sober story than the launch numbers. Artificial Analysis published its own evaluation of V4 Pro 0813 in 2026, placing it at 53 on its Intelligence Index at maximum reasoning effort. That index is a composite built from around ten separate benchmarks spanning reasoning, coding, agentic tool use and knowledge, run on a common harness across models.

Fifty-three is a solid mid-table result rather than a frontier one. The same index places Claude Opus 5 at 63, Claude Fable 5 at 62, GPT-5.6 Sol and Grok 4.6 at 61 and Kimi K3 at 60, with GLM-5.2 level with DeepSeek at 53. For a model at a fraction of frontier pricing, that is a genuinely strong value position, and it is a much more useful sentence than any preview-to-GA delta.

What Did DeepSeek Actually Publish, and What Is Still Missing in 2026?

More than the commentary suggests, and this is worth correcting because several write-ups in 2026 describe the release as entirely silent. DeepSeek published a news entry dated 13 August 2026 and a matching changelog entry, both covering the GA rollout, the agent improvements, the low, high and max thinking effort levels, native Responses API support optimised for Codex, and the upcoming pricing change. There is also a Hugging Face model card for DeepSeek-V4-Pro-0813 with open weights under an MIT licence.

What is thinner is the evaluation trail. The benchmark table in the announcement is published as an image rather than as text, which means the scores cannot be read by tooling, quoted precisely without transcription, or checked against a machine-readable record. The announcement carries no documented evaluation methodology, no description of the harness and no baseline conditions.

None of that is misconduct. Publishing a chart as an image is a formatting choice plenty of vendors make in 2026, and DeepSeek released the weights, the model card and a technical report, which is more openness than several better-funded competitors offer. It is a procurement consideration rather than a scandal: when the numbers live only in an image and the methodology is undocumented, there is less to hold the vendor to if performance shifts after you have built on it.

This is the same category of issue we flagged when GLM-5.3 shipped without its weights in 2026. The question is what documentation you could point at in six months if the model you deployed stopped performing the way it did at launch.

How Should Teams Evaluate a Cheap Model in 2026?

By measuring cost per completed task rather than cost per token, and by treating any published rate as a temporary condition. Per-token price is the input to a cost model, not the cost model itself. A model that is half the price but needs three attempts, longer prompts or heavier review is more expensive in practice, and that difference never shows up on a pricing page.

The expiry point generalises well beyond one vendor. Gemini 3.7 Flash launched in 2026 on an explicitly introductory rate that doubles on 1 January 2027, and DeepSeek is restructuring its card in August 2026. When two independent vendors reprice within months of each other, the reasonable conclusion is that cheap introductory AI pricing in 2026 is a market phase rather than a baseline, and budgets should be built on that assumption.

What Are the Common Mistakes in 2026?

The recurring errors are all versions of treating a launch-day figure as a durable fact. Each of these is avoidable with an hour of work, and each of them has cost teams real money in 2026.

• Budgeting on a rate card without checking its expiry. The rate you signed off on in July 2026 may not be the rate you pay in September.
• Reading a self-comparison as a competitive result. A model beating its own preview build says nothing about the alternative you are evaluating against.
• Quoting the largest multiplier as the overall increase. The 12x figure applies to cache-hit input at peak, not to a realistic blended bill.
• Ignoring time-of-day billing when scheduling. Batch jobs left on default schedules will drift into peak windows and cost double for no benefit.
• Assuming the flagship tier is required. In 2026 the independent index separates V4 Pro from V4 Flash by one point at several times the output price.
• Migrating production before running your own evaluation. Published benchmarks are measured on tasks that are not your tasks, on prompts that are not your prompts.
• Treating cost per token as cost. Retries, longer reasoning traces and human review time are the part of the bill that pricing pages never show.

Key Takeaways for 2026

The short version for anyone with a decision to make this week. DeepSeek V4 Pro remains a strong value proposition even after the change, and the argument here is about arithmetic and evidence quality rather than about avoiding the model.

• DeepSeek's published pricing changes at 16:00 UTC on 16 August 2026, which is 21:30 IST the same evening and midnight on 17 August in Beijing.
• Peak hours in 2026 are 01:00 to 04:00 and 06:00 to 10:00 UTC, and off-peak rates are exactly half of peak.
• Verified multipliers at peak are about 12.1x on cache-hit input, about 3.0x on cache-miss input and about 4.6x on output.
• A realistic blended bill lands closer to two or three times current cost than to twelve, so re-run your own numbers rather than reacting to the headline.
• An Indian working day overlaps DeepSeek's peak windows for roughly half its hours in 2026, so moving batch work off-peak is the fastest available saving.
• The headline benchmark gains are self-comparisons against DeepSeek's own preview build and have no competitor in the comparison.
• Independent evaluation in 2026 places V4 Pro 0813 mid-table at 53, one point above DeepSeek's much cheaper V4 Flash 0731.
• The release was documented, with a news post, changelog entry, model card and MIT-licensed weights. What is missing is a machine-readable benchmark table and a documented evaluation methodology.
• Introductory and promotional AI pricing in 2026 is a temporary market condition. Build budgets and exit plans that assume rate cards move.

If your model choice in 2026 was made on price, the work this week is small and specific: pull your last thirty days of token usage, split it by cache hit, cache miss and output, apply the peak and off-peak rates above, and see which side of your threshold the answer falls on. For a fuller picture of how the current model field compares, our breakdowns of what GPT-5.6 actually changed in 2026 and the Gemini 3.7 Flash release cover the same ground for the other major vendors.