DeepSeek Posted the V4-Pro GA Notice Two Days After Shipping It: Benchmarks Are Now Official and Cached Input Gets 12x More Expensive

DeepSeek published the DeepSeek-V4-Pro general availability entry in its API change log on August 13, 2026, and released the weights on Hugging Face under the MIT license the same day. The notice puts ten benchmark figures under DeepSeek's own name, including Terminal Bench 2.1 at 87.9, DeepSWE at 62.7, and Cybergym at 83.3, and it announces peak and off-peak pricing that takes effect at 16:00 UTC on August 16, 2026. Cached input for deepseek-v4-pro rises from $0.003625 per 1M tokens to $0.044 at peak hours, a 12.1x increase. ASAP works only from figures verifiable in DeepSeek's own documentation and repository to show what this notice settled and which adoption assumption it breaks.

The notice filled three gaps: release notes, weights, and scores

The August 13, 2026 entry in the DeepSeek API change log states that the V4-Pro GA release rolled out across the app, the web client, and the API. The calling convention is unchanged, and setting the model name to deepseek-v4-pro reaches the new build. The same entry says the DeepSeek API now natively supports the OpenAI Responses API format with specific adaptation for Codex, and that thinking mode on V4-Pro and V4-Flash accepts three reasoning effort levels named low, high, and max.

The Hugging Face repository deepseek-ai/DeepSeek-V4-Pro-0813 was created at 03:05 UTC on August 13, 2026 and last modified at 16:28 UTC the same day. The license is MIT, the repository holds 66 safetensors files, and those files total 892.7 GB. The model card links arXiv 2606.19348 as the technical report.

The benchmark table ASAP declined to cite on August 13 has now changed source grade. The Terminal Bench 2.1 value of 87.9 that circulated then matches the value in the official change log and model card. An unofficial figure turned out to be correct, and that outcome does not conflict with the rule that unverified numbers stay uncited.

The entire price sheet is rewritten at 16:00 UTC on August 16

DeepSeek's pricing document states that new rates take effect at 16:00 UTC on August 16, 2026, with off-peak rates set at half the peak rates. Peak hours run 01:00 to 04:00 and 06:00 to 10:00 UTC, and every other hour is off-peak. New rates for deepseek-v4-pro per 1M tokens are $0.022 off-peak and $0.044 peak for cached input, $0.66 off-peak and $1.32 peak for cache-miss input, and $1.98 off-peak and $3.96 peak for output. For deepseek-v4-flash the pairs are $0.007 and $0.014, $0.22 and $0.44, and $0.66 and $1.32.

Increases differ sharply by line item. For Pro, cache-miss input goes from $0.435 to $1.32 at peak, a 3.0x rise; output goes from $0.87 to $3.96, a 4.6x rise; cached input goes from $0.003625 to $0.044, a 12.1x rise. Measured off-peak only, cache-miss input rises a modest 1.5x while cached input still rises 6.1x. Flash moves the same direction, with cached input up 5.0x at peak against 3.1x for cache-miss input.

The steepest increase lands on the cached input line

ASAP's August 13 article proposed raising cache hit rates as the defense against the announced increase. The published table inverts that advice, because cached input carries the largest multiple of any line item. A design that pins a fixed context in front and varies only the query absorbs no shock at all and instead takes the full 12.1x.

The structure explains why. Inside Pro, cache-miss price divided by cached price equals 120x today and 30x under the new sheet, and the ratio is identical whether computed at peak or off-peak rates. The caching discount narrows to a quarter of its former width. Flash sees the same compression, from 50x to 31.4x. Caching remains 30x cheaper, and the room to hold a bill down through caching alone shrinks accordingly.

Tier multiples also get tidied. Under current rates Pro costs 3.1x Flash on cache-miss input and output but only 1.3x on cached input, which left a band where a cache-heavy workload ran the larger model at close to the smaller model's price. The new sheet converges all three lines to between 3.0x and 3.14x. What DeepSeek actually removed in this revision is that inversion band, and the designs that exploited it absorb the largest increase.

The peak window sits directly on Korean working hours

Converted to Korea Standard Time, the 01:00 to 04:00 UTC window becomes 10:00 to 13:00 and the 06:00 to 10:00 UTC window becomes 15:00 to 19:00. Seven peak hours per day cover most of a Korean working day. In China Standard Time the same windows land at 09:00 to 12:00 and 14:00 to 18:00, matching mainland office hours exactly. The schedule is drawn around domestic traffic, and Korean users inherit it with no timezone cushion.

Off-peak hours in Korea are the 15 hours from 19:00 to 10:00 the next morning plus the two hours from 13:00 to 15:00. Interactive calls with a person waiting at the screen fall almost entirely inside peak, while nightly batches, scheduled indexing, and bulk evaluation runs can move wholesale into the half-price band. After the increase, effective unit cost depends more on scheduling than on model choice. Moving a midnight batch to 11:00 doubles that job's bill.

The score jump came from post-training and the harness, not new pretraining

The model card states that the GA build sits on the DeepSeek-V4-Pro preview model structure with a DSpark speculative decoding module attached. Holding structure constant, the distribution of gains becomes legible. DeepSWE rose from 12.8 to 62.7, a gain of 49.9 points; Cybergym rose from 52.7 to 83.3; NL2Repo rose from 38.5 to 61.5; Terminal Bench 2.1 rose from 72.1 to 87.9. HLE without tools rose from 37.7 to 42.7, a gain of 5.0 points.

The gap between a near fivefold jump on agentic tasks and a five point move on a knowledge task defines what this build is. The model learned almost nothing new, and nearly all of the gain sits in the procedure of calling a tool, reading the result, and choosing the next move. The model card notes that code agent tasks were evaluated using the minimal mode of DeepSeek Harness as the agent framework at max reasoning effort with temperature 1.0 and top_p 0.95.

That harness shipped the same day. The GitHub repository deepseek-ai/deepseek-harness was created at 11:56 UTC on August 13, 2026, carries the MIT license, and holds roughly 95,000 stars as of August 15, 2026. The July 31 V4-Flash notice had described the same harness as forthcoming, so DeepSeek moved to releasing model and execution environment together. Because a substantial share of the scores depends on that pairing, reproductions on other agent frameworks have room to land elsewhere.

The real barrier to open weights is 893 GB and a 4xGB300 node

The MIT license on this repository permits redistribution and commercial use without restriction, so the practical barrier to running DeepSeek-V4-Pro-0813 in house is volume rather than terms. The Hugging Face tensor index counts 1,650,497,936,906 parameters against 892.7 GB of files, which works out to 0.54 bytes per parameter and means most weights ship below 8-bit precision, already compressed once. The artifact to download is still 893 GB.

The vLLM command in the model card serves the model on a single node of four GB300 accelerators with fp8 KV cache, a block size of 256, and DSpark speculative decoding enabled. The SGLang path takes draft weights from the same checkpoint rather than a separate draft model path. The number an organization weighing self-hosting actually confronts is this hardware specification, not a benchmark score.

This release ships no Jinja chat template. In its place an encoding folder carries Python scripts and test cases that turn OpenAI-format messages into the model's input string and parse the output text back. Tool chains that read a standard template and handle the rest automatically must implement that step themselves for this model. The cost of putting open weights into production includes that integration work.

Even the official table leaves first place occupied

The comparison table on the model card does not claim a sweep. Kimi K3 leads on Terminal Bench 2.1 at 88.3 against 87.9, and also on DeepSWE at 67.5, Toolathlon-Verified at 76.5, and Agents' Last Exam at 27.6. Opus-4.8 leads on NL2Repo at 69.7, DSBench-Hard at 71.7, and HLE without tools at 49.8. Fable-5 with fallback takes the table's top value in several rows. The GA build leads its own table on Cybergym at 83.3 and AutomationBench at 31.8.

Every number in the table is vendor-run on DeepSeek's own harness rather than third-party verified, and the card does not detail the conditions under which competitor values were measured. Stated precisely, the GA build is an open-weights model that belongs on the same table as the strongest closed systems, leads on a few rows, and does not lead overall.

Layering the price change on top leaves one open question. The adoption case for DeepSeek models has long been performance per dollar, and after August 16 peak output for Pro costs $3.96 per 1M tokens, 4.6x the current rate. The trade is a narrower price advantage in exchange for performance near the top tier, and whether that trade favors a given team flips entirely on where its traffic sits relative to the peak window and how much of its input is cache reuse. No general answer exists before the first bills after August 16 arrive.

Source: DeepSeek API change log entry dated August 13, 2026; DeepSeek API models and pricing documentation; Hugging Face deepseek-ai/DeepSeek-V4-Pro-0813 model card and repository metadata; GitHub deepseek-ai/deepseek-harness repository