DeepSeek released the general-availability build of its flagship model this evening. DeepSeek-V4-Pro-0813 is now what the deepseek-v4-pro API alias serves, and OpenRouter lists it as the GA release, closing a preview period that opened on April 24. It is a 1.6-trillion-parameter mixture-of-experts model with 49B active per token, a 1-million-token context window, MIT-licensed open weights, and a price of $0.435 per million input tokens and $0.87 per million output.

There is an awkward number inside that launch. DeepSeek's own model card puts V4 Pro at 80.6% on SWE-bench Verified. Twelve days ago the company shipped a much smaller, much cheaper sibling, V4 Flash, and last week the independent evaluator vals.ai measured that model at 88.8% on the same benchmark using a neutral bash-only harness. The flagship is arriving with a self-reported coding score eight points under what its budget model actually scores when somebody else runs the test.

RelatedQwen3.8-Max lands at 2.4T with no benchmark table

  • 1.6T total, 49B active. A hybrid attention stack combining Compressed Sparse Attention and Heavily Compressed Attention, with MoE experts in FP4 and the rest of the network in FP8.
  • 1M in, 384K out. The context window is the headline capability, and the maximum output length is unusually large for a model at this price.
  • MIT license, weights published. Available on Hugging Face and ModelScope, so self-hosting is a real option rather than a press-release option.
  • The coding claim is vendor-reported. No independent evaluator has run V4 Pro yet. vals.ai has run the plain DeepSeek V4, which scored 77.4% and is a different model.

What exactly shipped today?

Not a new architecture. The 0813 build is the production checkpoint of the V4 Pro that has existed in preview since late April, and DeepSeek's change log on July 31 said plainly that the official V4 Pro release would follow soon after the Flash update. That is what landed. The calling convention did not change: anything already pointed at deepseek-v4-pro now gets the 0813 weights without a code edit, which is convenient for existing integrations and slightly annoying for anyone who wanted to pin a version.

The specs are the interesting part. A 1.6T sparse model that activates 49B parameters per token is roughly a 33-to-1 sparsity ratio, and the FP4 experts are what make serving it at these prices arithmetically possible. DeepSeek is not selling capacity here so much as selling the absence of a capacity tax.

SWE-bench Verified: DeepSeek V4 Flash, V4 Pro and V4 comparedHorizontal bar chart. DeepSeek V4 Flash scores 88.8 percent, independently measured by vals.ai. DeepSeek V4 Pro claims 80.6 percent, vendor reported. Plain DeepSeek V4 scores 77.4 percent, independently measured.SWE-BENCH VERIFIEDThe GA flagship claims less than its own budget model measuresDeepSeek V4 Flashindependent88.8%DeepSeek V4 Provendor claim80.6%DeepSeek V4independent77.4%Independent figures: vals.ai, mini-swe-agent bash-only harness. Vendor figure:DeepSeek model card, Pro-Max mode. The two were not run on the same harness.genztech.blog
Fig 1 · benchmark The eight-point gap is real but not like-for-like. Only the two grey bars were produced by an outside evaluator on a neutral harness.

Why does the cheaper model score higher?

Two things are going on, and only one of them is about model quality.

The first is provenance. V4 Flash's 88.8% comes from vals.ai, which runs every model itself on one harness and publishes the error bars. V4 Pro's 80.6% comes from DeepSeek, produced with DeepSeek's own scaffolding in what the model card calls Pro-Max mode. Those two numbers were not generated under the same conditions and are not directly comparable. Across the models on our leaderboard, vendor-published SWE-bench figures have run 2.6 to 11.6 points above the neutral harness, because SWE-bench scores the model and its agent loop together. Read straight, the gap between Pro and Flash is eight points. Read with that correction in mind, the honest answer is that we do not yet know how V4 Pro does on a level harness, and history suggests the number will not go up.

The second is that V4 Flash was re-post-trained specifically for agentic work. It is a 284B model with 13B active, roughly a fifth the size of Pro, and DeepSeek tuned it against Terminal-Bench and Codex-style workflows. Size does not decide SWE-bench. Task-specific post-training frequently does.

How do the two DeepSeek models actually compare?

V4 Pro (0813)V4 Flash (0731)DeepSeek V4
SWE-bench Verified80.6%88.8%77.4%
Score provenanceVendor-reportedIndependent (vals.ai)Independent (vals.ai)
Parameters1.6T total / 49B active284B total / 13B activeNot published
Context1M1M1M
Input / 1M tokens$0.435$0.14Superseded
Output / 1M tokens$0.87$0.28Superseded
LicenseMIT, open weightsMIT, open weightsMIT, open weights

Pro costs about 3.1 times what Flash costs per million output tokens, and on the one benchmark most people buy a coding model for, it currently claims less. Where Pro clearly earns the premium is everywhere that is not agentic patch-writing: 90.1 on GPQA Diamond, 93.5 on LiveCodeBench, 87.5 on MMLU-Pro, a 3206 Codeforces rating, and 55.4 on SWE-bench Pro against Flash's unpublished figure. Those are hard-reasoning and competitive-programming numbers, and a 1.6T model should win them.

RelatedDeepSeek V4 Flash claims 82.7 on Terminal-Bench 2.1

How DeepSeek V4 Pro routes a one-million-token promptFlow diagram. A prompt of up to one million tokens enters a hybrid attention stage combining Compressed Sparse Attention and Heavily Compressed Attention, then a mixture-of-experts layer holding 1.6 trillion parameters that activates only 49 billion per token, producing up to 384 thousand output tokens.V4 PRO · TOKEN PATH1.6T parameters in the file, 49B doing work per tokenPrompt1M ctxHYBRID ATTENTIONCompressed SparseHeavily CompressedMoE LAYER · FP4 EXPERTS49B active1.6T total~1 in 33 parameters usedOutput384KSparsity plus FP4 experts is what makes $0.87 per million output tokens work.The rest of the network stays in FP8. Weights are MIT-licensed and downloadable.genztech.blog
Fig 2 · architecture The orange sliver is the share of the model that fires on any given token. Everything else sits on disk.

Who should care, and who should not?

If you run an agent that reads a repository, edits files, and runs tests, Flash is the better buy right now at a third of the price, and that conclusion holds whether you trust the vendor numbers or the independent one. If your workload is long-context analysis, competitive-programming-shaped problems, or single-shot reasoning over a million tokens of documents, Pro is the DeepSeek model that fits, and the price is still far below the closed frontier.

For teams self-hosting, the MIT license is the whole story. A 1.6T model in FP4 is not casual hardware, but the weights are genuinely published rather than gated behind a research license, which is a meaningfully different proposition from most models near the top of our AI coding leaderboard. On that board V4 Pro currently sits at rank 15 on its vendor score. V4 Flash sits at rank 6 on an independently measured one. Claude Opus 5 leads at 97.0%.

What it means for the market

DeepSeek is privately held, so there is no ticker to watch. The exposure sits with everyone selling inference. An MIT-licensed 1.6T model with a 1M context at $0.87 per million output tokens sets a public floor that closed-model pricing has to argue against, and it does so with weights anyone can download and serve. The signal for investors is margin compression at the commodity end of the API market, not at the frontier: Opus 5 and GPT-5.6 Sol are not competing for the same buyer as a self-hosted MIT model. Watch whether the major inference providers list V4 Pro quickly and at what markup, and watch whether the next closed-model price cut arrives inside a quarter.

The timeline that got here

  1. Apr 24, 2026V4 Pro and V4 Flash preview weights published Open-weight preview, no GA date given
  2. Jul 31, 2026V4 Flash updated to the 0731 build Change log states the official V4 Pro release will follow soon
  3. Aug 5, 2026vals.ai publishes 88.8% for V4-Flash-0731 Independent, mini-swe-agent bash-only harness
  4. Aug 12, 2026V4-Pro-0813 ships as GA API alias repointed, OpenRouter lists it the same day
  5. NextAn independent SWE-bench run on V4 Pro Until then the 80.6% stays a vendor figure on our board
What to watch · 2026
  • The independent number. When vals.ai runs V4-Pro-0813 on the bash-only harness, expect it below 80.6%. If it lands above, DeepSeek's scaffolding is unusually honest and that is a story in itself.
  • Whether Flash quietly becomes the default. A budget model outscoring the flagship on the benchmark that sells coding models is an awkward product line, and DeepSeek may resolve it with a Pro refresh rather than a price change.
  • Serving economics. FP4 experts at 49B active is what makes $0.87 per million output work. If third-party hosts cannot match that price, the open weights matter less than the API does.
  • Version pinning. The alias repoint means production traffic moved to new weights today without anyone asking. Teams that care about reproducibility should pin the dated build.

Our take

The launch is real and the model is good. The framing around it is where the care is needed. A 1.6T open-weight model with a million-token context at under a dollar per million output tokens is a serious piece of engineering, and the reasoning benchmarks back it up. But the coding claim is the one being carried in headlines, and it is the one number here that nobody outside DeepSeek has checked. We are leaving V4 Pro on our leaderboard at its vendor score with the provenance stated plainly, exactly as we do for every unverified figure, and Flash stays nine ranks above it on a number somebody else produced. That ordering may look strange. It is the only ordering the evidence supports today.

Primary sources

Original analysis by GenZTech, built from DeepSeek's primary documentation and independent benchmark data.