On August 12, 2026, the Chinese lab DeepSeek moved its flagship to general availability. The model, designated DeepSeek-V4-Pro-0813, appeared on the company's app, web interface, and API. There was no blog post, no changelog entry, and no press release. A revised set of benchmark figures went up, a note said prices would rise on August 16, and that was the announcement.
For a model that DeepSeek's own numbers place within a tenth of a point of the current frontier on a leading agent benchmark, the silence is itself a data point. It is worth separating the three distinct claims tangled together in this release: that the model is the largest open-weight system available, that it performs at frontier level on agent tasks, and that it does so far more cheaply than closed peers. The first is verifiable now. The second and third rest, so far, entirely on the vendor's own reporting.
What is verifiable: the largest open-weight model shipped
DeepSeek V4-Pro-0813 is a mixture-of-experts model with 1.6 trillion total parameters, of which roughly 49 billion are active on any given token. That total makes it the largest open-weight model currently available for download, ahead of the open releases from Meta, Alibaba, Moonshot, and others. The context window is one million tokens, and the model is tuned for agent work: tasks where a system uses tools, executes code, and completes multi-step workflows without a human in the loop.
The mixture-of-experts design is the reason the headline parameter count and the running cost point in different directions. A dense 1.6-trillion-parameter model would be prohibitively expensive to serve. By activating only about 49 billion parameters per token, DeepSeek keeps inference cost far below what the total size implies, which is how it can list the prices it does. This is the same architectural lever that made earlier DeepSeek releases notable, applied at a larger scale.
DeepSeek
Meta
Qwen
Moonshot AIThe open-weight framing matters beyond leaderboard bragging rights. A downloadable model can be hosted on private infrastructure, inspected, fine-tuned, and run without sending data to a vendor's API. That is a genuinely different deployment model from a closed frontier system, and it is the axis on which open releases compete even when their raw scores trail. The V4-Pro release keeps that option open at a size that was, until recently, the exclusive territory of closed labs.
What is claimed: frontier agent scores
Here the ground gets softer. DeepSeek reports that V4-Pro-0813 scores 87.9 on Terminal-Bench 2.1, an agent benchmark that tests a model's ability to operate a command line to complete real tasks. That figure sits one tenth of a point behind Anthropic's Claude Fable 5 at 88.0. The company also reports large gains over its own April V4 Preview: from 72.1 to 87.9 on Terminal-Bench 2.1, from 52.7 to 83.3 on CyberGym, and from 12.8 to 62.7 on DeepSWE, a software-engineering agent test.
The reported jump from Preview to GA is on agent tasks
DeepSeek's own figures for the V4-Pro-0813 general-availability build versus its April V4 Preview, on three agent-focused evaluations. These are vendor-reported and not yet independently replicated. Higher is better.
The largest reported gain is on DeepSWE, a software-engineering agent test, which the company says rose nearly fifty points.
If these numbers hold up under independent testing, they describe a meaningful release: a nearly fifty-point jump on DeepSWE in a few months is a large move on a hard evaluation. But the conditional is the point. As of this writing, no third-party evaluator has replicated the scores. Benchmark figures published by the lab that trained the model, without an accompanying technical report or independent confirmation, are a starting hypothesis, not a settled result. This is not a claim unique to DeepSeek; it is the standard caveat for any self-reported benchmark, and it applies with extra force when the release skips the usual documentation entirely.
A near-frontier score published by the lab that trained the model, with no technical report and no independent replication, is a hypothesis, not a result.
On reading vendor benchmarks
There is also a well-known gap between benchmark performance and production behavior. Agent benchmarks are constructed tasks with clear success criteria. Real agent work is messier: ambiguous instructions, brittle tool integrations, and long-horizon runs where small errors compound. A model can top a terminal benchmark and still be frustrating to deploy, or vice versa. The scores are worth watching precisely because they are so high, but they should be treated as a prompt for independent evaluation rather than a conclusion.
What follows from the price
The third claim is the one that would matter most if the second holds. DeepSeek lists V4-Pro-0813 at $0.43 per million input tokens and $0.87 per million output tokens, with cache hits far cheaper still. For comparison, the model it sits level with on Terminal-Bench, Claude Fable 5, lists at roughly $10 per million input and $50 per million output. On output tokens, which dominate the cost of generative and agentic work, that is about a 57-fold difference.
A near-identical headline score at a fraction of the output price
Listed price per 1M output tokens, with each model's Terminal-Bench 2.1 score for context. The bars are scaled to the higher price. Lower price for a comparable score is the entire argument for an open-weight challenger.
On output tokens, DeepSeek lists a price roughly 57 times lower than Fable 5 for a score within a tenth of a point. Sticker price is not total cost, but the gap is the story.
Two caveats keep this honest. First, list price is not total cost. Different models consume different numbers of tokens to finish the same task, and a cheaper-per-token model that takes more turns can erase part of its own advantage. A full cost comparison needs end-to-end task runs, not just the price card. Second, the announced August 16 price increase means the launch pricing is a promotional floor, not a fixed rate. Still, even a substantial increase would leave a very large gap. If the performance is real, the price is the disruptive part: it changes the arithmetic of running frontier-grade agents at scale.
Why ship it this way
The quiet release is not an accident of a busy news week. DeepSeek has a pattern of letting the weights and the numbers speak, releasing on its own cadence rather than orchestrating a launch. Shipping a general-availability model with a bare set of figures and a scheduled price rise fits that pattern. It also has a side effect worth naming: with no technical report, the burden of verification falls entirely on the outside community. The scores will be confirmed, revised, or contested by independent labs over the coming weeks, and until then the release is best read as an invitation to test rather than a finished claim.
The broader context is a frontier where the top scores are converging and the competition is shifting to cost and openness. A closed lab defends a score with a premium price. An open-weight lab attacks the same score with a downloadable model and a low price card, and lets the market decide whether the two are really equivalent on the work that matters. V4-Pro-0813 is a clean instance of that second strategy at the largest scale yet attempted in the open.
The practical read
For a team deciding what to actually run, the release changes the option set more than it settles anything. There is now a downloadable, million-context, agent-tuned model with vendor-reported scores at the frontier and a price card far below closed peers. Whether it belongs in production depends on evaluation you run yourself, on your own tasks, against the specific closed models you would otherwise pay for. The headline is a reason to run that evaluation, not a substitute for it.
That is the recurring lesson of a fast-moving, multi-model frontier: the right model is the one that wins on your workload, and the identity of that model keeps changing. An open-weight challenger that undercuts a closed leader by a wide margin is exactly the kind of development that rewards keeping the model layer swappable rather than hardwired. A model-agnostic approach, the design principle behind platforms like Metir AI, lets a new release like this one be tested and adopted where it wins without re-plumbing everything around it. On a frontier this crowded, the ability to switch is worth more than commitment to any single model, closed or open.
Sources:
- DeepSeek V4 Pro 0813 Goes GA: Benchmark Claims Await Independent Proof | Tech Times
- DeepSeek V4 Pro 0813 (max) - Intelligence, Performance & Price Analysis | Artificial Analysis
- DeepSeek prices its new V4-Pro-0813 model at $0.87 per 1M output tokens | Wccftech
- Chinese AI startup DeepSeek quietly released DeepSeek V4 Pro 0813 | Gigazine
Image credits
Header image: the skyline of Hangzhou, China, seen across West Lake. Hangzhou is DeepSeek's home city. By Windmemories via Wikimedia Commons, licensed under CC BY-SA 4.0.