DeepSeek shipped the first pieces of its long-delayed V4 model family on April 24, 2026, and the numbers attached to it did more to rattle the AI industry than the model's actual capabilities did. The Chinese lab released two open-weight systems, V4-Pro and V4-Flash, that land within fractions of a point of top-tier closed models from Anthropic, OpenAI, and Google on real-world coding benchmarks. Then it priced them at a level that made those comparisons almost beside the point.
V4-Pro, the flagship, is a 1.6 trillion-parameter mixture-of-experts model that activates 49 billion parameters per token. V4-Flash, the smaller sibling, runs 284 billion total parameters with 13 billion active. Both default to a 1 million-token context window with up to 384,000 tokens of output, and both were released under the permissive MIT license with open weights posted to Hugging Face — meaning any developer or company can download, modify, and run them without paying DeepSeek anything at all.
What the model can actually do
On SWE-bench Verified, the industry's standard test for whether a model can resolve real GitHub issues, V4-Pro scored 80.6%. Depending on which point release of Anthropic's Claude Opus line it's measured against, that figure lands within roughly a fifth of a point of the closed-source leader — a gap close enough that DeepSeek's own technical report and multiple third-party benchmark trackers describe the two as practically indistinguishable in production use. V4-Pro also posted a Codeforces competitive-programming rating of 3,206 in its maximum reasoning mode — the highest score any AI model had recorded on the platform at the time of release, and one that several trackers estimate would place it around 23rd among human competitors, in the low Grandmaster band.
The model is not uniformly ahead. DeepSeek's own release notes acknowledge V4 falls short of GPT-5.4 and Gemini 3.1 Pro on general world-knowledge benchmarks by an estimated three to six months' worth of model progress, and it trails on the hardest mathematics competitions, including the HMMT 2026 test. Where it consistently leads is agentic tool use: V4-Pro topped the MCPAtlas benchmark for multi-step agent workflows and posted the highest score of any open-weight model on LiveCodeBench. V4-Flash, despite its smaller size, scored 79.0% on SWE-bench Verified — just 1.6 points behind its larger sibling, which several reviewers have flagged as the more interesting data point than the Pro flagship's numbers themselves, since it means most production coding workloads may not need the expensive model at all.
The pricing is the real story
V4-Pro launched at $1.74 per million input tokens and $3.48 per million output tokens, already a steep discount against Western frontier pricing. It did not stay there. On May 22, DeepSeek announced that a promotional 75% discount, originally set to expire May 31, would instead become permanent. That dropped the standing rate to $0.435 per million input tokens and $0.87 per million output tokens, with cached input priced at $0.003625 per million tokens — a rate low enough that agent pipelines reusing the same system prompt or retrieved context pay a small fraction of standard cost on repeated calls.
The comparison to incumbent pricing is stark. Anthropic's Claude Opus models and OpenAI's GPT-5.5 have both priced output tokens in the range of $25 to $30 per million. At V4-Pro's permanent rate, that puts DeepSeek's flagship model somewhere between seven and nine times cheaper per output token than its closest Western competitors, and industry analyses have pegged a 100 million-token monthly agent workload at roughly $87 on V4-Pro against $2,500 or more on Opus or GPT-5.5. V4-Flash pushes the gap further still: at $0.28 per million output tokens, running the same 10 million-token workload that costs $250 on a comparable closed model costs under $3 on Flash.
DeepSeek escalated again on July 31, releasing an officially finalized build of V4-Flash — internally labeled V4-Flash-0731 — that kept the same architecture, parameter count, and pricing as the April preview but added a re-post-training pass focused on agentic performance, along with native support for OpenAI's Responses API and explicit compatibility with Codex. The release landed within roughly two days of OpenAI cutting prices on its own GPT-5.6 Luna model, and multiple industry trackers characterized DeepSeek's move as a direct countermove aimed at keeping its output pricing as the effective floor for enterprise agentic workloads. The gambit worked well enough to become a problem: usage data from the OpenRouter marketplace and developer reports both point to demand for V4-Flash outstripping what DeepSeek's roughly 20,000-GPU compute footprint could comfortably serve, with users describing inference slowdowns at peak hours in the weeks after launch.
The engineering behind the price cut
The pricing isn't purely a business decision — DeepSeek's technical documentation attributes much of the cost reduction to architectural changes rather than simply eating a lower margin. V4 introduces what the company calls Compressed Sparse Attention paired with Heavily Compressed Attention, along with a token-wise compression scheme layered on top of DeepSeek Sparse Attention, its successor to the attention mechanism used in the V3 generation. The stated goal is to cut the per-token compute cost of very long context windows, and DeepSeek has said the approach runs at roughly a quarter of the single-token compute and a tenth of the memory footprint of its predecessor once context length stretches into the hundreds of thousands of tokens. That matters because both V4-Pro and V4-Flash default to a 1 million-token context window rather than treating long context as a premium add-on, which is unusual among frontier-class models and is part of why DeepSeek is marketing V4 as built specifically for long-running agentic workflows rather than single-turn chat.
The training run itself also reportedly used a custom optimizer called Muon in place of more conventional approaches, though DeepSeek has released fewer specifics on that piece than on the attention architecture. Both models ship in dual modes — a "thinking" mode for extended reasoning and a faster non-thinking mode — and the API is compatible with both the OpenAI ChatCompletions format and Anthropic's API format, meaning developers can often swap DeepSeek in as a drop-in replacement in existing tooling, including agent frameworks like Claude Code and OpenCode, without rewriting integration code.
Where the industry stands now
The reaction from competitors has been pricing moves of their own rather than public benchmarking disputes. Analysts covering the enterprise AI market have said the durability of DeepSeek's cuts — positioned explicitly as permanent rather than promotional — is pushing Western labs toward reconsidering consumption-based pricing altogether in favor of outcome- or value-based billing models that are harder for a cheaper competitor to undercut directly. Nvidia and cloud infrastructure providers that built forecasts around inference remaining expensive through the back half of the decade are facing the same repricing pressure secondhand.
V4-Pro's own path to general release ran on a longer fuse than Flash's, and it closed only this week. The Pro endpoint on DeepSeek's API began pointing to a new build, designated V4-Pro-0813, on August 12 — ending a preview period that had run since the April 24 debut, with no public blog post or changelog entry marking the switch. The change surfaced first through DeepSeek's own API documentation and pricing pages, and was confirmed independently when OpenRouter listed the build as the official general-availability release the same day. DeepSeek later described the release publicly as focused on agent capabilities, adding native support for OpenAI's Responses API format and Codex integration — the same upgrades Flash received in July.
DeepSeek's own benchmark figures for the 0813 build show large jumps on agentic tests compared with the April preview: Terminal-Bench 2.1 rose from 72.1 to 87.9, DeepSWE from 12.8 to 62.7, and CyberGym from 52.7 to 83.3, according to the company. As of this writing, no independent evaluator has replicated those numbers, which is worth noting given that the earlier preview-era benchmarks discussed above did hold up under third-party scrutiny over the following months. Pricing on the new build carries over unchanged from preview — $0.435 per million input tokens, $0.87 per million output, $0.003625 on a cache hit — but DeepSeek has told developers that will not last. A broader price increase across the V4 family takes effect August 16 at 16:00 UTC, alongside a new peak- and off-peak billing structure that charges double during Beijing business hours and half price outside them. DeepSeek has not published the new baseline rates.
The legacy deepseek-chat and deepseek-reasoner API aliases, which had pointed to the company's V3-generation models, were fully retired on July 24, 2026, forcing any remaining holdout integrations onto the V4 naming scheme regardless of whether developers had migrated deliberately.
What DeepSeek has established, even before this week's benchmark claims are independently verified, is a pricing floor the rest of the industry has spent since April building around — and, now, a demand curve strong enough that the company is already walking part of that pricing back.
