DeepSeek V4 Release: Specs, Benchmarks, Peak Pricing

The most-used Chinese open-weight model of the year has finally left the lab. On Friday, July 24, 2026, DeepSeek moved its V4 model family from a months-long preview to general availability, retiring the legacy API aliases that cautious enterprises had been waiting on and switching on a new peak-hour pricing scheme. The graduation does not introduce a new model name, but it converts a fast-moving preview build into a stable, production-grade target — the thing large customers need before they will route real workloads through it.

DeepSeek first shipped V4 Preview on April 24, 2026, publishing two open-weight mixture-of-experts models under the permissive MIT license. Friday’s release keeps that same two-model lineup while adding infrastructure optimizations, a firmer service commitment, and the pricing changes the company had signaled since late June.

Two models, one architecture

The V4 family is built on a mixture-of-experts (MoE) design, in which only a fraction of the network’s parameters activate for any given token. That approach lets DeepSeek advertise very large total parameter counts while keeping the per-token compute cost — and therefore the price — comparatively low.

• DeepSeek V4-Pro carries 1.6 trillion total parameters with roughly 49 billion active per token.
• DeepSeek V4-Flash is the lighter variant at 284 billion total parameters and about 13 billion active.

Both models ship with a default 1-million-token context window, putting them in the same long-context tier as the frontier systems from U.S. labs and China’s other leading houses. If you want a refresher on the technique underpinning the lineup, our explainer on mixture of experts walks through why sparse activation has become the default recipe for frontier-scale models.

The benchmark that matters

The headline number is coding. DeepSeek reports that the top configuration, V4-Pro-Max, scores 80.6% on SWE-bench Verified, a widely watched test of a model’s ability to resolve real software-engineering issues. That figure is the strongest posted by any openly licensed model to date, and it lands level with Gemini 3.1 Pro — a closed, proprietary system. For a model whose weights anyone can download and run, matching a flagship on a hard agentic benchmark is the sort of result that resets expectations about what “open” can deliver.

The competitive pressure is not only about capability; it is about cost. On a per-output-token basis, DeepSeek claims V4-Pro is roughly 28.7 times cheaper than Claude Opus 4.8 and 34.5 times cheaper than GPT-5.5. Even allowing for the usual caveats around vendor-supplied comparisons, the gap is wide enough to change procurement math for teams running high-volume inference.

The new pricing — including a peak-hour surcharge

DeepSeek published API pricing that undercuts most Western frontier providers by a large margin:

• V4-Pro: $0.43 per million input tokens and $0.87 per million output tokens.
• V4-Flash: $0.14 per million input tokens and $0.28 per million output tokens.

The genuinely new wrinkle at general availability is peak-hour pricing. DeepSeek warned that it expects heavy demand at launch and will charge 2× the baseline rate during defined Beijing business windows, easing back to standard rates outside those hours. It is an unusual move for a frontier API — closer to how cloud providers price spot capacity or how utilities price electricity — and it signals that DeepSeek is managing a real capacity constraint rather than pricing purely to win share.

Alias retirement forces a migration

The release also carries a hard deadline for existing integrations. DeepSeek’s changelog confirms that the older deepseek-chat and deepseek-reasoner API aliases are fully retired on July 24, 2026, with traffic routing to deepseek-v4-flash in the interim. Teams that hard-coded the legacy names will need to point their clients at the explicit V4 model IDs. The forced cutover is the clearest sign that DeepSeek considers the preview era over: there is no longer a stable older model to fall back to.

Where V4 sits in the open-weight wave

DeepSeek’s general-availability launch arrives during the densest run of open-weight releases the industry has seen. It follows Moonshot AI’s Kimi K3, the 2.8-trillion-parameter model unveiled at the World Artificial Intelligence Conference, and it deepens a now-familiar pattern: Chinese labs, working under U.S. export controls that limit their access to the most advanced accelerators, are shipping frontier-class systems and giving the weights away.

That strategy has a market cost attached. Each major Chinese open-weight release this summer has coincided with fresh pressure on U.S. chip and AI equities, part of the broader chip-stock selloff as investors weigh how durable American labs’ lead really is. The through-line of our coverage on open-source models closing the gap is that the distance between the best closed and best open systems keeps shrinking — and V4’s SWE-bench result is one more data point on that curve.

DeepSeek itself has spent 2026 building out the rest of its stack, from a reported custom inference chip to persistent chatter about an eventual funding round and IPO. The V4 GA is the piece that turns all of that into a product businesses can actually commit to.

What it means

The move from preview to general availability is the whole story, even though nothing about the models’ weights changed. Enterprises do not deploy preview builds. By stabilizing V4, retiring the legacy aliases, and publishing a firm price sheet, DeepSeek is explicitly courting production workloads it could not win while the API was a moving target. The 80.6% SWE-bench figure gives buyers a capability justification; the sub-dollar output pricing gives them a cost justification. Together they make V4 a credible default for high-volume coding and agentic tasks where a 28-to-34× price gap against U.S. flagships compounds quickly.

The peak-hour surcharge is the most revealing detail. A lab confident in its supply does not meter demand by the clock. Charging double during Beijing work hours tells you DeepSeek is capacity-constrained at launch and would rather shape demand than degrade service — a constraint that likely traces back to the accelerator limits Chinese labs operate under. For customers, it introduces a planning wrinkle: batch and asynchronous jobs can be scheduled into cheap off-peak windows, while latency-sensitive traffic pays the premium. Expect sophisticated users to arbitrage the two.

For the competitive picture, V4’s release tightens the squeeze on Western providers at the low end of the price curve. It will not dislodge the top proprietary models from the frontier of raw capability, and enterprises with data-residency or provenance concerns will still hesitate over a Chinese-origin model. But for cost-sensitive teams that can self-host open weights, V4 makes the “good enough, far cheaper” option meaningfully better than it was a day ago. The watch items from here: whether V4-Pro’s benchmark strength holds up in independent evaluations, how quickly the peak-hour pricing pushes workloads onto self-hosted deployments of the MIT-licensed weights, and whether the next Chinese release — expected within days — pushes the open-weight frontier again before Western labs answer.