DeepSeek just changed the terms of its own bargain. On August 13, 2026, the Chinese AI lab pushed out DeepSeek V4 Pro 0813, an updated reasoning model that scored 53 on the Artificial Analysis Intelligence Index, according to the benchmarking firm. The catch: Reuters reported the model now costs $1.32 per million input tokens and $3.96 per million output tokens, up from $0.435 and $0.87 for the April 2026 version of V4 Pro. That is a 3.6x jump in list price for an 8-point benchmark gain, and it lands weeks after DeepSeek shipped its first native multimodal model, DeepSeek-V4-Flash-Vision-Exp, as a 305-billion-parameter open weights release under the MIT license.
Together, the two launches mark a shift in how DeepSeek is positioning itself. The company built its reputation in early 2025 on undercutting Western labs on price. Eighteen months later, its flagship reasoning tier is priced closer to a frontier product, while its open-weight releases keep expanding into new formats, including images. For engineers who route production traffic through DeepSeek’s API, and for anyone tracking the open-weight side of the AI model race, both moves carry real consequences.
What DeepSeek Shipped This Week
DeepSeek’s release cadence in 2026 has followed a pattern: point updates to the V4 family arrive every few weeks, each carrying a build tag tied to its release date. The DeepSeek V4 Flash line moved to the V4-Flash-0731 build on July 31, 2026, which DeepSeek described as the same architecture re-post-trained for agent-style tasks rather than a ground-up redesign. Two weeks later, on August 13, 2026, the reasoning-focused DeepSeek V4 Pro 0813 build landed, followed on August 30-31, 2026 by DeepSeek-V4-Flash-Vision-Exp, a multimodal model built on the V4-Flash base with a vision encoder added.
That vision model is the more unusual release of the two. Independent technical blog Cellcog reported that “DeepSeek published open weights for V4-Flash-Vision-Exp on Hugging Face on August 31, 2026, under the MIT license: its first multimodal model in the DeepSeek-V4 family” (Cellcog). Technology outlet Pandaily separately described it as DeepSeek’s first native multimodal model, distinguishing it from earlier DeepSeek systems that relied on separate vision components bolted onto a text-only base. DeepSeek shipped the model with serving recipes for both vLLM and SGLang, the two most widely used open-source inference engines for large language models, which lowers the bar for teams that want to self-host rather than call DeepSeek’s hosted API.
DeepSeek V4 Pro 0813: The Benchmark Numbers
Artificial Analysis, the benchmarking firm whose Intelligence Index has become a common reference point for model comparisons, published its read on the new reasoning build in a post on X. According to the firm, “DeepSeek V4 Pro 0813 scores 53 on the Artificial Analysis Intelligence Index, 8 points above April’s DeepSeek V4 Pro – but with a 3.6x price increase and only 1 point above DeepSeek V4 Flash 0731” (Artificial Analysis). That framing matters. An 8-point jump on a composite reasoning benchmark is a meaningful improvement in raw capability. A one-point edge over DeepSeek’s own cheaper flash-tier model, at more than three times the price, is a much harder sell for cost-conscious teams.
Reuters, reporting separately on the release, put DeepSeek V4 Flash at 40 on the same index, well below both the new V4 Pro 0813 score and the V4 Flash 0731 figure Artificial Analysis referenced, suggesting DeepSeek is now running several flash-tier configurations in parallel with meaningfully different capability profiles. Other benchmark trackers cited in the days after the release placed DeepSeek V4 Pro 0813 four points behind GPT-5.6 Terra and seven points behind Kimi K3, positioning the new build in the upper-middle of the current leaderboard rather than at the very top.
The Pricing Shock: API Costs Jump 3.6x
The pricing change is the part of this story that will hit engineering budgets first. DeepSeek built its early reputation on undercutting OpenAI and Anthropic by an order of magnitude on a per-token basis. The August 2026 reasoning tier narrows that gap substantially. Reuters reported that, depending on which earlier DeepSeek tier is used as the baseline, the new pricing runs as much as 14 times higher than some of the company’s older offerings.
The table below lays out the four current DeepSeek V4 configurations side by side, using the pricing figures reported by Reuters and DeepSeek’s own published rate cards as cited in industry coverage.
| Model | Release Date | Input ($/M tokens) | Output ($/M tokens) | Notes |
|---|---|---|---|---|
| DeepSeek V4 Pro (original) | April 2026 | $0.435 | $0.87 | Cached input $0.003625/M tokens |
| DeepSeek V4 Flash 0731 | July 31, 2026 | $0.14 | $0.28 | Cache-hit input $0.003/M tokens |
| DeepSeek V4 Pro 0813 (reasoning) | August 13, 2026 | $1.32 | $3.96 | 3.6x the April V4 Pro price, per Artificial Analysis |
| DeepSeek-V4-Flash-Vision-Exp | August 30-31, 2026 | Open weights (MIT license) | Open weights (MIT license) | Self-hosted via vLLM or SGLang, no DeepSeek-hosted rate published |
For teams that built pricing assumptions into a product roadmap around DeepSeek’s original cost advantage, the jump from roughly $0.87 per million output tokens to $3.96 is the kind of change that forces a re-evaluation of vendor mix, especially for high-volume reasoning workloads like agentic coding or multi-step planning where output tokens accumulate quickly.
V4-Flash-Vision-Exp: DeepSeek’s First Native Multimodal Model
Where the V4 Pro 0813 update raises questions about DeepSeek’s pricing discipline, DeepSeek-V4-Flash-Vision-Exp reinforces the company’s open-weight strategy. Cellcog and Pandaily both put the parameter count at roughly 305 billion, with one technical source citing an exact figure of 304,646,824,126 parameters, built on top of the existing V4-Flash architecture with a vision encoder layered in for image understanding. That is a large model to release under a permissive license: MIT terms allow commercial reuse, modification, and redistribution with minimal restriction, in contrast to the custom or noncommercial licenses several other open-weight labs attach to their largest releases.
DeepSeek published the weights to Hugging Face alongside deployment recipes for vLLM and SGLang, the inference frameworks most commonly used to serve open models at scale. That combination, large parameter count plus permissive license plus ready-made serving instructions, is designed to make V4-Flash-Vision-Exp usable by teams that want to run a multimodal model on their own infrastructure rather than send image data to a third-party API, a distinction that matters for companies with data residency or compliance constraints.
How V4 Pro 0813 Stacks Up Against Rival Models
The Artificial Analysis Intelligence Index has become one of the more widely cited composite benchmarks for comparing frontier and near-frontier models, blending performance across reasoning, coding, and knowledge tasks into a single score. Industry coverage citing the index in early September 2026 placed a range of current models on the same scale, giving a rough picture of where DeepSeek’s new reasoning build sits relative to the field.
| Model | Artificial Analysis Intelligence Index Score |
|---|---|
| Claude Opus 5 | 63 |
| Claude Fable 5 | 62 |
| GPT-5.6 Sol | 61 |
| Grok 4.6 | 61 |
| Kimi K3 | 60 |
| DeepSeek V4 Pro 0813 | 53 |
| GLM-5.2 | 53 |
| Gemini 3.6 Flash | 52 |
| DeepSeek V4 Flash | 40 |
Scores in this table are compiled from Artificial Analysis Intelligence Index figures as cited in press and industry coverage, not from a single primary source table, so treat exact rankings as directional rather than official. Even with that caveat, the pattern is clear: DeepSeek V4 Pro 0813 lands in a tight cluster with GLM-5.2 and just ahead of Gemini 3.6 Flash, several points behind Kimi K3, GPT-5.6 Sol, Grok 4.6, and the Claude 5 family. DeepSeek has closed part of the gap to frontier Western and rival Chinese models, but it has not closed all of it, and it did so while giving up much of its price advantage.
Historical Context: The January 2025 R1 Shock
To understand why any DeepSeek release still draws this much attention, it helps to go back to January 27, 2025. That day, DeepSeek’s R1 model, trained at a reported fraction of the cost of comparable U.S. systems, triggered a sudden reassessment of AI infrastructure spending among investors. Nvidia’s stock fell 16.9% in a single session, closing at $118.58 after a Friday close of $142.62, according to TechCrunch. CNBC and Reuters both reported the drop wiped out close to $600 billion in Nvidia’s market capitalization in one day, at the time the largest single-day market cap loss in U.S. corporate history. The episode became shorthand for a broader question: if a lower-cost, open-weight model could approach frontier performance, did the AI industry’s enormous capital spending on GPUs still make sense.
That history is the backdrop against which every subsequent DeepSeek release gets read. It is also why the August 2026 pricing increase is notable in the opposite direction: the company that once symbolized cheap frontier-adjacent AI is now charging closer to what its rivals charge for comparable reasoning performance.
Why the August 2026 Launch Didn’t Repeat the Selloff
Despite the historical weight attached to DeepSeek news, the August 2026 releases did not produce anything resembling the January 2025 market shock. Reuters’ coverage of V4 Pro 0813 focused on pricing and benchmark performance rather than framing the release as a threat to the AI infrastructure buildout, and no comparable single-day stock move has been reported in connection with either the V4 Pro 0813 or V4-Flash-Vision-Exp launches. The more defensible read is that the market has already priced in DeepSeek as a recurring, iterative competitor rather than a one-time shock. Investors reacted once, in January 2025, to the idea that cheap, capable open-weight models were possible at all. A model that gets modestly better while getting substantially more expensive does not carry the same disruptive charge.
DeepSeek’s Business Numbers: Funding, Valuation, and Revenue
DeepSeek’s commercial position has grown alongside its model releases. Research firm Sacra reported that DeepSeek closed a funding round of roughly RMB 50 billion, about $7.4 billion, in June 2026, with a post-money valuation near RMB 350.88 billion, or roughly $51.82 billion, based on a July 2026 regulatory filing (Sacra). Sacra also reported that DeepSeek reached an annualized revenue run-rate of about $500 million in August 2026, up sharply from roughly $12 million at the end of 2025.
Reported figures on realized, as opposed to annualized, revenue are more conservative. Cryptopolitan reported on August 26, 2026 that DeepSeek generated 475 million yuan, about $70.7 million, in recognized revenue across the first seven months of 2026, a figure Sacra also cited as roughly 10 times the company’s estimated full-year 2025 revenue of about $7 million. The gap between a $500 million annualized run-rate and $70.7 million in year-to-date recognized revenue reflects how quickly DeepSeek’s paid usage has been accelerating through the middle of 2026, rather than a contradiction between sources.
Market Impact: What This Means for the Open-Weight AI Race
DeepSeek’s dual move, a pricier reasoning tier alongside a large, permissively licensed multimodal release, illustrates a split strategy that is becoming common among open-weight labs in 2026. Frontier-adjacent reasoning capability now commands frontier-adjacent pricing, even from labs that built their name on discount economics. At the same time, giving away a 305-billion-parameter multimodal model under an MIT license keeps DeepSeek relevant to the self-hosting and fine-tuning community that a purely API-priced strategy would not reach.
The broader model market has also gotten more crowded and more fragmented. Benchmark aggregator BenchLM reported tracking 413 individual benchmarks across 10 categories for 415 distinct AI models as of September 2, 2026, with roughly 136 new model releases logged over the trailing 12 months, close to one release every three days. Of the top 50 models on BenchLM’s own leaderboard, 12 were open-weight as of that date. That density makes any single release, including DeepSeek’s, a smaller part of a much larger and faster-moving picture than it would have been in 2025.
Competitive Comparison: DeepSeek vs Qwen, Llama, and the Open-Weight Field
DeepSeek no longer has the open-weight lane to itself. Alibaba’s Qwen family, Zhipu’s GLM series, Moonshot AI’s Kimi models, and Meta’s Llama line all compete for the same developer attention, and each has taken a different approach to the cost-versus-capability tradeoff DeepSeek is now navigating. GLM-5.2 landed at 53 on the Artificial Analysis Intelligence Index, tied with DeepSeek V4 Pro 0813, while Kimi K3 scored meaningfully higher at 60, per the same round of industry citations. Meta’s Llama 3.1 family, still widely deployed as a self-hosted baseline across enterprise infrastructure, continues to post strong scores on public benchmark documentation for its 405B, 70B, and 8B variants, with the 405B model previously shown to match or beat several 2024-era frontier systems on a majority of public benchmarks according to analysis from DeepLearning.AI.
What differentiates DeepSeek’s current move is the combination of raising reasoning-tier prices while simultaneously shipping one of the largest permissively licensed multimodal models available for self-hosting. Most rivals have picked one lane or the other: either compete on hosted-API price, as DeepSeek did in 2025, or compete on open-weight generosity, as Meta has largely done with Llama. DeepSeek is now trying to do both at once, charging more for its hosted reasoning tier while giving away a bigger open model than most competitors offer.
Expert and Industry Reactions
Benchmarking firm Artificial Analysis was the most direct named voice assessing the release, framing the update in terms of capability-per-dollar rather than raw score. As the firm put it, DeepSeek V4 Pro 0813 “scores 53 on the Artificial Analysis Intelligence Index, 8 points above April’s DeepSeek V4 Pro – but with a 3.6x price increase and only 1 point above DeepSeek V4 Flash 0731” (Artificial Analysis). That is an unusually restrained assessment for a model release, and it reads as a deliberate signal to developers that the benchmark gain does not, on its own, justify the pricing change for every use case.
Coverage from Cellcog framed the V4-Flash-Vision-Exp release in more straightforwardly positive terms, noting DeepSeek published open weights for the model on Hugging Face under the MIT license as its first multimodal entry in the V4 family (Cellcog). The two reactions together sum up the split reception: enthusiasm for the open multimodal release, and more measured scrutiny of the reasoning tier’s new price.
What Comes Next: Predictions Through Early 2027
A handful of trends look likely to play out over the next several months based on the pattern DeepSeek has set with this release cycle.
- DeepSeek’s reasoning-tier pricing will likely keep drifting toward parity with OpenAI, Anthropic, and Google as the company chases higher benchmark scores, narrowing the cost gap that first made it famous.
- Expect DeepSeek to keep its flash-tier and open-weight releases aggressively priced or free to retain the developer and self-hosting audience that reasoning-tier price hikes risk alienating.
- Given DeepSeek’s reported $500 million annualized revenue run-rate and roughly $51.82 billion valuation, further funding activity or renewed IPO speculation into 2027 is plausible.
- Rival Chinese labs, including Zhipu and Moonshot AI, are likely to respond with their own large, permissively licensed multimodal releases to avoid ceding ground to DeepSeek’s open-weight strategy.
- With more than 400 models now tracked across public leaderboards, expect enterprise buyers to rely less on any single benchmark score and more on task-specific evaluation before choosing a model provider.
What Developers Should Watch For
Teams currently building on DeepSeek’s API should treat the V4 Pro 0813 pricing change as a prompt to re-audit which workloads actually need reasoning-tier performance. Output-token-heavy tasks, including long agentic chains and multi-step planning, will feel the roughly 4.5x jump in output pricing most acutely. For workloads that do not require the full reasoning tier, DeepSeek V4 Flash 0731 remains priced closely in line with the company’s original discount positioning, at $0.14 per million input tokens and $0.28 per million output tokens.
Teams evaluating DeepSeek-V4-Flash-Vision-Exp for self-hosted deployment should note that, at roughly 305 billion parameters, the model requires substantial GPU memory to serve even with the vLLM and SGLang recipes DeepSeek provided, and organizations without existing large-model serving infrastructure may find a hosted multimodal API more practical despite the licensing advantages of running the open weights directly.
Frequently Asked Questions
What is DeepSeek V4 Pro 0813?
It is an updated build of DeepSeek’s reasoning-focused V4 Pro model, released August 13, 2026, that scored 53 on the Artificial Analysis Intelligence Index, 8 points higher than the April 2026 version of V4 Pro, according to Artificial Analysis.
Why did DeepSeek’s API pricing increase so much?
Reuters reported the new V4 Pro 0813 reasoning tier costs $1.32 per million input tokens and $3.96 per million output tokens, versus $0.435 and $0.87 for the April 2026 version, a 3.6x increase Artificial Analysis attributed to the model’s improved benchmark performance.
What is DeepSeek-V4-Flash-Vision-Exp?
It is DeepSeek’s first native multimodal model in the V4 family, released August 30-31, 2026 as open weights on Hugging Face under the MIT license, with roughly 305 billion parameters and serving recipes for vLLM and SGLang.
Is DeepSeek-V4-Flash-Vision-Exp free to use?
The model weights are free to download and self-host under the MIT license, which permits commercial use. DeepSeek has not published a separate hosted-API rate for the vision model in the coverage reviewed for this article.
How does DeepSeek V4 Pro 0813 compare to GPT, Claude, and Gemini?
Based on Artificial Analysis Intelligence Index figures cited in industry coverage, DeepSeek V4 Pro 0813’s score of 53 trails Claude Opus 5 (63), Claude Fable 5 (62), GPT-5.6 Sol (61), Grok 4.6 (61), and Kimi K3 (60), while landing roughly even with GLM-5.2 and just ahead of Gemini 3.6 Flash (52).
Did the August 2026 DeepSeek release cause a stock market reaction like DeepSeek R1 did in January 2025?
No comparable market shock has been reported. The January 2025 DeepSeek R1 release triggered a reported $600 billion single-day loss in Nvidia’s market capitalization, per CNBC and TechCrunch, while coverage of the August 2026 releases has focused on pricing and benchmarks rather than a broad market reaction.
How much revenue does DeepSeek generate?
Research firm Sacra reported DeepSeek reached an annualized revenue run-rate of about $500 million in August 2026, while Cryptopolitan reported 475 million yuan, about $70.7 million, in recognized revenue across the first seven months of 2026.
What is DeepSeek’s current valuation?
Sacra reported a post-money valuation of roughly RMB 350.88 billion, about $51.82 billion, based on a July 2026 regulatory filing tied to a roughly RMB 50 billion, or $7.4 billion, funding round closed in June 2026.
