At 11 PM on August 13, DeepSeek quietly updated its API documentation. The model identifier changed to DeepSeek-V4-Pro-0813. No launch event. No social media teaser. No press release. By morning, the update had 692 upvotes on Hacker News and 246 comments — because the quiet part was loud: DeepSeek’s flagship model had left preview, and the pricing structure that made Chinese AI APIs the cheapest in the world is being dismantled.
What is DeepSeek V4 Pro? DeepSeek’s flagship model — a mixture-of-experts system with 1.6 trillion total parameters and 49 billion active per token. It went into preview in April 2026 with open weights under the MIT licence. The smaller Flash variant graduated to general availability first, on July 31. V4 Pro 0813 is the flagship’s turn, and it arrives with serious benchmark improvements and a new pricing structure that signals the end of the loss-leader era.
The Benchmark Leap
The headline number is DeepSWE: 62.7, up from 12.8 in preview — a five-fold jump. On Terminal Bench 2.1, it climbed from 72.1 to 87.9, roughly matching GLM-5.3 at 88.2 and Fable 5 at 88.0. On CyberGym, V4 Pro 0813 scored 83.3, edging past Fable 5’s 83.1. SWE-bench Verified landed at 80.6, level with Gemini-3.1-Pro and a fraction behind Claude Opus 4.6 at 80.8.
As Unite.AI reports, these are vendor-reported numbers from DeepSeek’s own harness runs. No independent evaluator has replicated them for the 0813 build yet. But the magnitude of the DeepSWE jump — from 12.8 to 62.7 — is hard to fake. That benchmark measures multi-step, long-horizon repository code repairs, the kind of task that separates coding assistants from coding agents.
The model supports a one-million-token context window with a 384,000-token maximum output. That’s enough to index a large enterprise codebase in a single call. It also ships with Anthropic API compatibility, pre-fill prompt continuation, and Fill-in-the-Middle code completion — features that were missing in preview.
The Pricing Tier
Here’s where it gets interesting. V4 Pro 0813 costs $0.435 per million input tokens (cache miss) and $0.87 per million output tokens. That’s roughly 3x the Flash variant’s rates of $0.14 and $0.28. The concurrency cap is 500 for Pro versus 2,500 for Flash.
According to SteinsLab, DeepSeek also introduced peak and off-peak pricing. Off-peak rates are 50 per cent lower than peak. The new pricing takes effect at 16:00 UTC on August 16, 2026.
And then there’s the footnote. DeepSeek’s pricing page explicitly states the company plans “a significant increase” in overall API pricing in the near future. The era of Chinese AI labs subsidising API access at a loss to build market share is drawing to a close.
The End of the Price War
For the past 18 months, DeepSeek has been the price disruptor of the AI industry. Their permanent 75 per cent discount on flagship models forced every western lab to justify their pricing. OpenAI cut API rates. Anthropic introduced cache-hit pricing. Google slashed Gemini API costs. The cascade was real, and DeepSeek was the catalyst.
Now DeepSeek is telling its customers: the subsidy is ending. The tiered structure — Flash for high-volume, low-cost workloads, Pro for high-value complex reasoning at 3x the price — is a commercial segmentation strategy, not a price war weapon. The upcoming “significant increase” confirms it.
This isn’t surprising. R&D costs and compute consumption have been mounting. The April preview of V4 was open-weight under MIT, which means no exclusivity premium. The loss-leader strategy worked — DeepSeek’s Hugging Face repositories logged more than 1.4 million downloads in the last month alone. But at some point, the bill comes due. That point is now.
What Ships Next
The open question is weights. The Hugging Face repositories still host the April preview builds. DeepSeek has not announced a timeline for publishing 0813 weights, nor said whether the GA build differs from preview beyond post-training. The company’s stated cadence for the V4 line runs through the API first.
The self-developed agent framework “DeepSeek Harness” — which the community has been anticipating — was not launched alongside the GA model. Official documentation still recommends third-party agent hosts like OpenAI Codex, Claude Code, or OpenCode. For enterprise developers, the absence of a bundled harness means building custom context-truncation and state-management pipelines on their own.
What Stands Out
What’s notable here is not the benchmarks — those are expected, and vendor-reported until proven otherwise. What’s notable is the commercial signal. DeepSeek was the lab that proved you could build frontier-class models at a fraction of western compute costs and give the API away cheap. Now they’re telling you the cheap part is over.
The tiered pricing, the concurrency caps, the explicit warning of price increases — this is a company transitioning from market-share acquisition to margin management. The same transition that every successful tech company makes. The difference is that DeepSeek’s price war already forced the entire industry to adjust. Western labs got leaner. Chinese labs got faster. The consumer won.
The question now is whether the consumer keeps winning when the cheapest provider in the market starts raising prices.
❓ FAQ
What changed between V4 Pro preview and GA? The 0813 build shows major jumps in agentic benchmarks — DeepSWE went from 12.8 to 62.7, Terminal Bench 2.1 from 72.1 to 87.9, CyberGym from 52.7 to 83.3. DeepSeek also added Anthropic API compatibility, pre-fill continuation, and Fill-in-the-Middle code completion. Whether the underlying weights differ from preview has not been disclosed.
How much does V4 Pro cost now? $0.435 per million input tokens (cache miss) and $0.87 per million output tokens. Cache-hit input is $0.003625 per million. Peak and off-peak pricing takes effect August 16, with off-peak rates 50 per cent lower. A “significant increase” in overall pricing is coming, per DeepSeek’s pricing page.
Are the 0813 weights open source? Not yet. The April preview weights are on Hugging Face under MIT licence with 1.4 million downloads. DeepSeek has not announced when 0813 weights will be published.
How does V4 Pro compare to GLM-5.3? Both shipped the same week. On Terminal Bench 2.1, they’re close (87.9 vs 88.2). On CyberGym, V4 Pro edges Fable 5 (83.3 vs 83.1) while GLM-5.3 leads both (84.5). GLM-5.3’s story is emergent cyber capability from post-training; V4 Pro’s is the agentic leap from preview to GA. Different angles, same week — the Chinese open-weight cadence is now matching or exceeding western release velocity.
Is the cheap API era really over? For flagship models, the signs point to yes. DeepSeek’s tiered pricing, concurrency caps, and explicit price-hike warning are clear signals. The smaller Flash model remains cheap, but the loss-leader strategy on flagships is winding down. This affects every developer who built infrastructure on the assumption that Chinese API prices would stay at rock bottom.
