On August 12, 2026, DeepSeek quietly promoted its flagship model to general availability. No press release, no launch event. The API documentation updated to list DeepSeek-V4-Pro-0813 as the production build behind the deepseek-v4-pro endpoint, and OpenRouter’s model page reflected the change within hours. Four months after its April preview debut, the largest open-weight model in production is now officially live.
For enterprise AI teams, the timing is deliberate and the implications are immediate: V4 Pro’s benchmark scores jumped sharply from preview to GA, its pricing undercuts every US frontier model by a wide margin, and the company has warned that a “significant price increase” is coming soon. The window to assess and adopt at current rates is open now, and probably not for long.
What V4 Pro 0813 Is
DeepSeek V4 Pro is a mixture-of-experts (MoE) architecture with 1.6 trillion total parameters and 49 billion active per forward pass. Those 49 billion active parameters place it in the same ballpark as a dense 50B model for compute purposes, while the full 1.6T parameter space allows for deep specialist routing across domains. The model pre-trained on more than 32 trillion tokens, with post-training that develops domain-specific experts independently before consolidating them through on-policy distillation.
The architectural highlight is what DeepSeek calls Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA): attention variants that cut single-token inference compute to 27 percent and KV cache memory to 10 percent of what the prior-generation V3.2 model required at the 1-million-token context setting. The result is that running V4 Pro at long context is materially cheaper than running a comparable dense model at the same length.
The API surface is broad: OpenAI ChatCompletions, Anthropic Messages format, and DeepSeek’s own Responses API. Tool calling and structured JSON output work on both V4 Pro and V4 Flash. Three operating modes ship: non-thinking (fast, cost-efficient), high reasoning effort, and a maximum reasoning effort that DeepSeek calls V4-Pro-Max. The context window is 1 million tokens, with a maximum output of 384,000 tokens.
The Benchmark Numbers
These are vendor-reported figures from DeepSeek’s own harness runs. Independent evaluators have not yet replicated the 0813 build specifically, so treat them as the company’s claims until external verification arrives. With that caveat: the numbers are large.
| Benchmark | V4 Pro Preview | V4 Pro 0813 | Reference |
|---|---|---|---|
| Terminal-Bench 2.1 | 72.1% | 87.9% | GPT-5.6 Sol Max ~88% |
| CyberGym | 52.7% | 83.3% | Security and agentic tasks |
| DeepSWE | 12.8% | 62.7% | Repo-level software engineering |
| SWE-bench Verified | (preview) | 80.6% | Claude Opus 4.6 at 80.8% |
| LiveCodeBench | (preview) | 93.5% pass@1 | Top of published leaderboard |
| GPQA Diamond | (preview) | 90.1% | Graduate-level science reasoning |
| Humanity’s Last Exam | (preview) | 37.7% | Gemini-3.1-Pro at 44.4% |
The DeepSWE jump (12.8% to 62.7%) is the most significant for enterprise agentic workloads. DeepSWE evaluates repository-level software engineering: multi-file edits, test execution, build verification. That kind of improvement moves V4 Pro from “interesting option” to “serious competitor” for coding agents. The CyberGym jump matters for security-adjacent agentic use cases.
Where V4 Pro still trails: Humanity’s Last Exam (37.7% vs Gemini-3.1-Pro’s 44.4%) and Terminal-Bench 2.1 (87.9% vs GPT-5.4’s 75.1% at xHigh, though the comparison is version-specific and not apples-to-apples).
The Price Picture
This is where V4 Pro’s GA becomes a genuine market event.
API pricing is unchanged from the preview: $0.435 per million input tokens on a cache miss, $0.003625 per million on a cache hit, and $0.87 per million output tokens. On a blended input-output rate, that lands around $0.65 per million tokens.
Comparable pricing for US frontier models at comparable capability: Claude Fable 5 costs $10 per million input and $50 per million output, blending to roughly $30 per million tokens. That is a 46x price gap on raw rates.
On a per-task basis, the gap compounds further. Artificial Analysis estimates the cost per benchmark task at around $0.03 for V4-Flash and $3.15 for Claude Fable 5. The per-task cost for V4 Pro 0813 is not yet published, but it sits between those. Hugging Face CEO Clement Delangue has noted that when full task economics are counted, the cost ratio can reach $0.04 for DeepSeek versus $31 for Fable 5, reflecting the longer inference chains the US frontier models tend to generate.
On performance, DeepSeek’s own benchmarks show Claude Fable 5 averaging 5.3 percent higher across 9 comparable agent evaluations where both models have published scores. Excluding the benchmarks with large outlier gaps, the remaining average difference narrows to approximately 2.8 percent. Neither number has been independently verified for the 0813 build against Fable 5.
The practical read: V4 Pro 0813 appears to sit within 5 percent of the best US frontier models on agentic benchmarks while costing 20 to 46 times less. For high-volume agentic workloads (coding agents, document pipelines, research automation) where the task economics matter more than peak capability, that is a significant shift in the calculus.
What Changed from Preview to GA
The preview-to-GA transition is not primarily a capability announcement. The April preview already shipped open weights and a live API. What the 0813 build adds:
Post-training improvements. The benchmark jumps above are the visible output of additional post-training runs. DeepSeek describes the GA post-training as targeting agentic, coding, and tool-calling workloads specifically.
Responses API support and Codex adaptation. Both are now live for V4 Pro, matching what V4 Flash gained on July 31. Enterprise teams using OpenAI-compatible tooling can now route to V4 Pro without API-level rewrites.
Production status. The preview label is gone. The endpoint now commits to the 0813 build as the stable version. Concurrency limits are 500 for Pro and 2,500 for Flash, with enterprise rate-limit increases available on request.
Open weights confirm MIT license. The weights that have been downloadable since April 24 are confirmed production-grade. Teams that need data residency can now self-host with confidence that the model is final.
The Price-Increase Warning
DeepSeek’s pricing page carries a notice that the company plans a “significant increase” in API pricing across its services, with specifics to come by official announcement. No timeline was given.
This is not unusual. DeepSeek has held preview-period pricing through the GA launch of both Flash (July 31) and Pro (August 12). Current rates reflect an early-adopter posture that is explicitly temporary.
For enterprise teams evaluating V4 Pro, this creates a practical window. Teams that want to benchmark V4 Pro against their current models, evaluate self-hosting economics, or lock in API usage patterns are best positioned to do so now, before pricing revises upward.
A related development: DeepSeek’s internal agentic coding is already running on V4 series models. The company’s own developer velocity is a live proof-of-concept for the cost and capability claims, though it is also the strongest argument for taking the vendor-risk questions seriously.
What Enterprise AI Teams Should Act On
Benchmark before the price changes. Run your highest-cost agentic workloads (large coding agents, document-heavy pipelines, multi-step research automation) against V4 Pro 0813 at current rates. The cost difference is large enough that even a workload that tolerates 10 percent quality degradation below Fable 5 may save 40 times the token budget.
Evaluate self-hosting for regulated workloads. The MIT license and open weights remove the API vendor-risk concern for teams that can run inference on their own infrastructure. The compute requirement is significant (FP8 hardware at scale), but the managed inference ecosystem (Together AI, Fireworks AI, and others) offers a middle path.
Treat vendor risk as a real constraint. DeepSeek is a PRC-jurisdiction AI lab. For workloads involving sensitive data, regulated content, or operations in jurisdictions with AI-origin restrictions, the open weights path is the safer route than the API. Legal review is not optional.
Watch for the pricing notice. When DeepSeek publishes its revised pricing, re-run the workload economics. If the increase is modest, V4 Pro may still lead on cost. If it narrows the gap materially, the decision calculus shifts.
For a deeper look at how the open-weight cost race is reshaping enterprise AI model economics, see our analysis of DeepSeek V4 Flash and the OpenAI Luna price cut. For context on how enterprises are rethinking AI model selection under pricing pressure, our enterprise AI pricing revolt analysis covers the structural forces at play. And for a comparison with the other major open-weight frontier model in this weight class, see our coverage of Kimi K3 from Moonshot AI.
Sources: Unite.AI, Digital Today, WebProNews, DeepSeek API documentation, Hugging Face model card.