The short answer
On 13 August 2026 DeepSeek moved its V4 Pro model from preview to general availability, deploying the 0813 checkpoint across its app, web, and API platforms. The model is a 1.6-trillion-parameter Mixture-of-Experts system that activates 49 billion parameters per token, supports a one-million-token context window, and ranks third on the Artificial Analysis Intelligence Index across 106 models. MIT-licensed open weights allow teams to self-host without royalties. A significant price increase takes effect at 16:00 UTC on 16 August — output costs rise 4.5x from $0.87 to $3.96 per million tokens at peak hours.
For teams already running AI-assisted development workflows, V4 Pro 0813 is the most capable open-weight agentic coding model available today. The urgency is the pricing deadline: teams that want to run cost projections before committing need to act by end of 15 August.
What DeepSeek actually released
The GA release carries three substantive changes over the preview build. First, DeepSeek reports "significantly enhanced agent capabilities" — the model's ability to use tools, execute code autonomously, and complete multi-step workflows without human intervention. Second, thinking-effort levels have been expanded to three settings (low, high, max), allowing developers to tune the computation budget to task complexity rather than choosing between on/off. Third, the API now supports the OpenAI Responses API format out of the box, with built-in Codex integration.
That last point has practical consequences for teams already in production. Existing codebases that call the OpenAI SDK can redirect requests to DeepSeek V4 Pro by swapping the base URL and API key — no structural refactoring of prompts, tool definitions, or agent loops required. The calling methods are unchanged from preview, so any integration already running in staging continues to work.
The model architecture remains the same as announced in April: 1.6 trillion total parameters with a Mixture-of-Experts design that activates only 49 billion parameters during inference. Context window is one million tokens (~1,500 pages of text). Maximum output length is 384,000 tokens. The model supports thinking and non-thinking modes, text input/output only — no multimodal input.
How capable is it?
On the Artificial Analysis Intelligence Index — an independent evaluation covering reasoning, coding, instruction-following, and multi-step tasks — V4 Pro 0813 scores 53 out of 100, placing it third across 106 evaluated models. That is well above the open-weight median of 27 and competitive with top proprietary systems.
On agent-specific benchmarks published by DeepSeek: Terminal Bench 2.1 (long-horizon agentic tasks) scored 87.9; DeepSWE (software engineering problem-solving) scored 62.7; NL2Repo (natural-language to repository-level code generation) scored 61.5. These figures are self-reported. As of 13 August 2026 no major third-party evaluator has published independent replication of the exact 0813 checkpoint — treat the vendor numbers as directionally useful but not confirmed.
Speed sits at 80.8 tokens per second on Artificial Analysis's infrastructure, above the market median of 66.3. A 97% cache-hit discount applies to repeated prompt prefixes — relevant for teams running batch inference over similar prompts or RAG pipelines with shared context.
The price change you cannot ignore
The biggest operational signal from the GA release is the pricing restructure that takes effect at 16:00 UTC on 16 August 2026. Current flat rates: $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. Post-change peak rates: $1.32 per million input tokens and $3.96 per million output tokens — a 4.5x increase on output. Off-peak prices are set at half the peak rate, rewarding teams that can schedule batch workloads during off-peak windows.
For context, at current rates V4 Pro is roughly 57x cheaper on output than Claude Opus 4 at list price. After 16 August, that gap narrows but the model remains substantially cheaper than most closed proprietary APIs at off-peak hours. The cache-hit discount is unchanged and remains the largest lever for cost control on repeated-query workloads.
Teams that have not run a monthly token audit should do so before end of 15 August. The post-increase bill for a high-volume agentic workload — SWE-agent loops, RAG pipelines, code review automation — can be significantly higher than current projections based on pre-GA pricing.
What it means for US & EU teams
Three decisions become concrete with GA status.
API versus self-hosted. GA confirms the model is production-stable and supported. For teams with standard data-residency requirements — GDPR Article 46 transfers, SOC 2 scope boundaries — running V4 Pro behind DeepSeek's API means data transits to DeepSeek's infrastructure. Self-hosting the MIT-licensed weights on your own cloud account (AWS, Azure, GCP, or on-premises Kubernetes) keeps inference inside your control perimeter. The GPU requirement for the full 1.6T model is substantial; managed inference services such as Together AI and others offer day-zero access for teams that want open-weight openness without owning the hardware directly.
Vendor independence. MIT licensing means no royalty lock-in. Engineering teams building GenAI integration layers that abstract the model provider benefit here: V4 Pro becomes a routeable backend alongside proprietary APIs, chosen per task based on cost, latency, and capability. The OpenAI-compatible API format makes this routing transparent to the application layer.
EU AI Act classification. V4 Pro is a general-purpose AI model. EU AI Act GPAI obligations apply to providers of such models, not necessarily to deployers. That said, if your product uses V4 Pro in a high-risk application under Annex III — recruitment, credit scoring, critical infrastructure — you carry obligations as the deployer. Self-hosting does not transfer provider-level obligations away from you; it changes which party is responsible for which layer.
How to evaluate it this quarter
The structured path: run V4 Pro 0813 through your existing evaluation harness against the specific task distribution your agents handle — not against general leaderboard prompts. Measure task success rate, output token count per task (verbose models are more expensive at scale), total wall-clock time per workflow, and estimated monthly cost at post-16-August rates. Compare to your current model, not to benchmarks.
For teams without an evaluation harness, the practical minimum is: pick five representative tasks, run each 10 times on both models, rate outputs blind, and calculate cost per rated-success at post-increase pricing. That 50-sample data point beats any leaderboard figure for predicting performance on your codebase.
Keep your integration model-agnostic. V4 Pro is the best open-weight option today; that ranking will shift. The team that built a routing layer routes to a better model next quarter without a rebuild.
FAQ
What is DeepSeek V4 Pro 0813 and when did it go GA?
DeepSeek V4 Pro 0813 is the general-availability checkpoint of DeepSeek's flagship model, officially released on 13 August 2026 across the DeepSeek app, web interface, and API. The model had been in preview since the V4 series debuted in April 2026. Calling methods are unchanged from the preview — existing integrations continue to work without code changes.
Are the open weights still MIT-licensed?
Yes. The V4 series was released with MIT-licensed weights permitting self-hosting and commercial use without royalties. The 0813 GA checkpoint retains that license. Teams should verify the exact scope against legal counsel for regulated environments, particularly given Chinese data-residency obligations that may apply to the model's origin, but the license itself is genuinely permissive.
How does V4 Pro 0813 compare to other frontier models for agentic coding?
According to Artificial Analysis, V4 Pro 0813 ranks third out of 106 models on the Intelligence Index, well above the open-weight median. On agent benchmarks reported by DeepSeek: Terminal Bench 2.1 scored 87.9, DeepSWE scored 62.7, NL2Repo scored 61.5. Independent third-party replication of the exact 0813 checkpoint was pending as of 13 August 2026 — treat vendor figures conservatively until corroborated by third parties.
What is the price increase on 16 August and how should teams plan?
At 16:00 UTC on 16 August 2026 DeepSeek introduces peak/off-peak billing. Peak output rises from $0.87 to $3.96 per million tokens — a 4.5x increase. Off-peak prices are set at half the peak rate. Cache-hit rates remain deeply discounted (97% off). Teams should audit monthly token volumes, calculate the post-increase bill, and decide whether to shift batch workloads to off-peak hours or evaluate self-hosting via the MIT-licensed weights.
What does native OpenAI Responses API compatibility mean for teams?
Any codebase already calling the OpenAI SDK can route requests to DeepSeek V4 Pro by changing the base URL and API key — no structural prompt refactoring required. This sharply lowers evaluation cost: teams can A/B test DeepSeek alongside existing model providers without changing their agent or tooling layer.
Sources: DeepSeek API Documentation — api-docs.deepseek.com (first-party release notes, 13 August 2026); Artificial Analysis — DeepSeek V4 Pro 0813 benchmarks and pricing; Quartz / QZ — DeepSeek officially launches V4-Pro AI model in August 2026 (13 August 2026).