DeepSeek V4 Pro Goes GA: Frontier-Level Agentic AI at 57x Lower Cost

DeepSeek Quietly Ships Its Production-Grade Flagship

On 13 August 2026, DeepSeek released V4 Pro 0813 as the general-availability production checkpoint for its flagship reasoning model, without a press release, blog post, or formal announcement. The release appeared first as an update to DeepSeek's own API documentation, listing DeepSeek-V4-Pro-0813 as the model version behind the deepseek-v4-pro endpoint, and simultaneously on OpenRouter. No changelog described what changed between the April 24 preview and the production build; the version number and API endpoint update were the announcement. For a model that positions itself at the frontier of AI reasoning and agentic capability, the understated launch signals a company increasingly focused on engineering output over marketing — a posture that has become characteristic of DeepSeek's release cadence since V4 Flash's equally quiet general availability in late July.

Architecture: 1.6 Trillion Parameters, 49 Billion Active per Token

DeepSeek V4 Pro is a Mixture-of-Experts system with 1.6 trillion total parameters and 49 billion parameters active per inference token. The MoE design is the same structural principle as DeepSeek's earlier flagship releases: a large total parameter count provides depth of specialised capability, while sparse activation limits compute cost at inference time by routing each token only through the relevant subset of model weights. The context window is 1,048,576 tokens — effectively one million tokens — with a maximum output of 384,000 tokens. For agentic workflows that require maintaining long context across extended multi-step tasks, tool-use sessions, and code repositories, the combination of a one-million-token context window and a 384,000-token output ceiling positions V4 Pro as one of the highest-capacity models for long-horizon agent tasks available in August 2026.

Benchmarks: Strong Terminal-Bench Score, Contested SWE-bench Claims

DeepSeek's self-reported benchmark figures place V4 Pro 0813 at 87.9 on Terminal-Bench 2.1 — a score that sits 0.1 points below Anthropic Fable 5 at 88.0 and 2.9 points above Claude Opus 4.8. The April preview build had scored 72.1 on this benchmark, meaning the production checkpoint represents a 15.8-point improvement. On SWE-bench Verified, DeepSeek's self-reported score is 80.6 per cent, claimed against real-world software engineering tasks including code repair and test generation. However, the independently run DeepSWE evaluation by yage.ai placed the preview build at just 8 per cent pass@1 — a gap from the DeepSeek self-report that no third-party evaluator has closed in the weeks since the GA release. Buyers evaluating V4 Pro for agentic coding workflows should treat the 80.6 per cent figure as vendor-claimed until externally replicated. On the Artificial Analysis Intelligence Index, V4 Pro 0813 scores 53, one point above V4 Flash 0731 at 52.

Terminal-Bench and Agent Tasks

Terminal-Bench 2.1 measures a model's ability to complete multi-step terminal sessions — tasks that require tool use, sequential decision-making, and error recovery in a shell environment. V4 Pro's 87.9 on this benchmark, within 0.1 of Fable 5, is the result that most justifies the model's positioning as a frontier-quality agentic model. Terminal-Bench is one of the more operationally representative benchmarks for agentic software development tasks, which is the primary use case DeepSeek has positioned V4 Pro to serve. V4 Flash 0731 scored 82.7 on the same benchmark, meaning V4 Pro 0813 gains 5.2 points over its sibling Flash model at roughly the same output token price.

Pricing and the Coming Increase

V4 Pro 0813 is priced at $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. DeepSeek announced alongside the GA release that pricing will rise at 16:00 UTC on 16 August 2026, with the specific new rates not disclosed at time of publication. The pre-increase output pricing of $0.87 per million tokens compares to approximately $50 per million output tokens for Anthropic Fable 5 — making V4 Pro approximately 57 times cheaper on output — and is 28.7 times cheaper per output token than Claude Opus 4.8. The cache-hit pricing of $0.003625 per million input tokens makes V4 Pro economically viable for agentic pipelines where system prompts and large contexts are reused across many calls, the scenario where context caching most dramatically reduces operational cost per inference.

What V4 Pro Means for Indian Software Teams Building AI Products

For Indian software teams building AI agents, coding assistants, and enterprise workflow automation, V4 Pro 0813 raises a specific question: at near-Fable-5 Terminal-Bench quality and 57 times lower output cost, should it replace a frontier model in the inference layer of a production agentic system? The answer depends on which quality dimension matters most for the specific workflow. For agentic tasks measured by Terminal-Bench 2.1 — tool use, multi-step execution, code generation and repair in a shell — the gap between V4 Pro and the current frontier is 0.1 points on one benchmark. The SWE-bench Verified claim of 80.6 per cent remains contested and should not be taken as a decided comparison point for coding workflows until independently reproduced. For teams currently paying frontier model pricing to power long-running agent loops, V4 Pro's one-million-token context window and cache-hit pricing make it the clearest cost-optimisation candidate as of August 2026 — provided teams independently validate the benchmarks relevant to their specific workflows before committing to production infrastructure changes.

The Bottom Line

DeepSeek V4 Pro 0813 reached general availability on 13 August 2026 with no formal announcement, updating only its API documentation. The model is a 1.6-trillion-parameter Mixture-of-Experts system with 49 billion active parameters per token, a one-million-token context window, and a maximum output of 384,000 tokens. It scores 87.9 on Terminal-Bench 2.1, compared to 88.0 for Anthropic Fable 5 and 82.7 for V4 Flash 0731. DeepSeek's self-reported SWE-bench Verified score of 80.6 per cent has not been independently replicated. Current pricing is $0.435 per million input tokens and $0.87 per million output tokens, with a price increase taking effect at 16:00 UTC on 16 August 2026. For Indian software teams building cost-sensitive agentic pipelines, V4 Pro 0813 offers near-frontier agentic benchmark quality at a price point that substantially changes the economics of running large-context agent loops in production.