A GA label, a dated suffix, and a vendor change log that says nothing changed

OpenRouter listed DeepSeek V4 Pro 0813 as a general-availability endpoint on August 12, with a 1-million-token context window and posted rates of $0.435 per million input tokens and $0.87 per million output tokens ^[1]. DeepSeek's own July 31 change log said the V4 Pro API was unchanged and that an official release would follow soon, and DeepSeek's homepage, captured the same day the listing appeared, also said V4-Pro had not changed ^[2][3].

The gap between those two statements is the story. What the reviewed materials establish is a new versioned slug, deepseek/deepseek-v4-pro-0813, fronted by OpenAI-compatible chat completions and Responses endpoints plus an Anthropic Messages endpoint, with reasoning controls, streaming, tool calls and structured output exposed ^[5]. What they do not establish is any 0813-specific model, post-training, capability, latency, throughput or reliability change, and DeepSeek has not explained in the reviewed documentation what the 0813 suffix denotes ^[4].

Compare how DeepSeek handled the small model. The July 31 notice said DeepSeek-V4-Flash-0731 kept the preview model's architecture and size while receiving additional post-training, and it explicitly limited that update to Flash while stating the V4 Pro API, app and web models were unchanged ^[7]. That is what a documented dated revision looks like. There is no equivalent entry for V4 Pro 0813 in the reviewed change log, and the available evidence does not distinguish between OpenRouter versioning an existing deployment, DeepSeek changing serving infrastructure without documenting it, or DeepSeek simply not having published the note yet ^[8].

Nothing else in the listing closes the gap. OpenRouter displayed no latency or throughput measurements when the endpoint appeared, and DeepSeek supplied no 0813-specific coding, reasoning, agent-task or reliability results in the reviewed materials ^[9]. The GA label therefore describes availability on an aggregator, not verified performance against the previous V4 Pro endpoint ^[10]. Calling V4 Pro was already a model-parameter change through OpenAI ChatCompletions or Anthropic-compatible interfaces without touching the base URL, per DeepSeek's April change log for the preview family ^[6]. Re-pinning to a dated slug buys you a name, not a measured delta.

Two operational details are worth more attention than the version string. OpenRouter names DeepSeek as the sole provider and forwards every request directly, with no alternate host listed, so a degraded or unavailable DeepSeek endpoint cannot be routed around; that puts monitoring and fallback on you ^[11]. And DeepSeek documents a maximum output length of 384,000 tokens, roughly 38 percent of the advertised context, which still has to share the allowance with prompt, history and tool results ^[12][3].

The economics are legible but temporary. The pricing page lists $0.435 per million uncached input tokens, $0.003625 per million cache-read tokens and $0.87 per million output tokens, and warns of a future increase ^[13]. Cache reads are about one one-hundred-twentieth of the uncached input rate ^[1], and output costs twice input ^[2], so a maximal single call, a full million uncached input tokens plus 384,000 output tokens, prices out near $0.77 ^[4]. OpenRouter showed a usage-weighted effective input price below the posted rate on launch day because of caching ^[15]. That flatters nobody's forecast unless you model your own cache-hit rate and output consumption, and agent loops multiply tokens across planning, tool calls and retries ^[14].

Watch for a DeepSeek change log entry that actually names 0813 and says what moved, the promised official V4 Pro release, the flagged price increase, and whether latency and throughput numbers ever attach to the listing.