Sometime around 11 p.m.
Beijing time on August 13, 2026, the `deepseek-v4-pro` endpoint quietly became DeepSeek V4 Pro 0813 and nobody sent a memo.
The model had been living in preview for nearly four months. Then it graduated to general availability. Chat interface picked it up, API picked it up. And the market found out from tracker sites after the fact.
You keep hitting the same `deepseek-v4-pro` model name and the backend silently flipped to `DeepSeek-V4-Pro-0813`.
No migrations. No name swaps. If you had traffic running before the rollout, you woke up already on the new checkpoint.
For an operator routing client work, that’s the friendliest versioning call DeepSeek could’ve made.
Zero risk that a checkpoint string change breaks a pipeline overnight.
The OpenRouter model page actually listed 0813 as the GA build a day earlier, on August 12, 2026.
And at least one tracker notes the API pricing hasn’t been formally revised either.
This is a lab that ships into the product first and lets somebody else handle the announcement.
DeepSeek V4 Pro 0813 Pricing and Token Economics
Three numbers set the bill.
$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.
The cache-hit rate runs dramatically cheaper than the miss rate. That’s the entire production story right there, and most routing decisions gloss right over it. The miss rate — $0.435/M. Is your ceiling, what you eat on cold starts and queries you’ve never seen. The cache-hit rate is where your real ongoing cost gets set, assuming your workload looks like mine: a stable system prompt, a cached context blob, repeated calls. Neither number sells itself in isolation. The gap between them is the part that decides your invoice at the end of the month.
One throughput constraint to flag before you scale anything. The Pro endpoint reportedly caps at 500 concurrent requests, versus 2,500 on the Flash tier. Fan out a few hundred parallel agent calls and that 500 ceiling bites first — before price is even a factor.
DeepSeek V4 Pro 0813 Context Window and Specs
1-million-token context window. 384,000-token maximum output (deepseekv4pro.com). Enough room to load a sizable codebase, a long call transcript, or an entire documentation set into one request. On paper, that’s a genuine production-grade spec sheet.
Then you check the composite leaderboard and the picture gets muddy.
DeepSeek V4 Pro 0813 scores 60.9 out of 100 and sits at #48 of 217 across 34 benchmark rows, with Knowledge as its best category at #17. Solid. Mid-pack. Unremarkable in the aggregate.
You’d scroll right past it if you were shopping purely on headline scores.
DeepSeek V4 Pro 0813 Agent Benchmarks: Where the Jumps Are
This is where the release stops being boring.
Terminal Bench 2.1 rises from 72.1 to 87.9. DeepSWE moves from 12.8 to 62.7. DSBench-FullStack improves from 41.8 to 71.1 versus the preview checkpoint (deepseekv4pro.com). DeepSWE nearly quintupling isn’t a tuning tweak. That’s the difference between a model that can’t finish an agentic coding loop and one that can.
The same reporting says V4 Pro 0813 beats Opus 4.8 on Terminal Bench 2.1, Cybergym, DeepSWE.
And AutomationBench, and clears its own preview across those same agent benchmarks too (deepseekv4pro.com).
The point isn’t that V4 Pro 0813 is secretly the best model on earth. It’s that aggregate scores are actively misleading for a release like this. A model ranking #48 overall can still be the right tool for a narrow, high-value agent job at a fraction of what a top-tier model charges per token. The leaderboard is a tiebreaker for general chat. The category benchmarks are the decision tool for routing real workloads.
Half these tracker sites copy each other’s numbers verbatim, which is its own separate mess. And there’s no DeepSeek model card behind any of these figures yet. Treat the deltas as reported until you validate them against your own evals.
Routing Production Traffic to DeepSeek V4 Pro 0813
If you move production AI traffic, this is the week for a focused eval. Not a stack rewrite. Point your existing `deepseek-v4-pro` calls at your real agentic tasks and log whether the new backend regressed anything. The endpoint name didn’t change, which means your regression risk stays invisible until you measure it. Build the billing model around the cache-hit rate, not the miss rate. And watch that 500-request Pro concurrency ceiling before you scale. Because on this tier throughput is the first wall you hit, not price.
The pattern across these releases is consistent: ship first, announce if asked, let the benchmark sites carry the marketing.
Whether the reported agent jumps survive independent testing is the open question. And it’s exactly the one I want answered before I move paid client traffic onto this checkpoint. The smart play is a parallel eval, same prompts and same rubric as whatever you run today, scored on your actual outputs. If the numbers hold, V4 Pro 0813 is a serious option for agent work at a price where the math genuinely works. If they don’t, you’ve spent a few dollars to dodge an expensive migration.
Run the eval before you move traffic. Trust your own numbers over the leaderboard.
DeepSeek V4 Pro 0813 FAQ
Does DeepSeek V4 Pro 0813 support a 1-million-token context window?
Yes. V4 Pro 0813 keeps the 1-million-token context window with a 384,000-token maximum output.
Enough for a full codebase or a long transcript in a single call (deepseekv4pro.com).
What’s the API pricing for DeepSeek V4 Pro 0813?
$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. The cache-hit rate runs dramatically cheaper than the miss rate. So prompt caching is where the real cost picture gets decided.
Do I need to change my API calls for DeepSeek V4 Pro 0813?
No. You keep using the `deepseek-v4-pro` model name and the backend silently became `DeepSeek-V4-Pro-0813` with no migration or model-name swap required.
Is DeepSeek V4 Pro 0813 strong on agent and coding benchmarks?
Reported figures show major gains: Terminal Bench 2.1 rose from 72.1 to 87.9, DeepSWE from 12.8 to 62.7. And DSBench-FullStack from 41.8 to 71.1 (deepseekv4pro.com). V4 Pro 0813 reportedly beats Opus 4.8 on Terminal Bench 2.1, Cybergym, DeepSWE, and AutomationBench.
These are single-source numbers, though — validate on your own evals before moving production traffic.
