DeepSeek launched V4-Pro today, the production checkpoint (dated 0813) following the preview build referenced in its own comparison table, alongside updates to V4-Flash that add tunable reasoning effort. Two headline features: flexible reasoning effort (low for simple tasks, high for daily agent workflows, max for complex tasks) across both V4-Pro and V4-Flash, and native OpenAI Responses API support, "optimized for Codex with one-click setup." It's live now in app/web under "Expert Mode" and via API, with model names left unchanged — existing integrations don't need to update anything to get the new checkpoint.
Where it actually stands
DeepSeek's own table compares V4-Pro-0813 and V4-Flash-0731 against both models' preview builds and three rivals — GLM-5.2, Kimi-K3, Claude Opus 4.8, and Claude Fable 5 (the last one flagged "w/ fallback," an asterisk DeepSeek doesn't explain anywhere in the post, worth noting as an unexplained caveat rather than glossing over it). Across ten agentic and reasoning benchmarks, V4-Pro-0813 leads outright on two — Cybergym (83.3, edging Fable 5's 83.1) and AutomationBench Public (31.8, ahead of Kimi-K3's 30.8) — and is genuinely competitive on several more, beating Claude Opus 4.8 on HLE-with-tools (60.0 vs 57.9) and Terminal Bench 2.1 (87.9 vs 85.0). But Fable 5 and Kimi K3 lead most of the table: DeepSWE (70.0 and 67.5 vs V4-Pro's 62.7), Toolathlon-Verified (77.9 and 76.5 vs 74.1), DSBench-FullStack (77.2 and 73.7 vs 71.1), Agents' Last Exam (Kimi-K3's 27.6 vs V4-Pro's 25.7). This is a model that's closed most of the distance to the frontier, not one that's claiming it.
The generational jump is the more honest headline
The more interesting comparison in DeepSeek's own table is V4-Pro-0813 against its own preview build, and it's dramatic: DeepSWE 12.8 → 62.7, nearly 5x. Terminal Bench 2.1 72.1 → 87.9. Cybergym 52.7 → 83.3. AutomationBench 12.8 → 31.8, two and a half times. That's the same shape we've now seen from Grok 4.6 against 4.5 and Ling-3.0-tiny's local-deployment jump this month — the preview-to-launch gap on agentic benchmarks is consistently larger than the launch-to-rival gap, across labs. Whatever's happening in the last mile of post-training between "preview" and "shipped" is doing more work right now than the base model generation is.
One methodology footnote worth flagging: the code-agent numbers were run on "our upcoming DeepSeek Harness (minimal mode)" — a harness DeepSeek describes as still forthcoming, being used to generate official comparison numbers before it's itself released. This "DeepSeek Harness" was already mentioned as upcoming when V4-Flash launched in July; it's still "upcoming" a month later, which either means it's close or means "upcoming" is doing some marketing work of its own.
The real story is in the pricing graphic
DeepSeek's separate pricing announcement is the part worth sitting with. New API pricing, effective 16:00 UTC on August 16, introduces peak and off-peak rates — peak hours defined as 01:00–04:00 and 06:00–10:00 UTC, everything else off-peak:
| Model | Off-peak output | Peak output |
|---|---|---|
| DeepSeek-V4-Flash | $0.66 | $1.32 |
| DeepSeek-V4-Pro | $1.98 | $3.96 |
Set that against the $0.87-per-million-output figure this blog logged for DeepSeek V4 pricing two weeks ago: V4-Flash's new off-peak rate is actually a modest cut, to $0.66. Its peak rate is a 52% increase over that old flat number. And V4-Pro is an entirely new, separately priced tier that didn't exist as a distinct line before — at up to $3.96 per million output tokens during peak hours, more than four and a half times the old flat DeepSeek rate.
This matters beyond DeepSeek's own price sheet. Time-of-day surge pricing is not something we've seen from any other major lab's consumer API — OpenAI, Anthropic, and Google all price flat per-token regardless of clock time, sometimes with batch or priority tiers, never with an hourly demand curve. DeepSeek built its reputation as the lab that broke the price floor; introducing peak-hour pricing, even at rates still cheap by frontier standards, is DeepSeek borrowing a mechanism from cloud compute and ride-hailing rather than from any other model vendor. It's a rational response to the same problem Nemotron 3.5 Lightning and NeMo Switchyard were built to solve — demand isn't flat, and flat pricing under variable demand means either overprovisioning or degraded service at peak — but it's a structural first for this tier of the market, worth watching for whether it spreads.
The model card confirms the weights, the license, and one real friction point
The full model card is now readable, and it closes out most of what this post flagged as unconfirmed. License: MIT — the least encumbered choice available, more permissive than the revenue-gated licenses this month's coverage has been tracking from other labs. The card also names the architectural change behind the "0813" release directly: V4-Pro-0813 is "built on the DeepSeek-V4-Pro (Preview) model structure, with a DSpark speculative decoding module attached" — the same base model, plus an inference-acceleration component. The card doesn't claim DSpark itself explains the dramatic preview-to-launch benchmark jump highlighted above, but it's a plausible partial contributor worth naming: DeepSeek's own framing is that the improvements are "especially pronounced in production environments," and speculative decoding buys more usable reasoning within the same wall-clock budget on exactly the long-horizon agentic tasks where the biggest jumps showed up.
One genuine friction point, worth flagging rather than skipping past: this release ships with no standard Jinja chat template. Instead of the chat_template.jinja file nearly every Hugging Face model card includes, DeepSeek provides a dedicated Python encoding module with custom functions to serialize messages into the model's input format and parse its output — a bespoke toolchain rather than the ecosystem-standard one. That's a real, if minor, tax on "open means anyone can just load it": vLLM and SGLang both get first-class support with specific launch flags for DSpark (--speculative-config with method: dspark on vLLM; --speculative-algorithm DSPARK on SGLang, target and draft weights sharing the same checkpoint), so the major serving frameworks are covered — but a hobbyist reaching for the usual AutoTokenizer.apply_chat_template() workflow will hit a wall this specific release doesn't smooth over.
Recommended local settings: temperature 1.0, top_p 0.95 for agentic use (1.0 otherwise), and a 384K-token maximum output length for the high and max reasoning-effort tiers — a genuinely large budget, consistent with a model built for long-horizon agent work rather than quick answers.
What to expect next
- Watch whether peak/off-peak pricing spreads. If a second lab introduces time-based API pricing within the next quarter, DeepSeek didn't just cut a price today, it introduced a pricing mechanism the rest of the market may adopt.
- Watch the DeepSeek Harness actually ship. It's been "upcoming" since at least July while already generating official benchmark numbers — the gap between those two states is worth closing.
- Watch the Fable 5 "w/ fallback" asterisk get explained, by DeepSeek or by Anthropic. An unexplained caveat on the one model DeepSeek is most directly competing with is exactly the kind of detail worth someone else's independent verification.
- Watch whether the custom encoding toolchain becomes a standard DeepSeek pattern or was specific to shipping DSpark quickly. A permanent departure from Jinja templates would be a small but real ongoing cost for every downstream tool that assumes one exists.
References: DeepSeek (@deepseek_ai) — V4-Pro launch announcement · deepseek-ai/DeepSeek-V4-Pro-0813 on Hugging Face · related coverage: The Price War Nobody Is Actually Fighting · Grok 4.6 launch · Qwen3.8-Max's open weights land · Ling-3.0-tiny · Nemotron 3.5 Lightning · Frontier Arcade: trends & predictions