DeepSeek V4 Pro API 2026: Peak-Valley Pricing & the New Agent Flagship

On 2026-08-13, DeepSeek pushed DeepSeek-V4-Pro-0813 to its official release across the app, web, and API — and it came with a pricing change that every heavy API user will feel within hours. Starting 16:00 UTC on August 16, 2026 (that’s 00:00 Beijing time on August 17), DeepSeek moves the whole V4 line to peak / off-peak billing: run your agents during off-peak windows and you pay half the peak rate.

This is the story to read before your next batch job, because it changes how you should schedule DeepSeek workloads — and because V4 Pro finally closes the agent gap that kept it one step behind V4 Flash. Here’s what shipped, what the new price table actually looks like, and how to route around peak hours. Everything here is verified against DeepSeek’s official pricing page and the official releases changelog.

What 2026-08-13 Actually Shipped

The official deepseek-v4-pro model (version DeepSeek-V4-Pro-0813) is the flagship of DeepSeek’s V4 line. Calling it is unchanged — set model="deepseek-v4-pro" and you get the latest version. Beyond the version bump, three changes matter to developers:

• Agent capability jump. DeepSeek says the production agent performance of the official build is “significantly improved” over the preview. The headline numbers: HLE 42.7/60.0 with tools (and 42.7 without), Terminal Bench 2.1 at 87.9, NL2Repo 61.5, Cybergym 83.3, DeepSWE 62.7, Toolathlon-Verified 74.1, Agents’ Last Exam 25.7, AutomationBench (Public) 31.8, DSBench-FullStack 71.1, and DSBench-Hard 67.2.
• Native Responses API + Codex. deepseek-v4-pro now natively supports the OpenAI Responses API format and is “specifically adapted for Codex,” with a one-script configuration installer. This closes the exact gap we flagged at the V4 Flash launch, when Responses API was Flash-only.
• Three thinking intensities. Both V4-Pro and V4-Flash now support low / high / max thinking modes, so you can pick the reasoning budget per task — low for simple tasks, high for daily agent work, max for the hardest problems.

The model spec sheet keeps the same generous ceiling as the rest of the V4 line: 1M input context, 384K max output, JSON output, tool calls, Anthropic API format, chat prefix completion (Beta), and FIM completion (Beta, non-thinking mode only).

The Big One: Peak-Valley Pricing Starts Today

DeepSeek’s pricing page now carries a prominent notice: “DeepSeek API pricing will be updated to peak / off-peak billing, with off-peak rates at half the peak rates.” The change is effective 16:00 UTC on August 16, 2026 (00:00 Beijing time, August 17), so the table below is live as of today.

Peak hours are 01:00–04:00 UTC and 06:00–10:00 UTC each day, which corresponds to 09:00–12:00 and 14:00–18:00 Beijing time. Every other hour is off-peak — and off-peak is half the peak price. Here is the full table:

Compare that with the flat prices that applied until today: deepseek-v4-pro was $0.003625 (cache hit) / $0.435 (cache miss input) / $0.87 (output) per 1M, and deepseek-v4-flash was $0.0028 / $0.14 / $0.28. The new off-peak Pro price is materially higher than the old flat rate, while peak Pro output jumps to $3.96/M — more than 4.5x the old $0.87. The 50% discount is real, but it applies to a re-tiered baseline, not to last week’s prices.

The Off-Peak Playbook: Cut AI API Spend by 50%

The new billing model turns when you run a DeepSeek workload into a first-class cost lever. If an agent job is not latency-sensitive, moving it out of the peak windows (01:00–04:00 and 06:00–10:00 UTC) halves the token bill.

• Nightly re-indexing and embedding backfills — schedule for off-peak and take the 50% discount for free.
• Batch evaluation / regression suites — off-peak Pro output at $1.98/M instead of $3.96/M compounds on long eval runs.
• Code review and diff-analysis agents — queue them overnight so the morning starts with results at half price.
• Cache-hit traffic stays cheap (Pro $0.022/M off-peak) — keep prompt prefixing on to maximize cache hits.

The one catch: DeepSeek’s peak windows overlap the US evening / EU business morning for many teams, so “real-time” multi-region products will frequently land in peak hours. For those, model selection (Flash vs Pro) and cache-hit engineering matter more than scheduling.

There is also a strategic angle worth naming: peak-hour pricing is a polite way to shape demand toward off-peak capacity. DeepSeek is effectively publishing a time-of-use tariff — the same lever cloud GPU providers increasingly use to smooth load. For API consumers the takeaway is that latency-insensitive workloads should be treated as schedulable batch, and your orchestration layer should be able to move a job across the peak boundary automatically rather than paying the 2x surcharge by default.

V4 Pro Now Drives Responses API & Codex

At the V4 Flash launch in late July, native Responses API support was Flash-only — V4 Pro trailed. The 08-13 update removes that gap: deepseek-v4-pro now natively supports the OpenAI Responses API format, and DeepSeek has tailored it for OpenAI Codex with a one-script configuration installer (curl the setup script, and Codex CLI, the ChatGPT desktop app, and the VS Code Codex extension all point at deepseek-v4-pro). If you’ve been running Codex-style agent loops, Pro is now a first-class citizen instead of a workaround.

Thinking Mode: low / high / max

Both V4 models now expose three thinking intensities. The practical guidance from DeepSeek itself: use low for simple tasks (faster, cheaper), high for everyday agent work, and max for complex, multi-step reasoning. Because thinking effort directly trades cost against reasoning depth — and now rides on top of peak/off-peak pricing — pairing thinking level with time-of-day gives you two independent cost dials on the same job.

# OpenAI Chat Completions — thinking mode + model
curl https://api.deepseek.com/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $DEEPSEEK_API_KEY" \
  -d '{
    "model": "deepseek-v4-pro",
    "messages": [{"role": "user", "content": "Refactor this diff and explain the trade-offs."}],
    "thinking": { "type": "max" }
  }'

# Responses API (OpenAI format) — near-identical payload
from openai import OpenAI
client = OpenAI(api_key="your-key", base_url="https://api.deepseek.com")

resp = client.responses.create(
    model="deepseek-v4-pro",
    input="Review this pull request: " + open("pr.diff").read(),
    reasoning={"effort": "high"},
)
print(resp.output_text)

Worked Cost Calculation

A typical agentic code-review pass over a large repository might consume ~900K input tokens (750K cache-hit + 150K cache-miss) and 80K output tokens. Here’s what that costs on deepseek-v4-pro under the new schedule:

The same job that cost about $0.14 on the old flat rate now runs $0.27 off-peak (roughly 2x) and $0.55 at peak (roughly 4x). The discount is real, but the new baseline is higher — which is why teams with large DeepSeek bills should re-budget now rather than assume “off-peak = cheaper than before.”

V4 Pro vs V4 Flash: Which One, When?

With both models under peak/off-peak pricing and both now supporting Responses API and three thinking modes, the Pro-vs-Flash decision is purely about capability vs throughput:

• Choose V4 Pro for the hardest agent tasks — complex software engineering, multi-step reasoning, anything where DeepSWE 62.7 / DSBench-Hard 67.2 / HLE 42.7-with-tools headroom pays off. You get 500 concurrency, which is plenty for sequential agent loops.
• Choose V4 Flash for high-fan-out parallelism and latency-sensitive calls — 2,500 concurrency and the one-third cost of Pro (off-peak Flash output $0.66 vs Pro $1.98) make it the default for batch and pipeline work.
• Hybrid pattern: run the orchestrator on V4 Pro (deep reasoning) and farm out mechanical sub-tasks to V4 Flash (cheap, parallel). This is the same split teams have used since the V4 Flash launch.

For context against the wider 2026 agent-API landscape, DeepSeek’s V4 line now brackets the same terrain as Kimi K3 (Moonshot’s 1M-context agent model) and the GPT-5.6/Claude pricing tiers we reviewed in our V4 Flash launch deep-dive. DeepSeek’s edge is the peak/off-peak lever: no Western provider currently lets you literally half your bill by shifting hours.

Frequently Asked Questions

Is the pricing change retroactive or forward-looking?

Forward-looking. It takes effect at 16:00 UTC on August 16, 2026. Usage before that is still billed at the old flat rates; anything after is priced under the new peak/off-peak table.

Which hours count as off-peak?

All hours outside 01:00–04:00 UTC and 06:00–10:00 UTC are off-peak. In Beijing time those peak windows are 09:00–12:00 and 14:00–18:00.

Does cache-hit pricing also get the 50% discount?

Yes — cache-hit input is $0.007/$0.014 for Flash and $0.022/$0.044 for Pro (off-peak/peak). Cache hits were always cheap; they stay the cheapest tier under the new schedule.

Can I still call deepseek-chat and deepseek-reasoner?

For now, yes, but DeepSeek has said the legacy names will be retired three months after the V4 announcement. Point new code at deepseek-v4-pro or deepseek-v4-flash to avoid a breaking switch later.

Does deepseek-v4-pro support the Anthropic API format too?

Yes. Alongside the OpenAI Chat Completions and Responses formats, DeepSeek exposes Pro at the Anthropic-format endpoint (https://api.deepseek.com/anthropic), so Claude SDKs and agent frameworks that speak the Anthropic protocol can target deepseek-v4-pro with minimal changes.

Why is Flash capped at a higher concurrency than Pro?

DeepSeek limits deepseek-v4-pro to 500 concurrent in-flight requests versus 2,500 for deepseek-v4-flash. The higher Pro ceiling is reserved for the most demanding agent loops; Flash’s larger concurrency budget and one-third cost make it the default for fan-out-heavy pipeline work.

Verdict: Re-Budget Now, Schedule Around Peak Hours

DeepSeek V4 Pro’s official release is a genuine agent-capability step up — HLE 42.7 with tools, Terminal Bench 87.9, native Responses API and Codex support, and three thinking intensities on a 1M-context flagship. But the news that changes your bill today is the peak/off-peak restructure: off-peak output is 2.3x the old flat rate, and peak is 4.5x. That’s not a “coming soon” footnote — it lands at 16:00 UTC on August 16.

Act this week: (1) re-budget any DeepSeek spend you currently forecast at the old flat numbers, (2) move batch and overnight agent workloads into off-peak windows to lock in the 50% discount, and (3) for latency-sensitive peak-hour traffic, keep the Flash-vs-Pro split sharp (Flash at 1/3 the cost and 5x the concurrency is the sensible default).

If you prefer not to manage DeepSeek direct billing alongside OpenAI, Anthropic, and Google traffic, FreeModel exposes deepseek-v4-pro (and V4 Flash, Kimi K3, Claude, and GPT-5.6 models) through a single API key, with usage dashboards that surface cost-per-task so you can see exactly what the peak/off-peak change does to each workload.