DeepSeek API Pricing Update
TL;DR Highlight
DeepSeek launched V4-Pro GA and introduced peak and off-peak API pricing
Who Should Read
Backend developers using DeepSeek API in agents or coding tools who are worried about cost changes
Core Mechanics
- V4-Pro is now generally available
- DeepSeek says it brings major agent upgrades.
- It is in Expert Mode on app/web, and API model names are unchanged.
- Reasoning effort has three levels
- Both V4-Pro and V4-Flash support low, high, and max.
- Low is for simple tasks, high for daily agent workflows, max for complex tasks.
- Native OpenAI Responses API support
- It is optimized for Codex with one-click setup.
- Peak and off-peak pricing is new
- Off-peak rates are 50% lower than peak.
- The aim is to allow more flexible workload scheduling.
- New pricing starts 16:00 UTC, Aug 16, 2026
- The post has no price table and points to the pricing docs.
Community Reactions
- Peak hours fall in China's work hours, per commenters
- Some said this favors US customers.
- Others noted Europe's morning overlaps with peak.
- One commenter posted a price breakdown
- Flash input goes from 0.14 to 0.44 at peak (3.1x), output from 0.28 to 1.32 (4.7x).
- Pro input goes from 0.435 to 1.32 (3.0x), output from 0.87 to 3.96 (4.6x).
- Cache-hit price on Pro rose 6x off-peak and 12x peak, a commenter noted
- Long agent sessions can be 90%+ cache hits, so the impact is large.
- They said it is still cheaper than other providers.
- Some say competitiveness dropped
- One user said gpt-5.6-luna got 5x cheaper and is twice as fast, so it now looks like the better deal.
- Another said the cheapest OpenRouter provider charges 0.16/M output for Flash.
- Frequent pricing changes are tiring, one commenter said
How to Apply
- Recompute last month's bill at new prices
- Pull input, output, and cache-hit tokens per model from your logs today.
- Multiply by the peak and off-peak rates from the comment table and compare.
- Move batch jobs to off-peak hours
- Pick latency-insensitive jobs like nightly summaries, eval runs, or bulk preprocessing and set their schedule.
- Confirm the off-peak window in the official pricing docs (quick_start/pricing).
- Split reasoning effort by task
- Use low for classification and extraction, high for normal agent loops, max only for hard problems.
- Compare cost and accuracy per effort level on the same task set.
- Add time-based rates to per-request cost logging
- Store request time in UTC and apply peak or off-peak rates.
- Keep separate rate tables before and after 16:00 UTC on Aug 16.
Try It
# Paste into Claude or ChatGPT (cost check)
DeepSeek API pricing changed to peak/off-peak. Calculate my monthly cost from the usage below.
Rates (per million tokens, input/output/cache read):
- V4-Flash before: 0.14 / 0.28 / 0.0028
- V4-Flash off-peak: 0.22 / 0.66 / 0.007
- V4-Flash peak: 0.44 / 1.32 / 0.014
- V4-Pro before: 0.435 / 0.87 / 0.003625
- V4-Pro off-peak: 0.66 / 1.98 / 0.022
- V4-Pro peak: 1.32 / 3.96 / 0.044
My monthly usage (million tokens): {{model}}, input {{input}}, output {{output}}, cache read {{cache}}, peak share {{peak%}}
Show a table with cost before vs after and the multiplier.
Next: rerun with peak share at 0%, 50%, and 100%.Terminology
Related Papers
Mercury 3: Diffusion at scale
Scaling diffusion LLM 40x gives Mercury 3: 4–8x faster than speed-tier models, similar quality
Timestral: Mistral's time-series foundation models
A 209M time-series model ranks first on three benchmarks and runs on CPU.
DeepSeek V4.1 Flash 1017 (Max)
DeepSeek V4.1 Flash 1017 Max scores 45 on the Intelligence Index at 211 tok/s and $0.30 input.
If you don't run your own LLM proxy, how do you know whats going on?
Running your own LLM proxy shows which keys, models, and cache misses drive your usage.
DeepSeek: DeepSeek v4.1 Flash 1017
DeepSeek V4.1 Flash 1017: $0.05/M input, 1M context, built for agents
DeepSeek-v4.1-Flash-1017
DeepSeek V4.1 Flash minor update lifts coding and agent benchmarks while cutting run time and output length
Show HN: Llmspeed – feel how fast a local LLM runs before you buy the hardware