AI Knowledge HubDeepSeek-V4.1-Flash-1017 Review: It Gets Truly Cheap When You Use It Outside the Seven Peak Hours
Conclusion first: the version stays the same, and knowing the hours that cost half is the whole trick
DeepSeek-V4.1-Flash-1017 is an updated snapshot of V4.1-Flash released on October 17, 2026. DeepSeek called it a minor update, and the API model name stays a single string: deepseek-flash. Leave your code alone and it connects to the latest version automatically. Pricing is identical to the previous release. Only the performance benchmarks and the output length changed.
There are two things I want to leave in this post. One is what 1017 actually improved, and the other is how to read this model's price table. The second is what really cuts your bill in half.
1. The version name 1017 is a date
| Item | Detail |
|---|---|
| Model name | DeepSeek-V4.1-Flash-1017 |
| Release date | October 17, 2026 |
| API call name | deepseek-flash (unchanged) |
| Previous version | DeepSeek-V4.1-Flash (September 10, 2026) |
| Context | 1M tokens |
| Max output | 384K tokens |
| Parameters | 552B total, 8B active on input / 16B on output |
| Architecture | Sparse MoE on a Causal Encoder-Decoder (CED) backbone, native image understanding |
1017 is a snapshot label meaning the October 17 release, the same scheme as last month's V4-Flash-0731 and V4-Pro-0813. On the API, deepseek-flash always serves the newest snapshot, and the legacy slug deepseek-v4-flash is kept for compatibility while the old model is retired and routed to 1017.
2. Official benchmarks: numbers copied as they are
These are the figures from DeepSeek's official documentation compared against the previous release. They are not a separate test but the values published in the update announcement.
| Benchmark | V4.1-Flash (9/10) | 1017 (10/17) |
|---|---|---|
| DeepSWE v1.1 | 74.2 | 77.0 |
| Automation-Bench | 54.8 | 60.5 (v1.0.6) |
| Agents' Last Exam | 31.8 | 35.4 |
| ProgramBench | 20.3 | 29.1 |
| NL2Repo-Bench | 65.4 | 68.2 |
| Terminal-Bench 4.0 | 31.2 | 42.3 |
The one closest to daily feel is Terminal-Bench 4.0, which measures how far an agent can carry a task through a terminal: it went from 31.2 to 42.3. DeepSeek presented the update as a Pareto improvement, higher performance with shorter output and faster runtime. Since it means finishing the same work with fewer tokens, the effect on your invoice matters more than the benchmark climb.
3. How to read the price table: only seven peak hours pay full price
The DeepSeek API uses two rates by time window. Peak hours are UTC 01:00-04:00 and 06:00-10:00, Monday through Friday (excluding Chinese public holidays). Everything else, the entire weekend, and Chinese public holidays run at the off-peak rate of half price.
USD per 1M tokens:
| Item | Peak | Off-peak |
|---|---|---|
| Input (cache miss) | $0.30 | $0.15 |
| Input (cache hit) | $0.006 | $0.003 |
| Output | $1.20 | $0.60 |
Converted to Korean time (UTC+9):
| Window | Korean time |
|---|---|
| Peak (full price) | Weekdays 10:00-13:00 and 15:00-19:00 |
| Off-peak (half price) | Every other hour plus 24 hours on weekends |
In short, every moment except those seven daytime hours on weekdays is half price. Night work goes without saying, and Saturday and Sunday are half price all day. There is no separate signup or switch. If the call time falls in the peak window you pay full price; otherwise you pay half.
Put into feel: at 1M output tokens, running at night turns $1.20 into $0.60. One habit, batching work for the night, cuts the API invoice in half. Bulk translation, batch document generation, report cleanup, all of it can be pushed to the night.
4. Notes left from using it directly
Three usage patterns I confirmed.
First, a cache hit costs one fiftieth. Input cache hits run $0.003 off-peak against $0.15 for a miss, one fiftieth the price. In agent workflows that repeat the same system prompt, keeping the prompt cache on is the biggest lever on cost. Cache status is a price difference, not a model difference.
Second, the shorter output shows up on the actual bill. The official announcement emphasizes reduced output length, and for an agent, not wasting tokens mid-run and going to the end is token savings itself. Run the same work twice and the time-window difference comes out larger than the version difference.
Third, if urgent calls pile into the peak window, only the price doubles. Your working hours are the peak window. Any automation that runs while people are in the office is designed away from half price from the start. Pin the scheduler's timezone to UTC and avoid only the off-peak gaps.
5. What falls short
Because peak hours are defined in UTC, they overlap exactly with domestic working hours. Full price runs from 10 a.m. to 7 p.m., which is a theoretical disadvantage for anyone running automation during the day. DeepSeek also does not state whether a request crossing a peak boundary is billed by its start or end time, so leaving a few minutes of margin before the boundary is safer.
All benchmarks above are official figures, not measurements from my environment. Actual feel and billing differences vary with usage patterns.
Closing
DeepSeek-V4.1-Flash-1017 is not a big change. The version name is a date, the price is unchanged, and performance rose a little. The real content is in the small print of the price table. Avoid the seven peak hours and use the weekend, and the same model costs half. This is not a reason to switch to a new version; it is a reason to finally read the existing price table correctly.
Visitor Comments