What shipped on 17 October
The announcement is titled "Wish-Driven Programming Made Real". DeepSeek describes 1017 as stronger on benchmarks, faster to finish tasks and shorter in its outputs than the previous Flash, which it calls a Pareto improvement in performance and cost.[3] The model card calls it a minor update to DeepSeek-V4.1-Flash. The previous Flash was released on 10 September according to DeepSeek's changelog.[3]
The weights are on Hugging Face under the MIT license, listed at 763B total parameters. The card does not state an active-parameter count.[2] The same release opened DeepSeek's Search and Fetch APIs to all users at $0.75 per 1,000 calls per endpoint; calls made through server-side tools in DeepSeek Harness, Claude Code and Codex are free until billing starts on 31 October 2026 at 16:00 UTC.[3]
The benchmark claims, with the caveat attached
Every number below comes from DeepSeek's own model card, not an independent run. Treat them as the vendor's before-and-after, which is useful for judging whether upgrading from the old Flash matters, and not as a ranking against other labs.
The Terminal-Bench jump of 11.1 points is the headline, and it is a coding-agent score, which is exactly the traffic that burns tokens. The smaller DeepSWE gain (2.8 points) is a reminder that gains are uneven by task.[2]
The price table
DeepSeek lists two API models. Prices are USD per 1M tokens, and DeepSeek states that off-peak rates are half of peak rates.[1]
Both models have a 1M context window and a 384K maximum output. Concurrency limits are 2,500 for deepseek-flash and 500 for deepseek-v4-pro.[1] Our earlier post on the DeepSeek V4 Pro price cut covers how Pro got to this rate card; the short version for today is that Pro costs 4.4x Flash on cache-miss input and 3.3x on output.
Peak hours in your timezone
DeepSeek defines peak as "01:00 - 04:00 and 06:00 - 10:00 UTC, Monday through Friday, excluding Chinese public holidays". All other hours are off-peak, including weekends and Chinese public holidays in full.[1] The page states the days against UTC. Plan on the UTC calendar: Friday 01:00 UTC is still Thursday evening in the Americas, and Monday 01:00 UTC is Sunday evening there, so some peak hours land on your Sunday night.
US daylight saving time ends on the first Sunday of November, which is 1 November in 2026, so the US rows shift by an hour next weekend. India has no DST. The sharp edge for IST teams: the second peak window, 11:30-15:30, sits in the middle of the Indian working day. Interactive tools your team uses at lunchtime run at the doubled rate; a US East Coast team mostly sees peak in the small hours and pays the doubled rate only on night jobs.
One thing the pricing page does not say is whether a request is billed at the rate in force when it starts or when it finishes. If you run long agent loops across a window boundary, test it on a small job and read the usage record before relying on either assumption.
Monthly bills: peak, off-peak and mixed
Take a coding-agent backend that reads a lot of repeated context. Per month: 300M cache-hit input tokens, 100M cache-miss input tokens and 60M output tokens. All arithmetic below is dollars per month, tokens in millions.
Workload A on deepseek-flash
- Off-peak: 300 x 0.003 = 0.90, 100 x 0.15 = 15.00, 60 x 0.60 = 36.00. Total $51.90.
- Peak: 300 x 0.006 = 1.80, 100 x 0.30 = 30.00, 60 x 1.20 = 72.00. Total $103.80.
- Spread evenly across the clock: 35 of 168 hours are peak, so 51.90 x (1 + 35/168) = $62.71.
- Skewed to working hours, with 30 percent of tokens in peak: 51.90 x 1.30 = $67.47.
- Scheduled so only 5 percent lands in peak: 51.90 x 1.05 = $54.50.
Workload A on deepseek-v4-pro
- Off-peak: 300 x 0.022 = 6.60, 100 x 0.66 = 66.00, 60 x 1.98 = 118.80. Total $191.40.
- Peak: double, $382.80. With 30 percent in peak: 191.40 x 1.30 = $248.82.
Workload B: a nightly batch with no caching
A document-processing job sends 1,000M cache-miss input tokens and produces 100M output tokens a month.
- Flash off-peak: 1,000 x 0.15 = 150.00, plus 100 x 0.60 = 60.00. Total $210.00.
- Flash peak: 1,000 x 0.30 = 300.00, plus 100 x 1.20 = 120.00. Total $420.00.
- Pro off-peak: 1,000 x 0.66 = 660.00, plus 100 x 1.98 = 198.00. Total $858.00.
The lever is clear. For a batch job, scheduling is a 50 percent discount you can apply with one line of cron. For an interactive tool you cannot schedule, the realistic goal is shifting what you can (evals, indexing, summarising backlogs) and accepting the rest at peak. Plug your own token counts into the AI model cost calculator, and use the agent run cost simulator if your spend comes from multi-step agent runs where the context grows each turn.
Scheduling batch work off-peak
Off-peak, by the pricing page, is 04:00-06:00 UTC and 10:00 UTC to 01:00 UTC the next day on weekdays, plus all of Saturday and Sunday. The long window is 15 hours, so one nightly start at 10:15 UTC (15:45 IST) has room to run until 01:00 UTC. Cron uses the server timezone, so either run the host in UTC or set it explicitly:
CRON_TZ=UTC
15 10 * * * /opt/jobs/nightly-batch.sh
For workers that run continuously, put a guard at the top of the loop so they pause instead of paying double. This version exits when it is inside a peak window (Chinese holidays are off-peak, so the guard errs on the side of waiting, never of overpaying):
#!/usr/bin/env bash
# peak-guard.sh: exit 0 means "do not run now"
dow=$(date -u +%u) # 1=Mon ... 7=Sun
hr=$((10#$(date -u +%H))) # strip leading zero
if (( dow <= 5 && ( (hr >= 1 && hr < 4) || (hr >= 6 && hr < 10) ) )); then
exit 0
fi
exit 1
Use it as if ./peak-guard.sh; then sleep 900; continue; fi in a worker loop. Pair it with a job-queue size cap so a paused worker does not come back to a backlog that overruns the 15-hour window. If a single request must not be interrupted, check the guard before sending each request, not once per job.
Migrating off the legacy model names
The names deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted. The pricing page says the corresponding models are retired, that their requests are served by DeepSeek-V4.1-Flash-1017, and that they bill at the Flash price.[1] The announcement adds that the calling method is unchanged.[3] I could not find a date after which those aliases stop working, so treat them as a convenience, not a contract.
- Change the model string to
deepseek-flash in config, not code, so you can roll back.
- Re-run your own eval set. The card says outputs are shorter, which changes token counts and your bill in your favour, but it can also break parsers that expected the old verbosity.
- If you used the vision-exp alias, test image inputs explicitly. The card reports visual benchmark gains (Chartography 83.0 vs 78.9, BabyVision 92.5 vs 89.6) but those are DeepSeek's numbers.[2]
- Check your concurrency. The 2,500 limit is per the pricing page; confirm it applies to your account tier before you scale a worker pool.
How the rate card compares
Here are list prices per 1M tokens on uncached input and output, from each vendor's own pricing page. Gemini's output price includes thinking tokens.
Sources: DeepSeek[1], Anthropic[6], Google[5]. Two honest observations. First, Haiku 5.5 undercuts Flash on both sides at list price for prompts under 100K tokens, and Flash peak is more than twice Haiku's output rate. Flash wins where cache hits dominate ($0.003 vs Haiku's $0.01 cache read), where your prompts run past 100K (Haiku jumps to $0.50 / $2.50, Flash does not), and where you want open weights. Second, Gemini 4 Argon is a different tier of model with a different price; the roughly 17x output gap to Flash off-peak ($10.00 vs $0.60) is not a like-for-like comparison. Our Gemini 4 Argon pricing breakdown has the December doubling math.
Worked check on Workload A with Haiku 5.5 (prompts under 100K, cache read $0.01): 300 x 0.01 = 3.00, 100 x 0.10 = 10.00, 60 x 0.50 = 30.00, total $43.00, against $51.90 for Flash off-peak. Haiku is cheaper here on paper. Quality per task, not list price, decides this, so measure on your own eval set.
Self-hosting the weights
The MIT license means you can run it yourself, and the card points to vLLM, SGLang, Docker Model Runner and quantized builds for llama.cpp, Ollama and LM Studio.[2] The card gives no hardware requirements, so I will not invent any. The only number to anchor on is arithmetic from the stated size: 763B parameters at 8 bits is roughly 763 GB for weights alone, before KV cache for a long context. That is a multi-GPU server, not a workstation. If you are not already running one, the API at $0.15 / $0.60 off-peak is cheaper than renting that hardware for any bursty workload. For the free-tier route to try the model without paying, see our guide to using DeepSeek for free.
Frequently asked questions
What is DeepSeek V4.1 Flash 1017?
It is a minor update to DeepSeek-V4.1-Flash, released 17 October 2026 under the MIT license. The API model name is deepseek-flash, with a 1M context window and up to 384K output tokens.[1][2]
When are DeepSeek API peak hours?
01:00-04:00 and 06:00-10:00 UTC, Monday to Friday, excluding Chinese public holidays. In IST that is 06:30-09:30 and 11:30-15:30. Weekends and Chinese public holidays are off-peak all day.[1]
How much cheaper is off-peak?
Half. Flash cache-miss input drops from $0.30 to $0.15 per 1M tokens, and output from $1.20 to $0.60.[1]
Do I have to change my code from deepseek-v4-flash?
No. The old names are still accepted and served by 1017 at the Flash price. Moving to deepseek-flash is still the safer option because no cutoff date for the aliases is published.[1]
Are the benchmark numbers independent?
No. They are from DeepSeek's model card, including the Claude Code harness for Terminal-Bench 4.0. Run your own tasks before switching production traffic.[2]
Sources
- DeepSeek API Docs: Models and Pricing
- DeepSeek-V4.1-Flash-1017 model card, Hugging Face
- DeepSeek news, 17 October 2026: DeepSeek-V4.1-Flash-1017
- Hacker News discussion: DeepSeek-v4.1-Flash-1017
- Gemini API pricing, Google AI for Developers
- Claude API pricing, Anthropic
Comments · 0
Beta: comments are stored locally on your device and not visible to other readers.
No comments yet. Be the first to share your thoughts.