DeepSeek V4.1 Flash 1017 Review: What Changed, Tested vs V4 Pro

DeepSeek upgraded V4.1 Flash to DeepSeek-V4.1-Flash-1017 on October 17, 2026, with the same model ID (deepseek-flash) and the same prices. DeepSeek's own scores rise on every benchmark it lists. In our release-day tests it matched V4 Pro on every task both could run, finished first each time, and cost 4 to 11 times less.

Both models are on OmniaKey: DeepSeek V4.1 Flash and DeepSeek V4 Pro.

What changed in 1017

DeepSeek shipped 1017 on October 17, 2026, as an update to the V4.1 Flash it released on September 10 (our September review of V4.1 Flash). The new model card calls 1017 a minor update that still delivers substantial gains in reasoning, coding, agentic tasks and visual tests. The scores below are DeepSeek's own.

Every row moves up. Terminal-Bench 4.0 gains 11.1 points, ProgramBench 8.8 and Apex Bench v1.1 7.5. The † marks the text-only subset of HLE.

NL2Repo-Bench needs a footnote of its own: the 1017 model card puts the September model at 65.4, while the earlier V4.1 Flash model card lists 64.0, so the gain is 2.8 points on one baseline and 4.2 on the other.

DeepSeek says it also cut task runtime and output length, and calls the result a new performance–cost optimum across harnesses. It names DeepSeek Harness, Tencent CodeBuddy and WorkBuddy as platforms it tuned for. The weights are on Hugging Face as deepseek-ai/DeepSeek-V4.1-Flash-1017 under the MIT license. The model card suggests temperature 1.0, top_p 1.0 and at least 256K max_tokens when you run them yourself.

Do you need to change your code?

No, if you already call deepseek-flash. The change log says the calling method and prices stay the same, and that ID now serves 1017.

• The old names deepseek-v4-flash and deepseek-v4-flash-vision-exp still work and are served by 1017 at the Flash price. DeepSeek calls the arrangement temporary and publishes no end date.
• Thinking is on by default at effort high. reasoning_effort accepts low, high or max, and the default is a reasonable starting point for coding.
• Long thinking runs need room and a streamed connection. Our parser task used 20,709 output tokens on 1017, and V4 Pro returned 70,649 even though we set max_tokens to 32,000. Our non-streamed V4 Pro request was cut off after 10 minutes, and the streamed retry finished in 20.7 minutes.
• Peak hours cost double. At DeepSeek's list prices, checked October 17, Flash cache-miss input is $0.15 per million tokens and output is $0.60 per million off-peak, against $0.30 and $1.20 at peak. V4 Pro is $0.66 and $1.98 off-peak.

Peak hours are 01:00–04:00 and 06:00–10:00 UTC on weekdays, excluding Chinese public holidays. DeepSeek Harness users should move to 0.2.1-rc.1 or later, per the change log.

Our tests: 1017 vs V4 Pro

We started the five tasks on October 17, 2026, between 13:00 and 13:20 UTC, a Friday inside DeepSeek's off-peak window.

How we tested

• Model IDs: deepseek-flash, which DeepSeek's pricing page maps to DeepSeek-V4.1-Flash-1017, and deepseek-v4-pro, whose replies reported deepseek-v4-pro-ga-260813 (DeepSeek-V4-Pro-0813 on the pricing page).
• Calls went to OmniaKey's OpenAI-compatible /v1/chat/completions endpoint.
• Thinking on at the default effort high, with no temperature change. One run per task per model, tasks in parallel. Only the V4 Pro coding retry was streamed.
• Scoring: task 1 against a 30-case test file, with 19 results that must match Python exactly and 11 inputs that must raise the right error. Task 2 by every tool call and its arguments, task 3 by 10 fields, task 4 by three badge codes, task 5 by four answers.
• Time runs from sending the request to the full reply. Cost is the reply's token counts times DeepSeek's off-peak list prices for cache-miss input and output. No cache hits occurred.
• OmniaKey sells API access to both models, so we ran the tests through it. Every price here is DeepSeek's own list price, not ours.
• Sources, all checked October 17, 2026: change log, pricing, model card, thinking guide, Search and Fetch guide.

One run per task is a spot check, not a benchmark.

Task 1: a parser with Python's exact rules

What it checks: precedence, right-associative **, -2**2 = -4, floor division, modulo and the right exceptions, with no eval.

Prompt:

Write a Python 3 function evaluate(expr: str) that evaluates an arithmetic expression without using eval, exec, ast or any parsing library.

• 1017: 30 of 30 cases in 87.7 s, for $0.0125, using 20,709 output tokens, 19,734 of them thinking.
• V4 Pro: 30 of 30 cases, on a streamed retry: 1,239.5 s (20.7 minutes), $0.1401, 70,649 output tokens, 69,456 of them thinking. The first, non-streamed try was cut off after more than 10 minutes with no reply.

Verdict: 1017. Both wrote a correct parser, but 1017 took 88 seconds and V4 Pro almost 21 minutes, at 11 times the cost.

Task 2: parallel tool calls

What it checks: four calls in one reply, two cities and two tools, with the right city in each argument.

Prompt:

I'm flying out tonight. What's the weather right now in Lisbon and in Osaka, and what's the local time in each city? Use the tools.

• 1017: all four calls with the right arguments, 2.6 s, $0.0002, 195 output tokens.
• V4 Pro: all four calls with the right arguments, 5.2 s, $0.0008, 265 output tokens.

Verdict: a tie on correctness, a win for 1017 on time and cost.

Task 3: invoice fields from a messy email

What it checks: 10 fields from a German-English invoice with "2.500 pcs at 0,42 EUR", a "3rd of November 2026" date, 14-day terms and a 5% discount on one line.

Prompt:

Extract this invoice email into JSON with exactly these keys:

• 1017: 10 of 10 fields, 5.2 s, $0.0005, 779 output tokens.
• V4 Pro: 10 of 10 fields, 22.1 s, $0.0034, 1,607 output tokens.

Verdict: 1017. Same answers, four times faster and 6.7 times cheaper.

Task 4: a 263K-token service log

What it checks: the current badge code for Buildings A, B and C, while ignoring a superseded code for B and an unasked code for D.

Prompt: none to quote. The input is a 262,816-token service log, as counted by the API, with five NOTE lines.

• 1017: A, B and C right, the old B and D skipped, 24.1 s, $0.0417, 3,771 output tokens.
• V4 Pro: all right, 31.7 s, $0.1759, 1,221 output tokens.

Verdict: 1017. It thought longer here and still finished first, at about a quarter of the price.

Task 5: a chart

What it checks: the highest month, the lowest month, the six-month average to one decimal, and June minus January. DeepSeek lists vision on Flash only, so 1017 ran alone.

Prompt:

Look at the chart. Which month had the most API calls and which had the fewest? What is the average of the six months, rounded to one decimal? By how many thousand calls did June exceed January? Reply in four short lines.

• 1017: 4 of 4 answers, 3.8 s, $0.0002.

Verdict: 1017 by default. V4 Pro cannot take the image.

Results

V4 Pro took 14.1 times as long on the parser, twice as long on the tool task, 4.2 times as long on the invoice and 1.3 times as long on the log. Across tasks 1 to 4, 1017 cost $0.055 and V4 Pro $0.320, 5.8 times as much at DeepSeek's list prices.

DeepSeek's benchmark claims, read carefully

DeepSeek's 1017 table puts the new model ahead of Claude Opus 5 on four of the eleven tests and behind on seven. These are DeepSeek's numbers, run with its own harnesses.

1017 leads on DeepSWE v1.1 at 77.0 against 74.0, Automation-Bench v1.0.6 at 60.5 against 50.3, Agents' Last Exam at 35.4 against 28.6, and Apex Bench v1.1 at 72.1 against 65.8. Opus 5 still leads on HLE at 56.3 against 42.6, Terminal-Bench 4.0 at 51.8 against 42.3, ProgramBench at 37.0 against 29.1 and NL2Repo-Bench at 75.3 against 68.2. It also leads on the three visual tests: Chartography 84.0 against 83.0, BabyVision 94.1 against 92.5 and ZeroBench-main 52.0 against 51.0.

DeepSeek published no official table that pairs 1017 with V4 Pro; the only V4 Pro scores it published sit in the September V4.1 Flash model card. The chart that maps DeepSWE score against cost per task is an image only, with no values in the page text.

For your own code, the five tests above are a better guide.

The new Search and Fetch API

The same October 17 change log opened DeepSeek's in-house search engine as an API. Search is POST https://api.deepseek.com/v1/tools/search and Fetch is POST https://api.deepseek.com/v1/tools/fetch. Both take the same bearer key as the rest of the DeepSeek API and return raw JSON.

Each endpoint costs $0.75 per 1,000 successful requests at DeepSeek's list prices. A request is billed only when it returns HTTP 200 with a result, so failed or empty calls cost nothing. Both endpoints allow 5 queries per second per account, and Fetch also has a per-domain limit.

Search takes a query of 1 to 128 characters and returns 10 results by default, adjustable from 1 to 50. Fetch takes a URL of up to 8,192 characters and returns the page as Markdown.

Server-side tool calls through DeepSeek Harness, Claude Code and Codex are free until billing starts on October 31, 2026 at 16:00 UTC. That is November 1, 00:00 Beijing time.

It is DeepSeek's own API, sold from api.deepseek.com.

Pros and cons

Pros

• Matched V4 Pro's answers on every task both models could run.
• Faster on all four shared tasks, by 1.3 to 14 times.
• 4 to 11 times cheaper per task at DeepSeek's list prices.
• Reads images, which V4 Pro does not.
• Drop-in for existing deepseek-flash code.
• Open weights under MIT.

Cons

• Thinks long on hard tasks, at 19,734 thinking tokens for our parser.
• DeepSeek's own table still shows Claude Opus 5 ahead on HLE, Terminal-Bench 4.0, ProgramBench and NL2Repo-Bench.
• Every improvement number comes from DeepSeek.

Which one for which job

V4 Pro is still on sale as of October 17, 2026, with its own prices and a 500 concurrency limit against 2,500 for Flash. Keep it where your own tests show a win.

Call it through the API

from openai import OpenAI

client = OpenAI(api_key="YOUR_OMNIAKEY_KEY", base_url="https://api.omniakey.com/v1")

stream = client.chat.completions.create(
    model="deepseek-flash",           # serves DeepSeek-V4.1-Flash-1017
    messages=[{"role": "user", "content": "Write a Python function that parses ISO 8601 durations."}],
    max_tokens=32000,                 # room for long answers
    stream=True,                      # long thinking runs can outlast proxy timeouts
)
for chunk in stream:
    delta = chunk.choices[0].delta if chunk.choices else None
    if delta and delta.content:
        print(delta.content, end="")

1. Create a key in the OmniaKey dashboard and paste it in place of YOUR_OMNIAKEY_KEY.
2. Send the request above with deepseek-flash. Keep max_tokens near 32,000 for coding work and stream long runs.
3. Log the returned token counts and the model name on the first real task, then compare the cost against DeepSeek's off-peak rates.

Common questions

Did DeepSeek raise the price with 1017?

DeepSeek's October 17 change log says prices did not change. The pricing page that day lists Flash at $0.15 per million cache-miss input tokens and $0.60 per million output tokens off-peak. It shows current rates only.

Can I run 1017 on my own hardware?

The weights are on Hugging Face under MIT. The release ships reference Python encoding code instead of a Jinja chat template, so a local serving stack needs its own prompt encoding.

Does 1017 read images through the API?

Yes. DeepSeek's pricing page marks vision as supported on Flash and not on V4 Pro, and our chart task ran on Flash only.

Update log

• October 17, 2026: first published, with tests run the same day.