DeepSeek V4.1 Flash 1017: Same Price, Shorter Outputs, and AIHubMix vs OpenRouter Pricing

On Saturday, October 17, 2026, DeepSeek swapped the model behind its deepseek-flash API name for a new checkpoint, DeepSeek-V4.1-Flash-1017. The price did not move, the context window did not move, and the calling method did not move. What changed is the model: DeepSeek reports double-digit gains on terminal and program-building benchmarks, and says the new checkpoint finishes tasks faster with shorter outputs.

The short answer: if you run coding agents or tool-heavy workflows on the September V4.1 Flash, switch. You pay the same per token and, by DeepSeek's own cost chart, far less per finished task. If your workload is knowledge-heavy Q&A, the case is weaker, and frontier models still lead there.

On AIHubMix, the new checkpoint has its own ID, deepseek-v4.1-flash-1017, so nothing changes under you until you change the model name. Through October 25 its ByteDance channel is also 30% off, about a third below DeepSeek's own endpoint on OpenRouter; at peak hours, OpenRouter's quantized third-party hosts go lower still. The full price comparison is below.

What changed and what stayed the same

DeepSeek's release note calls 1017 "a Pareto improvement in performance and cost", and the Hugging Face card calls it "a minor update". Both are fair: the architecture and the API surface are untouched, and the post-training is new (DeepSeek describes "reinforcement learning with self-feedback to improve end-to-end project development").

The old names deepseek-v4-flash and deepseek-v4-flash-vision-exp still work on DeepSeek's API, and both now run on 1017 at Flash prices. DeepSeek offers no way to pin the September checkpoint, so anyone calling deepseek-flash directly has already been moved.

Where the gains landed

All figures below come from DeepSeek's model card and changelog; nobody has reproduced them independently yet. The Decoder and Kingy.ai have compiled the same table and flag the same caveat.

Terminal-Bench 4.0 was run in the Claude Code harness; DeepSWE in mini-SWE-v2.

The pattern is clear: the biggest jumps are in long terminal sessions and in building a program from scratch, which matches what DeepSeek says it trained for. Repository-level bug fixing (DeepSWE, NL2Repo) and vision moved by two to three points.

Shorter outputs are the real upgrade

The September checkpoint was expensive to run for a cheap model because it talked a lot. When Artificial Analysis ran its Intelligence Index on it at max effort, it generated about 250M output tokens against a median of 130M, and 62% more than DeepSeek's own V4 Pro 0813.

1017 targets exactly that. DeepSeek's DeepSWE chart (cost per task on the x-axis, as read by The Decoder) puts the September model at 74.2% for about $0.50 a task, 1017 at its top setting at 77.0% for just under $0.20, and 1017 at its low setting above 72% for about $0.10. Since the price per token is identical, nearly all of that saving comes from the model using fewer tokens and finishing in fewer turns.

Two cautions. DeepSeek has not said how much shorter the outputs are, and these per-task costs are read from a vendor chart, not measured by a third party. Run your own workload before you plan around a 60% saving.

How close it gets to the tier above

Inside DeepSeek's lineup, the tier above is V4 Pro 0813, and 1017 has already passed it on every agentic benchmark DeepSeek publishes for both (Terminal-Bench 4.0: 42.3 vs 12.4; DeepSWE: 77.0 vs 62.7). That gets its own post in this series.

Against frontier models the picture is mixed. DeepSeek's table beats Claude Opus 5 on DeepSWE, Automation-Bench, Agents' Last Exam, and Apex, but Opus 5 still leads on Terminal-Bench 4.0 (51.8), ProgramBench (37.0), NL2Repo (75.3), and HLE (56.3). DeepSeek also compared against older models: Opus 5.5, Sonnet 5.5, and the GPT-6 generation are not in its table, and Google separately reports Opus 5.5 at 66.4 on Terminal-Bench 4.0.

Price and real cost

Peak prices per 1M tokens, cache miss. Peak hours are 01:00–04:00 and 06:00–10:00 UTC on weekdays; every other hour, weekends, and Chinese public holidays are off-peak at half price. Flash cache hits cost $0.006 on DeepSeek and $0.0062 on AIHubMix at peak. AIHubMix is currently running a limited-time 30% discount on 1017; the figures here use list prices.

One agent session, costed. No published test gives per-task token counts for 1017, so this example uses stated assumptions: a bug-fix session of 40 model calls, 1.2M input tokens in total with 90% served from cache, and 60K output tokens including thinking. At AIHubMix list prices:

Arithmetic for 1017 at peak: 0.12M × $0.3098 + 1.08M × $0.0062 + 0.06M × $1.2392 = $0.118.

The same token counts cost the same on the September model, because the price is identical. The difference shows up in the token counts themselves: if 1017 needs fewer turns and shorter answers for the same fix, the bill shrinks with them. Scheduling batch work outside peak hours halves it again.

Effort is the other lever. DeepSeek's V4.1 Flash paper maps the API's low, high, and max tiers to internal effort values of 50, 75, and 100, and reports that the last step to max "lengthens agent trajectories by 1.6–1.8× for only marginal improvements". Start agents at high, which is the default, and reserve max for the hardest tasks.

AIHubMix vs OpenRouter: what 1017 costs on each

OpenRouter lists 1017 as deepseek/deepseek-v4.1-flash-1017 and serves it through 15 providers, from DeepSeek itself to hosts running quantized copies, so there is no single OpenRouter price. Two comparisons matter: DeepSeek's own endpoint, which is like for like, and the cheapest hosts.

Against DeepSeek's endpoint on OpenRouter

Peak prices per 1M tokens, checked October 19, 2026; both routes halve them off-peak on the same weekday schedule. AIHubMix serves 1017 through DeepSeek, ByteDance, and Alibaba Cloud channels. The ByteDance channel lists at $0.2816 / $1.1264 and is 30% off from October 19 to October 25 (UTC); the discount applies automatically when you call deepseek-v4.1-flash-1017. That channel also bills cache storage at $0.0024 per 1M tokens per hour, which adds well under $0.001 to the agent session above. AIHubMix charges no top-up fee. OpenRouter passes DeepSeek's price through without markup but charges 5.5% when you buy credits by card.

• Through October 25, AIHubMix is about a third cheaper. The ByteDance channel is 34% below DeepSeek's endpoint on OpenRouter per token, and 38% below once OpenRouter's card fee is counted.
• At list price, AIHubMix is still slightly cheaper. The DeepSeek channel is 3.3% higher per token, but with no top-up fee against OpenRouter's 5.5%, it works out about 2% cheaper. The ByteDance channel is 6% lower per token before any fee.
• For the agent session above, that is $0.075 on AIHubMix against $0.114 on OpenRouter's DeepSeek endpoint at peak ($0.121 with the card fee), and $0.038 against $0.057 off-peak.

The cheaper hosts on OpenRouter

DeepSeek's endpoint is not where OpenRouter sends you by default. With no provider settings, its routing picks among stable providers weighted by the inverse square of price, so an unpinned request usually lands on a cheaper host. Here is what the main hosts charge; the first five undercut every peak price above:

Prices per 1M tokens, checked October 19, 2026. These hosts charge the same rate around the clock. "Weights" is the quantization OpenRouter lists; the session is the same 40-call bug fix as above.

Off-peak, the gap narrows. AIHubMix's discounted ByteDance channel costs $0.0986 / $0.3942, and its $0.038 session comes in below Wafer, Morph, and InferenceNet; only Relace and DeepInfra stay cheaper.

The trade-off is what you are running. fp8 and mxfp4 are compressed versions of the model's weights, and DeepSeek's benchmark figures were not measured on any of these hosts, so test quality on your own workload before moving traffic. DeepInfra and InferenceNet also cap output at 131K tokens on OpenRouter, below the 256K the model card recommends for long agent runs. To stay on DeepSeek's endpoint, set "provider": {"only": ["deepseek"]} in the request, and the first table applies.

In short: against DeepSeek's own endpoint, AIHubMix is about a third cheaper through October 25 and still slightly cheaper after that. The lowest token prices are on OpenRouter's quantized third-party hosts, if their output holds up on your workload.

Where it is not the best choice

• Knowledge-heavy questions. HLE without tools is 42.6, well behind Opus 5 (56.3) and below Claude Haiku 5.5 (45.9, Anthropic's figure). Raw knowledge is still the Flash line's weak spot.
• Long, hard shell sessions. Terminal-Bench 4.0 at 42.3 is a big jump, but Sonnet 5.5 (70.6) and Opus 5.5 (66.4) are in a different league on that benchmark.
• Workloads that need a frozen model. On DeepSeek's own API, deepseek-flash is a rolling alias and the September checkpoint cannot be pinned. If you need reproducible behaviour, call a dated model ID.
• Traffic that all lands in peak hours. Off-peak pricing is half of peak; a workload stuck in the 01:00–10:00 UTC weekday window pays the full rate.

Who should switch, who should wait

• Switch now: coding agents, terminal automation, and tool-calling pipelines on the September Flash. Same price, better scores, and fewer tokens per task by DeepSeek's measure.
• Switch after a regression run: production prompts with strict output formats or evaluation suites. Shorter outputs can break parsers that expected the old verbosity.
• Stay put or look elsewhere: knowledge-heavy Q&A and hard debugging that already needs a frontier model. 1017 does not change that.

Try it on AIHubMix

AIHubMix serves 1017 on Chat Completions, the Responses API, and Anthropic Messages. Create a key and follow the quick start, then call it with the openai SDK:

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["AIHUBMIX_API_KEY"],
    base_url="https://aihubmix.com/v1",
)

resp = client.chat.completions.create(
    model="deepseek-v4.1-flash-1017",
    messages=[{"role": "user", "content": "Write a Python function that merges overlapping intervals."}],
    reasoning_effort="high",
    max_tokens=32000,
)
print(resp.choices[0].message.content)

In a test run on October 19, AIHubMix accepted low, high, and max for reasoning_effort, and every response's model field read deepseek-v4-1-flash-261017, which is how you can confirm the 1017 checkpoint answered. Thinking is on by default and the reasoning tokens count toward max_tokens, so keep it generous; the model card recommends at least 256K for long agent runs. To see both checkpoints side by side, open the AIHubMix compare page.

FAQ

Is DeepSeek V4.1 Flash 1017 a new model or an update?
It is an updated checkpoint of V4.1 Flash with the same architecture, context window, and API surface. DeepSeek calls it a minor update; the benchmark gains come from new post-training.

Did the price change with 1017?
No. DeepSeek kept the Flash price at $0.30 input and $1.20 output per 1M tokens at peak, half that off-peak. AIHubMix lists it at $0.3098 and $1.2392 at peak, with a limited-time 30% discount running now.

Is DeepSeek V4.1 Flash 1017 cheaper on AIHubMix or OpenRouter?
It depends on which OpenRouter provider serves the request. Against DeepSeek's own endpoint on OpenRouter, $0.30 input and $1.20 output per 1M tokens at peak, AIHubMix's ByteDance channel is about 34% cheaper through October 25. At list price AIHubMix is still about 2% cheaper, since it charges no top-up fee and OpenRouter charges 5.5% on card purchases. At peak hours, third-party hosts on OpenRouter that run fp8 or mxfp4 weights charge less than both.

What model ID do I use on AIHubMix?
Use deepseek-v4.1-flash-1017. The deepseek-flash name belongs to DeepSeek's own API, where it now points to 1017.

How much shorter are the outputs?
DeepSeek has not published a number. Its DeepSWE chart implies the cost per task fell from about $0.50 to under $0.20 at the same price per token, which points to much shorter runs, but that is a vendor figure.

Which reasoning effort should I use?
Start with high, the default. Use low for simple, latency-sensitive calls and max only for the hardest tasks, since DeepSeek's own paper says max lengthens agent runs by 1.6 to 1.8 times for small gains.

Is 1017 better than Claude Opus 5?
On some agentic benchmarks DeepSeek reports it ahead, such as DeepSWE and Automation-Bench. Opus 5 still leads on Terminal-Bench 4.0, ProgramBench, NL2Repo, and HLE, and newer frontier models were not part of DeepSeek's comparison.

Keep reading: the DeepSeek V4.1 Flash 1017 series

• Moving an existing integration over? The migration guide covers the reasoning_content 400, the silent alias swap, peak-hour billing, and coding-tool setup: Migrating to DeepSeek V4.1 Flash 1017: The reasoning_content 400, the Silent Alias Swap, and Peak-Hour Billing
• Weighing it against Anthropic's cheapest model for agent work? The head-to-head runs both on a terminal coding task: DeepSeek V4.1 Flash 1017 vs Claude Haiku 5.5: Which Cheap Model for Terminal Coding Agents?
• Still paying for V4 Pro? This post checks whether the Flash has made it redundant: DeepSeek V4.1 Flash 1017 vs V4 Pro: Is the Flash Enough Now?

Sources

• DeepSeek-V4.1-Flash-1017 release note (DeepSeek)
• DeepSeek-V4.1-Flash-1017 model card (Hugging Face)
• DeepSeek V4.1 Flash 1017 on AIHubMix (AIHubMix)
• Deepseek V4.1 Flash 1017 beats larger rivals in agentic tasks (The Decoder)
• DeepSeek V4.1 Flash 1017 on OpenRouter (OpenRouter)