DeepSeek V4.1 Flash 1017 keeps the September checkpoint's per-token price and API surface, and by DeepSeek's own chart it cuts the cost of a DeepSWE task from about $0.50 to just under $0.20 while the score rises from 74.2% to 77.0%. Terminal-Bench 4.0 goes from 31.2 to 42.3. All of these are DeepSeek's numbers; nobody has reproduced them yet.
The engineering problem is that the upgrade ships as a checkpoint swap, not as a new API. On October 17, 2026, DeepSeek pointed deepseek-flash at 1017, so anything calling DeepSeek directly has already moved, and the September checkpoint can't be pinned. On AIHubMix, 1017 has its own ID, deepseek-v4.1-flash-1017, so you choose when to switch. Either way, three things need checking on your own workload:
- which checkpoint actually answered;
- whether shorter outputs break anything downstream;
- what a finished task costs now, not what a token costs.
This post covers what changed, a working request through AIHubMix, how to confirm the checkpoint, the cost math, how AIHubMix and OpenRouter compare on price, and the failure modes worth testing for.
What changed, in one table
| Item | September V4.1 Flash | V4.1 Flash 1017 |
|---|---|---|
| Architecture | 552B MoE, 8B / 16B active | Same |
| Context / max output | 1M / 384K | Same |
| Input | Text and images | Same |
| Thinking default | On, effort high | Same |
| Effort levels | low / high / max | Same |
| DeepSeek API name | deepseek-flash | deepseek-flash (now 1017) |
| List price | Unchanged | Unchanged |
| Output length | Long | Shorter (DeepSeek, not quantified) |
The Hugging Face model card calls it "a minor update", and DeepSeek's release note calls it "a Pareto improvement in performance and cost". The architecture and API are untouched; the post-training is new, described by DeepSeek as "reinforcement learning with self-feedback to improve end-to-end project development".
The legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp still resolve on DeepSeek's API, and both now run on 1017 at Flash prices.
The benchmark deltas
Figures come from DeepSeek's model card and changelog, compiled the same way by The Decoder and Kingy.ai; nobody has reproduced them independently.
| Benchmark | September | 1017 | Change |
|---|---|---|---|
| Terminal-Bench 4.0 | 31.2 | 42.3 | +11.1 |
| ProgramBench | 20.3 | 29.1 | +8.8 |
| Apex Bench v1.1 | 64.6 | 72.1 | +7.5 |
| HLE, no tools | 36.8 | 42.6 | +5.8 |
| Automation-Bench | 54.8 | 60.5 | +5.7 |
| Agents' Last Exam | 31.8 | 35.4 | +3.6 |
| DeepSWE v1.1 | 74.2 | 77.0 | +2.8 |
| NL2Repo-Bench | 65.4 | 68.2 | +2.8 |
| BabyVision, with tools | 89.6 | 92.5 | +2.9 |
Terminal-Bench 4.0 ran in the Claude Code harness; DeepSWE in mini-SWE-v2.
Read it by workload. Long terminal sessions and building a program from scratch moved the most, which matches the training DeepSeek describes. Repository-level bug fixing (DeepSWE, NL2Repo) and vision moved two to three points. If your agent mostly patches existing repos, expect a smaller step than the headline number.
Prerequisites
- An AIHubMix API key exported as
AIHUBMIX_API_KEY. The AIHubMix quick start walks through creating one. - The OpenAI Python SDK (
pip install openai), pointed athttps://aihubmix.com/v1. - The model ID
deepseek-v4.1-flash-1017. AIHubMix serves it on Chat Completions, the Responses API, and Anthropic Messages; the sample below uses Chat Completions.
The request
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)
reasoning_effort="high" is the default anyway; it's set explicitly so the level shows up in your code review. Thinking is on by default, and reasoning tokens count toward max_tokens, so leave room. The model card recommends at least 256K for long agent runs.
Confirm which checkpoint answered
AIHubMix ran a set of requests against deepseek-v4.1-flash-1017 on October 19, 2026, one run per check:
- Every response's
modelfield readdeepseek-v4-1-flash-261017. Logresp.model: that string is how you know the 1017 checkpoint answered. -
reasoning_effortvalueslow,high, andmaxwere all accepted with HTTP 200. - The response included
reasoning_content, andusage.completion_tokens_details.reasoning_tokensreported the thinking share of the output, which is the number to watch when you compare effort levels. - Chat Completions, the Responses API (
/v1/responses), and Anthropic Messages (/v1/messages) all returned 200.
Cost: the token price is fixed, the token count isn't
| Model | DeepSeek input | DeepSeek output | AIHubMix input | AIHubMix output |
|---|---|---|---|---|
| V4.1 Flash (September) | $0.30 | $1.20 | $0.3098 | $1.2392 |
| V4.1 Flash 1017 | $0.30 | $1.20 | $0.3098 | $1.2392 |
| V4 Pro 0813 | $1.32 | $3.96 | $1.3944 | $4.1832 |
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 running a limited-time 30% discount on 1017; the figures here use list prices.
Since the price per token is identical, any saving has to come from fewer tokens. The September checkpoint was verbose: when Artificial Analysis ran its Intelligence Index on it at max effort, it generated about 250M output tokens against a 130M median, and 62% more than DeepSeek's V4 Pro 0813. DeepSeek's DeepSWE chart, 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.
To turn that into a per-session number, here is an estimate from stated assumptions, because no published test gives per-task token counts for 1017: 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:
| Model | Peak | Off-peak |
|---|---|---|
| V4.1 Flash 1017 | $0.118 | $0.059 |
| V4 Pro 0813 | $0.469 | $0.233 |
1017, peak: 0.12M × $0.3098 + 1.08M × $0.0062 + 0.06M × $1.2392 = $0.118
The same token counts cost exactly the same on the September model. What changes is the token count itself: if 1017 needs fewer turns and shorter answers for the same fix, the bill shrinks with them. Running batch work off-peak 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 and reserve max for the hardest tasks.
AIHubMix vs OpenRouter
OpenRouter serves 1017 as deepseek/deepseek-v4.1-flash-1017 through 15 providers, so its price depends on which one answers. The like-for-like row is DeepSeek's own endpoint:
| Route | Input | Output | Cache hit |
|---|---|---|---|
| OpenRouter, DeepSeek endpoint | $0.30 | $1.20 | $0.006 |
| AIHubMix, DeepSeek channel | $0.3098 | $1.2392 | $0.0062 |
| AIHubMix, ByteDance, Oct 19–25 | $0.1971 | $0.7885 | $0.0039 |
Peak prices per 1M tokens, checked October 19, 2026; both halve off-peak on the same schedule. AIHubMix serves 1017 through DeepSeek, ByteDance, and Alibaba Cloud channels. The ByteDance channel lists at $0.2816 / $1.1264, is 30% off October 19–25 (UTC) with the discount applied automatically to deepseek-v4.1-flash-1017, and adds a cache storage fee of $0.0024 per 1M tokens per hour. AIHubMix has no top-up fee; OpenRouter adds no markup to DeepSeek's price but charges 5.5% when you buy credits by card.
Through October 25 that makes AIHubMix about 34% cheaper per token, or 38% once OpenRouter's card fee is counted. At list price, the DeepSeek channel is 3.3% higher per token but about 2% cheaper after the card fee, and the ByteDance channel is 6% lower per token. For the session above: $0.075 against $0.114 at peak ($0.121 with the fee), and $0.038 against $0.057 off-peak.
OpenRouter won't route to DeepSeek's endpoint by default, though. Without provider preferences it picks among stable providers weighted by the inverse square of price, and several third-party hosts undercut every peak price above:
| Host on OpenRouter | Weights | Input | Output | Session |
|---|---|---|---|---|
| Relace | Not stated | $0.05 | $0.10 | $0.023 |
| DeepInfra | fp8 | $0.098 | $0.294 | $0.033 |
| Wafer | mxfp4 | $0.10 | $0.50 | $0.045 |
| Morph | fp8 | $0.043 | $0.60 | $0.047 |
| InferenceNet | fp8 | $0.08 | $0.50 | $0.048 |
Flat rates per 1M tokens around the clock, checked October 19, 2026. "Weights" is the quantization OpenRouter lists. Off-peak, the discounted ByteDance channel ($0.0986 / $0.3942, $0.038 per session) beats Wafer, Morph, and InferenceNet; Relace and DeepInfra stay cheaper.
Check two things before routing production traffic there: output quality on quantized weights, since DeepSeek's benchmarks weren't run on these hosts, and the output cap, since DeepInfra and InferenceNet stop at 131K tokens on OpenRouter, below the 256K the model card recommends for long agent runs. To pin DeepSeek's endpoint instead, pass "provider": {"only": ["deepseek"]} in the request body (extra_body in the OpenAI SDK).
Failure modes and fixes
A parser that worked on September output starts failing
Shorter outputs can break parsers and evaluation checks written around the old verbosity. Before moving production prompts with strict output formats, run your regression suite against deepseek-v4.1-flash-1017 and diff the failures against your September baseline.
Your "September" traffic on DeepSeek is already 1017
On DeepSeek's own API, deepseek-flash is a rolling alias, and deepseek-v4-flash and deepseek-v4-flash-vision-exp now run on 1017 too. There is no pinned name for September. If you need reproducible behaviour, call a dated model ID; on AIHubMix that's deepseek-v4.1-flash-1017, and log resp.model so a change would show up.
Answers come back short or cut off at a low max_tokens
Thinking is on by default and its tokens count toward max_tokens. A budget sized for a non-thinking model leaves little room for the answer. Keep it generous (the sample uses 32,000), go to 256K or more for long agent runs as the model card recommends, and watch reasoning_tokens in the usage block.
The bill doesn't drop the way the chart suggests
Two usual causes. Traffic that lands in the weekday peak window (01:00–04:00 and 06:00–10:00 UTC) pays twice the off-peak rate, so move batch jobs out of it. And max effort lengthens agent runs by 1.6–1.8× by DeepSeek's own measure; if an agent framework sets it globally, drop it back to high and raise it per task.
Where it is not the right model
- Knowledge-heavy questions. HLE without tools is 42.6, well behind Claude Opus 5 (56.3) and below Claude Haiku 5.5 (45.9, Anthropic's figure).
- Long, hard shell sessions. 42.3 on Terminal-Bench 4.0 is a big jump, but Sonnet 5.5 (70.6) and Opus 5.5 (66.4) are in a different league there.
- Workloads that need a frozen model on DeepSeek's API, for the alias reason above.
For context on the comparisons DeepSeek chose: inside its own lineup, 1017 has passed V4 Pro 0813 on every agentic benchmark DeepSeek publishes for both (Terminal-Bench 4.0: 42.3 vs 12.4; DeepSWE: 77.0 vs 62.7). Against Claude Opus 5 it leads on DeepSWE, Automation-Bench, Agents' Last Exam, and Apex, while Opus 5 leads on Terminal-Bench 4.0 (51.8), ProgramBench (37.0), NL2Repo (75.3), and HLE (56.3). Opus 5.5, Sonnet 5.5, and the GPT-6 generation are not in DeepSeek's table; Google separately reports Opus 5.5 at 66.4 on Terminal-Bench 4.0.
What hasn't been verified
- Every benchmark number above is DeepSeek's own. No third party has reproduced them.
- How much shorter the outputs are. DeepSeek hasn't published a figure, and the per-task costs are The Decoder's reading of a vendor chart, not a measurement.
- Per-task token counts for 1017. The session cost above is an estimate from assumptions.
- Effort cost through AIHubMix. The October 19 test used trivial prompts, and the reasoning token counts (68, 39, and 54 for low, high, and max) don't separate the levels. It confirms the parameter passes through, nothing more.
- Quality on OpenRouter's quantized hosts. No published test compares their output with DeepSeek's endpoint.
Decision rules
- 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 anything gated by an evaluation suite.
- Stay put or look elsewhere: knowledge-heavy Q&A and hard debugging that already needs a frontier model.
The AIHubMix compare page shows the two Flash checkpoints' specs and prices side by side.
Keep reading: the DeepSeek V4.1 Flash 1017 series
- Switching an existing tool loop? The
reasoning_content400, the alias swap, and peak-hour billing each have their own fix in the migration guide: Migrating to DeepSeek V4.1 Flash 1017: The reasoning_content 400, the Silent Alias Swap, and Peak-Hour Billing - If Terminal-Bench is the number you care about, this runs 1017 and Anthropic's cheapest model on the same terminal coding task: DeepSeek V4.1 Flash 1017 vs Claude Haiku 5.5: Which Cheap Model for Terminal Coding Agents?
- Running V4 Pro 0813 next to the Flash? This one checks whether Pro still earns its 3.3× output price: 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)
Top comments (0)