By Thorsten Meyer | October 18, 2026 | ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
DeepSeek’s October update makes its small model better and cheaper per task. DeepSeek’s own chart also shows where it still falls short.
On October 17, 2026, DeepSeek shipped DeepSeek-V4.1-Flash-1017, a “minor update” to the V4.1 Flash model it released on September 10. It replaces the September version on DeepSeek’s website, its app and its API. Anyone calling the `deepseek-flash` model now gets 1017 automatically, at an unchanged price. The weights are open under the MIT license.
DeepSeek reports gains on all eleven benchmarks it published. The biggest are in long, multi-step agent work: Terminal-Bench 4.0 rose from 31.2 to 42.3, ProgramBench from 20.3 to 29.1. Outputs are also shorter, so tasks finish faster and cost less.
That is the headline. Two things complicate it. Every number comes from DeepSeek, and none has been independently reproduced. And DeepSeek compared its model against rivals that are a generation old.
What DeepSeek shipped
- Model: DeepSeek-V4.1-Flash-1017, served as `deepseek-flash`. The older names `deepseek-v4-flash` and `deepseek-v4-flash-vision-exp` are also routed to it now.
- Architecture: a sparse Mixture of Experts with native image input. According to the technical report, as summarised by Kingy AI, it is DeepSeek’s first model built on a Causal Encoder-Decoder (CED) design: 40 layers, split 20/20 between encoder and decoder, with 384 routed experts, 6 of them active per token, plus one shared expert.
- Size: a 552B backbone with about 8B active parameters during prefill and 16B during decode. Hugging Face lists 763B, counted differently.
- Context: 1M tokens, with up to 384K tokens of output on the hosted API.
- Memory: the technical report is titled Pushing the Limits of KV Cache Compression. The KV cache is reported at about 890 bytes per token, under 1 GB for a full million-token context.
- Weights: MIT-licensed on Hugging Face and ModelScope, as an FP8 checkpoint with FP4 experts.
DeepSeek doesn’t say what changed in training. The Decoder reports that the gains come from post-training with reinforcement learning with self-feedback. In that method the model scores its own outputs from its confidence or its own critiques, instead of relying on human raters or a separate reward model. DeepSeek also says it tuned 1017 for its own agent harness, DeepSeek Harness (version 0.2.1-rc.1 or later recommended), and for Tencent’s CodeBuddy and WorkBuddy.
As an affiliate, we earn on qualifying purchases.
The benchmark table, read honestly
These are DeepSeek’s figures from the model card. They put 1017 next to the September version and the four rivals DeepSeek chose:
| Benchmark | 1017 | V4.1 Flash (Sep) | Kimi K3 | GLM-5.3 | GPT-5.6 Sol | Claude Opus 5 |
|---|---|---|---|---|---|---|
| HLE (no tools) | 42.6 | 36.8 | 43.5 | 42.0 | 44.5 | 56.3 |
| Terminal-Bench 4.0 | 42.3 | 31.2 | 12.6 | 37.9 | 39.9 | 51.8 |
| DeepSWE v1.1 | 77.0 | 74.2 | 67.5 | 66.9 | 73.0 | 74.0 |
| ProgramBench | 29.1 | 20.3 | 17.5 | 19.0 | 23.0 | 37.0 |
| NL2Repo-Bench | 68.2 | 65.4 | 58.0 | 58.0 | 56.8 | 75.3 |
| Automation-Bench v1.0.6 | 60.5 | 54.8 | 46.7 | 48.8 | 45.8 | 50.3 |
| Agents’ Last Exam | 35.4 | 31.8 | 27.6 | 28.5 | 26.7 | 28.6 |
| Apex Bench v1.1 | 72.1 | 64.6 | 47.8 | 56.6 | 54.3 | 65.8 |
| Chartography (tools) | 83.0 | 78.9 | 68.1 | n/a | 79.9 | 84.0 |
| BabyVision (tools) | 92.5 | 89.6 | 85.7 | n/a | 88.9 | 94.1 |
| ZeroBench-main (tools, Pass@5) | 51.0 | 49.0 | 41.0 | n/a | 53.0 | 52.0 |
Bold marks the best score in each row. DeepSeek’s HLE scores are on the full set; on the text-only subset they are 44.0 (1017) and 39.1 (September), and GLM-5.3’s 42.0 is a text-only score. Terminal-Bench ran in the Claude Code harness, DeepSWE in Mini-SWE-V2.
Three things stand out:
- The gains over September are real but uneven. Terminal-Bench rose 11.1 points, ProgramBench 8.8 and Apex Bench 7.5. DeepSWE, the coding benchmark DeepSeek leads with, gained only 2.8. This update mostly improves long agent runs, not raw reasoning.
- It is not the strongest model in its own chart. 1017 leads 4 of 11 rows. Claude Opus 5 leads 6, including reasoning (HLE 56.3 vs 42.6), ProgramBench, NL2Repo and two of the three vision tests. GPT-5.6 Sol leads the third.
- The rivals are a generation old. DeepSeek compared against GPT-5.6 Sol and Opus 5, not GPT-6.1 Sol, GPT-6 Astra, Opus 5.5 or Gemini 4 Argon. On Terminal-Bench 4.0, figures reported separately for the current frontier models sit between 57 and 66, according to Kingy AI. They come from different evaluations, so they aren’t a direct match, but the gap is large.
Infographic · DeepSeek V4.1 Flash 1017
Better on every benchmark, a fraction of the cost per task, and still behind the frontier.
DeepSeek’s October update improves all eleven scores it published over the September version. Against the four rivals DeepSeek picked, it leads four benchmarks and trails seven. Its real advantage is the price of a finished task.
1017 vs the September version
Score per benchmark (0–100), sorted by gain. Hover a row for the exact values.
- Terminal-Bench 4.042.331.2+11.1
- ProgramBench29.120.3+8.8
- Apex Bench v1.172.164.6+7.5
- HLE (no tools)42.636.8+5.8
- Automation-Bench60.554.8+5.7
- Chartography83.078.9+4.1
- Agents' Last Exam35.431.8+3.6
- BabyVision92.589.6+2.9
- DeepSWE v1.177.074.2+2.8
- NL2Repo-Bench68.265.4+2.8
- ZeroBench-main51.049.0+2.0
Against the best rival per benchmark
Points ahead of or behind the strongest of Kimi K3, GLM-5.3, GPT-5.6 Sol and Claude Opus 5 on each row.
- Automation-Benchbest rival: Claude Opus 5 50.3+10.2
- Agents' Last Exambest rival: Claude Opus 5 28.6+6.8
- Apex Bench v1.1best rival: Claude Opus 5 65.8+6.3
- DeepSWE v1.1best rival: Claude Opus 5 74.0+3.0
- Chartographybest rival: Claude Opus 5 84.0-1.0
- BabyVisionbest rival: Claude Opus 5 94.1-1.6
- ZeroBench-mainbest rival: GPT-5.6 Sol 53.0-2.0
- NL2Repo-Benchbest rival: Claude Opus 5 75.3-7.1
- ProgramBenchbest rival: Claude Opus 5 37.0-7.9
- Terminal-Bench 4.0best rival: Claude Opus 5 51.8-9.5
- HLE (no tools)best rival: Claude Opus 5 56.3-13.7
Cost per DeepSWE task
USD per task, linear scale to $10, as read from DeepSeek’s chart.
- 1017, low effortabove 72% of DeepSWE solvedabout $0.10
- 1017, default77.0% of DeepSWE solvedjust under $0.20
- V4.1 Flash (September)74.2% of DeepSWE solvedabout $0.50
- Frontier models on the chartGPT-6 Astra, Gemini 3.8 Flash, Claude Opus 5$2 to $10+
None of the frontier models on DeepSeek’s chart reached 77%.
API price per 1M tokens
| Token type | Off-peak | Peak |
|---|---|---|
| Input, cache miss | $0.15 | $0.30 |
| Input, cache hit | $0.003 | $0.006 |
| Output | $0.60 | $1.20 |
Peak: weekdays 01:00–04:00 and 06:00–10:00 UTC, excluding Chinese public holidays.
All scores are DeepSeek’s own and not independently reproduced; rivals are the ones DeepSeek chose (GPT-5.6 Sol and Claude Opus 5, not the current GPT-6 and Opus 5.5 models). Sources: DeepSeek model card, DeepSeek API pricing, cost per task via The Decoder (17–18 Oct 2026).
Powered by Thorsten Meyer AI
As an affiliate, we earn on qualifying purchases.
Where it wins: cost per finished task
The case for 1017 isn’t leaderboard rank. It’s economics. On DeepSeek’s DeepSWE chart, as reported by The Decoder, 1017 reaches 77.0% at just under $0.20 per task. A low-effort setting still clears 72% at about $0.10. The September model needed about $0.50 per task for 74.2%. The frontier models on the same chart cost between $2 and more than $10 per task.
The official price list makes the gap concrete:
| Per 1M tokens | Off-peak | Peak |
|---|---|---|
| Input, cache miss | $0.15 | $0.30 |
| Input, cache hit | $0.003 | $0.006 |
| Output | $0.60 | $1.20 |
Peak hours are 01:00–04:00 and 06:00–10:00 UTC on weekdays, excluding Chinese public holidays. That window falls in the European morning. Batch jobs scheduled outside it pay half. A job with 1M uncached input tokens and 200K output tokens costs $0.27 off-peak. At list prices of $2/$10 per million tokens, the same job costs $4.
An early independent pilot by Kingy AI supports the cost story and adds nuance. Over seven small tasks, each run twice, 1017 passed 14 of 14, as did GPT-6.1 Sol and Claude Sonnet 5.5, at roughly a quarter of their cost. It kept unsupported facts marked as unknown and ignored instructions injected into the input. It was not always fastest, though: one task took about 27 seconds against about 3 seconds for Sonnet. And because three models scored perfectly, the pilot can’t rank them.
As an affiliate, we earn on qualifying purchases.
DeepSeek moves up the stack: Search and Fetch
The second announcement may matter as much as the model. DeepSeek opened its own search engine to developers as two agent tools. Search returns real-time web results in many languages and filters out SEO spam and low-quality AI-generated content. Fetch retrieves pages.
- Price: $0.75 per 1,000 successful requests for each endpoint.
- Free period: server-side calls through DeepSeek Harness, Claude Code or Codex are free until October 31, 2026, 16:00 UTC (November 1, 00:00 Beijing time). Calls made directly to the endpoints are billed from the start.
For agent builders this means model and retrieval come from one provider. In tight agent loops, though, per-call fees can add up faster than the token bill.
As an affiliate, we earn on qualifying purchases.
What to do if you run it
- Pin what you depend on. `deepseek-flash` is a rolling alias. It changed behaviour overnight on October 17 and will do so again.
- Re-run your own evaluations. Results measured on September’s V4.1 Flash don’t carry over to 1017, and DeepSeek’s scores depend heavily on the harness. I made the same point after testing 71 model configurations on my own news pipeline: the benchmark winner is rarely the best choice for a specific workload.
- Schedule batch work off-peak. Moving it out of the peak windows halves the bill.
- Keep the reasoning trace in tool calls. In thinking mode the API expects `reasoning_content` to be passed back; dropping it can cause errors.
- If you self-host, budget for the full model. There’s no ready-made Jinja chat template, only DeepSeek’s reference encoder and its `deepseek-recipe` toolkit. A small KV cache doesn’t make a small model: 16B active parameters still means hundreds of billions to load.
- Watch the policy risk. The Decoder reports that China is considering tighter export controls on the most capable open-weight models. That could affect future releases more than this one.
Verdict
DeepSeek-V4.1-Flash-1017 doesn’t take the frontier crown, and DeepSeek’s own table shows it. What it offers is agentic coding near frontier level at a tenth of the cost per task or less, with open weights and now search built for agents.
For high-volume agent work where cost per finished task decides, it is the model to test this month. For the hardest reasoning and vision work, the closed frontier models still lead. That is why the current frontier models remain in my own October stack for that kind of work. Test 1017 on your own tasks before you switch.
Sources
- DeepSeek: API changelog (October 17, 2026)
- DeepSeek: API pricing
- DeepSeek: DeepSeek-V4.1-Flash-1017 model card on Hugging Face
- DeepSeek: DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression (technical report)
- DeepSeek: Announcement on X
- The Decoder: Deepseek V4.1 Flash 1017 beats larger rivals in agentic tasks
- Kingy AI: DeepSeek V4.1 Flash 1017 Tested: Specs, Benchmarks & Price
- Bloomberg: DeepSeek pushes out updated AI model it says is faster, leaner
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
