DeepSeek V4.1 Flash 1017 Ships MIT Weights and a 77% DeepSWE Score at About $0.20 a Task, but Claude Opus 5 Still Wins 6 of DeepSeek’s 11 Benchmarks
DeepSeek made V4.1 Flash 1017 its default web, app and API model on 17 Oct 2026 with MIT weights and unchanged prices. We read its benchmarks, the $0.20 a task chart, peak hours in IST and the harness version catch.
DeepSeek pushed a new build of its smallest current model on 17 October 2026. DeepSeek-V4.1-Flash-1017 is now the model behind the DeepSeek website, app and API, and the weights are on Hugging Face under the MIT license.
The headline claim is a 77.0 percent pass rate on the DeepSWE v1.1 coding benchmark for roughly $0.20 per task, according to DeepSeek’s own cost chart. The company also says API pricing has not changed.
We went through the release note, the model card, the pricing page, the change log and the npm registry. The gains over the September build are large, but the comparison table leaves out the newest US flagships, and Claude Opus 5 still beats it on 6 of the 11 benchmarks DeepSeek chose to publish.
What DeepSeek actually shipped
The model card calls 1017 a “minor update” to DeepSeek-V4.1-Flash, which launched on 10 September. DeepSeek says it improves reasoning, coding, agent and vision results while cutting task time and output length.
The DeepSeek change log says “the API calling method and pricing remain unchanged”. You call it as deepseek-flash. The old names deepseek-v4-flash and deepseek-v4-flash-vision-exp still work, but the pricing page says those requests are now served by 1017 and billed at the Flash price.
That second point matters if you pinned a legacy name in production. Your app is already running on the new model whether you planned a migration or not.
On X, the official @deepseek_ai thread also opened DeepSeek’s Search and Fetch APIs to everyone. More on that below.
DeepSeek’s benchmark table, read line by line
All numbers below come from DeepSeek. We found no independent results for the 1017 build when we checked, and Artificial Analysis had no page for it yet.
| Benchmark | V4.1 Flash 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) | 51.0 | 49.0 | 41.0 | n/a | 53.0 | 52.0 |
Source: DeepSeek model card and release note, 17 Oct 2026. *Text-only HLE subset (DeepSeek scores 44.0 on that subset). Terminal-Bench 4.0 used the Claude Code harness, DeepSWE used Mini-SWE-V2 and ZeroBench used Pass@5.
The jump from the September build is real on paper. Terminal-Bench 4.0 rises 11.1 points, ProgramBench nearly 9 and Apex Bench 7.5.
Against the rivals DeepSeek picked, 1017 leads on four agent and coding tests: DeepSWE, Automation-Bench, Agents’ Last Exam and Apex Bench. Claude Opus 5 leads on six, including Terminal-Bench, ProgramBench and NL2Repo. GPT 5.6 Sol takes ZeroBench.
The $0.20 a task chart, and who is missing
DeepSeek’s second chart plots DeepSWE v1.1 pass rate against average cost per task. Read from that chart, 1017 at max effort lands at 77.0 percent for about $0.20 a task, and its low setting sits above 72 percent for a little over $0.10.
On the same chart, Claude Opus 5 tops out between 73 and 74 percent at roughly $10 or more a task, and GPT-6.1 Sol peaks around 75 percent at well under $1. The fine print says the rival points come from the public DeepSWE leaderboard with data dated 22 September, plus OpenAI’s official chart for the GPT-6 family.
That date is the catch. Anthropic’s newer Claude Opus 5.5 appears in neither the table nor the chart. GPT-6.1 Sol is on the chart but not in the table, where DeepSeek compares against the older GPT 5.6 Sol.
The 1017 release also reports no security numbers. The September V4.1 Flash launch listed CyberGym 88.1, SEC-Bench Pro 62.8 and ExploitGym 15.3. This update is silent on all three.
Confirmed vs not yet known
| Claim | Status |
|---|---|
1017 is live on web, app and API as deepseek-flash |
Confirmed (DeepSeek change log and pricing page) |
| API prices unchanged from the September build | Confirmed by DeepSeek |
| MIT license for code and weights | Confirmed (Hugging Face model card) |
| 77.0 on DeepSWE v1.1 at about $0.20 per task | Vendor reported, read from DeepSeek’s chart |
| Beats Claude Opus 5.5 or Gemini 4 Argon | Not tested in DeepSeek’s materials |
| Security benchmark results for 1017 | Not published |
| Independent benchmark results | None found at publication time |
Pricing in USD and INR, and when peak hours hit
The DeepSeek pricing page lists these per million tokens for deepseek-flash. INR figures use 95.11 rupees to the dollar.
| Per 1M tokens | Peak | Off-peak |
|---|---|---|
| Input, cache hit | $0.006 (about ₹0.57) | $0.003 (about ₹0.29) |
| Input, cache miss | $0.30 (about ₹28.5) | $0.15 (about ₹14.3) |
| Output | $1.20 (about ₹114) | $0.60 (about ₹57) |
Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, excluding Chinese public holidays. Weekends are off-peak in full. We covered how this split started in our DeepSeek V4 Pro peak and off-peak pricing report.
| Location | Weekday peak windows (local time) |
|---|---|
| India (IST) | 6:30 to 9:30 AM and 11:30 AM to 3:30 PM |
| New York and Toronto (EDT) | 9 PM to midnight (Sunday to Thursday) and 2 AM to 6 AM |
| Sydney (AEDT) | Noon to 3 PM and 5 PM to 9 PM |
North American daytime work falls entirely in the cheaper window, at least until US clocks change on 1 November. Indian and Australian teams pay double for part of the working day.
The context window is 1M tokens and maximum output is 384K, per the pricing page. The concurrency limit for Flash is 2,500.
Third-party hosts already list the build. On OpenRouter on 17 October, providers included DeepSeek at $0.15 in and $0.60 out, Cloudflare at $0.10 and $0.60, and DeepInfra from $0.112 and $0.336 in FP8. One provider caps context at 262K, so check before you route long jobs.
Search and Fetch APIs open up
The same release opened DeepSeek’s in-house Search and Fetch APIs. Each costs $0.75 per 1,000 successful requests, or about ₹71. Search returns up to 50 results per query, and Fetch returns one page as Markdown, according to the Search and Fetch guide.
Server-side search calls made through DeepSeek Harness, Claude Code or Codex on the DeepSeek API are free for now. Billing starts at 16:00 UTC on 31 October 2026, which is 9:30 PM IST that day.
The harness version trap
DeepSeek tells users to run DeepSeek Harness 0.2.1-rc.1 or later for the best results. The npm registry shows the default latest tag still points to 0.2.0-rc.2. Version 0.2.1-rc.1 went up at 03:17 UTC on 17 October and 0.2.1-rc.2 followed at 08:36 UTC under the next tag.
So a plain install gets you the older harness. Use npm i -g @deepseek-ai/dsh@next if you want the version DeepSeek tuned for.
Running the weights yourself
Hugging Face lists about 763 billion stored parameters and roughly 510 GB of files, with FP8 weights and FP4 experts. The September card describes a 552B backbone plus a 196B Engram memory, activating 8B parameters per token in prefill and 16B in decode. That keeps per-token compute low, but you still have to store and load the full checkpoint.
There is no Jinja chat template. DeepSeek ships a Python reference encoder and a separate deepseek-recipe toolkit instead, and its inference example uses eight-way tensor parallelism and calls itself “a readable reference implementation rather than a production serving engine”. The download is bigger than Reflection’s smallest Beam build but smaller than Meta’s Muse Spark Open, which needs about a terabyte of Mac memory to run locally.
What it means for Indian developers
For most Indian teams the API is the practical route. A job with 10M uncached input tokens and 2M output tokens costs about $5.40 (around ₹514) at peak and $2.70 (around ₹257) off-peak. Moving batch work to evenings or weekends halves the bill.
The MIT license lets startups fine-tune and ship the weights commercially without a usage cap. Before switching a production coding agent, run your own repos through it. DeepSeek’s numbers are strong, but every one of them is self-reported for now.
FAQ
Is DeepSeek V4.1 Flash 1017 free?
The weights are free to download under the MIT license. The API is paid, starting at $0.15 per million uncached input tokens off-peak and $0.003 for cache hits.
Do I need to change my API code?
No. DeepSeek says the calling method is unchanged. Requests to deepseek-flash and the legacy V4 Flash names are already served by 1017.
Does it beat Claude Opus 5.5?
Unknown. DeepSeek compared it with Claude Opus 5, not Opus 5.5, and Opus 5 still wins six of the eleven benchmarks in DeepSeek’s table.
How big is the download?
About 510 GB on Hugging Face. DeepSeek’s reference inference example runs it with eight-way tensor parallelism.
When do Search and Fetch calls start costing money inside Claude Code or Codex?
At 16:00 UTC on 31 October 2026, which is 9:30 PM IST. After that, each endpoint costs $0.75 per 1,000 successful requests.
1 comment