DeepSeek V4.1 Flash 1017: Reviewing A Lower-Cost Near-Frontier Model

🔍 Read the full analysis: DeepSeek V4.1 Flash 1017: Reviewing A Lower-Cost Near-Frontier Model on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on monitors, keyboards and dev gear

• Business-only prices and quantity discounts
• Tax-exempt purchasing
• Multiple users, one account, clear invoices

Create a free Business account →Free to register

As an affiliate, we earn on qualifying purchases.Hide

DeepSeek released V4.1 Flash 1017 on October 17, replacing the September version across its website, app and API while keeping token prices unchanged. The company says the update improves results on all 11 benchmarks it published and makes some agent tasks cheaper, though the scores are company-reported and have not been independently reproduced at scale.

DeepSeek describes 1017 as a minor update to V4.1 Flash, first released September 10. The API name deepseek-flash now routes to the new version automatically; DeepSeek says the older names deepseek-v4-flash and deepseek-v4-flash-vision-exp also route to it. The model’s weights are available under the MIT license.

In DeepSeek’s model card, the largest gains over September appear on longer agent and coding tasks. Terminal-Bench 4.0 rose from 31.2 to 42.3, and ProgramBench from 20.3 to 29.1. DeepSWE v1.1, a coding benchmark highlighted by DeepSeek, rose more modestly from 74.2 to 77.0. These are score-point changes in the published results, not independently verified measures of real-world performance.

The model card also reports shorter outputs, which DeepSeek says can reduce completion time and cost. Its published API rates are $0.15 per million uncached input tokens off-peak and $0.30 at peak; output costs $0.60 off-peak and $1.20 at peak. Cached input is cheaper. DeepSeek’s stated peak windows are 01:00–04:00 and 06:00–10:00 UTC on weekdays, excluding Chinese public holidays.

Lower Bills for Agent Work

The release’s practical case rests on cost per completed task, not on leading every benchmark. The Decoder reported that, on DeepSeek’s DeepSWE chart, 1017 reached a 77.0% score at just under $0.20 per task; a low-effort setting scored above 72% at about $0.10. The September model’s 74.2% result was shown at about $0.50 per task. These figures come from DeepSeek’s evaluation and depend on its task and pricing assumptions.

If those economics hold in a user’s workload, teams running frequent coding or computer-use agents could try more tasks for the same budget. DeepSeek also recommends its own DeepSeek Harness version 0.2.1-rc.1 or later and says it tuned the update for Tencent’s CodeBuddy and WorkBuddy. That makes the release relevant to developers evaluating both the model and the surrounding agent setup.

Kingy AI described an early independent pilot involving seven small tasks, each run twice. It reported that 1017, GPT-6.1 Sol and Claude Sonnet 5.5 each passed all 14 runs, with 1017 costing roughly a quarter as much. The pilot also found that 1017 was not always fastest: one task took about 27 seconds, compared with about three seconds for Sonnet. The small sample supports a limited cost observation, but cannot establish broad reliability or rank the three models.

Products Worth Considering

Samsung 49-Inch Odyssey OLED G9 (G91SD) Dual QHD Curved Gaming Monitor

AmazonView Latest Price

Samsung Galaxy Tab A11+ 11-Inch Wi-Fi Tablet, 6GB RAM, 128GB Storage, Gray

AmazonView Latest Price

EcoFlow DELTA 3 Portable Power Station, 1024Wh, 1800W

AmazonView Latest Price

As an affiliate, we earn on qualifying purchases.

A Benchmark Table With Limits

DeepSeek’s comparison table places 1017 beside its September model and four rivals: Kimi K3, GLM-5.3, GPT-5.6 Sol and Claude Opus 5. In that table, 1017 leads four of 11 rows; Claude Opus 5 leads six, and GPT-5.6 Sol leads one. Opus 5 scores 56.3 on HLE, compared with 42.6 for 1017, and leads on ProgramBench and NL2Repo-Bench as well as two of the three vision tests.

The comparisons have qualifications. DeepSeek’s HLE result uses the full set, while the GLM-5.3 figure shown is text-only; DeepSeek gives text-only scores of 44.0 for 1017 and 39.1 for the September version. Terminal-Bench was run in the Claude Code harness, while DeepSWE used Mini-SWE-V2. The table also does not include newer models such as GPT-6.1 Sol, GPT-6 Astra, Claude Opus 5.5 or Gemini 4 Argon. Kingy AI has cited separate Terminal-Bench results of 57 to 66 for current frontier models, but those evaluations are not directly comparable.

The technical report, as summarized by Kingy AI, describes a sparse mixture-of-experts model with image input and a one-million-token context. It reportedly has a 552-billion-parameter backbone, with about eight billion active during prefill and 16 billion during decoding; Hugging Face lists 763 billion under a different counting method. DeepSeek has not specified what changed in training. The Decoder attributes the gains to post-training with reinforcement learning and self-feedback, but that account is not a detailed explanation from DeepSeek in the supplied materials.

““minor update””

— DeepSeek

Independent Results Remain Sparse

DeepSeek’s benchmark results remain unverified by a broad, independent evaluation. The company has not explained in the supplied materials what changed in training, and its comparison set does not include several newer models. Differences in harnesses, task sets and scoring also limit conclusions from comparing figures across sources.

The Kingy AI pilot is an early, small test of seven tasks, not evidence that the reported pass rate or cost advantage will generalize. It also cannot settle speed comparisons: one reported task took 27 seconds for 1017 versus about three seconds for Claude Sonnet 5.5. The available information does not establish how the update performs across larger workloads, different agent frameworks or sustained production use.

Testing in Real Workflows

Developers can access the updated model through DeepSeek’s website, app and API, or download its MIT-licensed weights from Hugging Face and ModelScope. The next useful evidence will come from independent evaluations that publish task sets, harnesses, token use, latency and full cost calculations, including comparisons with current models.

Teams considering a deployment can compare 1017 with their existing model on representative tasks and record completion quality, retries, latency and total cost. DeepSeek has not announced a date for a further update or an independent validation program in the supplied materials, so those milestones remain unknown.

Key Questions

When did DeepSeek release V4.1 Flash 1017?

DeepSeek released it on October 17, 2026. It replaces the September version on the company’s website, app and API.

Did DeepSeek raise its API prices?

No price change was reported. DeepSeek’s published rates vary by token type and time: uncached input is $0.15 per million tokens off-peak and $0.30 at peak, while output is $0.60 and $1.20, respectively.

Are the benchmark improvements independently confirmed?

Not at broad scale. The 11 benchmark results are from DeepSeek’s own model card. Kingy AI reported a small pilot, but its seven-task sample does not independently reproduce the full benchmark table.

Does 1017 lead DeepSeek’s comparison table?

It leads four of the 11 listed benchmark rows. Claude Opus 5 leads six and GPT-5.6 Sol leads one, according to DeepSeek’s table.

What is still unknown about the update?

DeepSeek has not detailed the training changes behind 1017 in the supplied materials. Its performance across broader production workloads, newer-model comparisons and different agent setups also remains to be established.

Source: ThorstenMeyerAI.com