Before You Switch AI Models, Compare DeepSeek V4.1 Flash 1017

🔍 Read the full analysis: Before You Switch AI Models, Compare DeepSeek V4.1 Flash 1017 on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on office and shipping supplies

• 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 model on its website, app and API while keeping the listed price unchanged. The update matters to developers weighing low-cost models for agent and coding work: DeepSeek reports gains across all 11 benchmarks in its model card, with the largest increases in two multi-step task tests, though the results have not been independently reproduced.

The new release is served under the API name deepseek-flash; the older names deepseek-v4-flash and deepseek-v4-flash-vision-exp also route to it. DeepSeek describes 1017 as a minor update to V4.1 Flash, first released September 10. Its weights are available under the MIT license, according to the source material, and the model supports image input alongside text.

In DeepSeek’s model-card table, Terminal-Bench 4.0 rises from 31.2 for the September model to 42.3 for 1017, while ProgramBench increases from 20.3 to 29.1. DeepSWE v1.1, a coding benchmark, moves from 74.2 to 77.0. Those are score-point changes reported by DeepSeek, not independently verified measurements. The source says Terminal-Bench ran in the Claude Code harness and DeepSWE in Mini-SWE-V2.

DeepSeek also says the model produces shorter outputs, which can reduce completion time and token charges. Its published API rates remain $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 listed at $0.003 off-peak and $0.006 at peak. The off-peak rate applies outside weekday peak hours, which DeepSeek lists as 01:00–04:00 and 06:00–10:00 UTC, excluding Chinese public holidays.

Lower Costs for Agent Tasks

The update’s practical case rests on cost per completed task, not a claim that it leads every capability test. The source reports that DeepSeek’s DeepSWE chart puts 1017 at 77% for just under $0.20 per task; a low-effort setting reaches more than 72% at about $0.10. The September model reached 74.2% at about $0.50 per task in that comparison. These figures are attributed to DeepSeek’s chart as reported by The Decoder.

For teams that run many coding or agent tasks, lower per-task expense could make experimentation and routine automation less costly. But benchmark scores do not establish that a model will complete a particular company’s work reliably, and the listed token prices do not capture every cost, such as orchestration or review. A sample calculation in the source puts one million uncached input tokens and 200,000 output tokens at $0.27 off-peak using DeepSeek’s rates, compared with $4 at list prices of $2 input and $10 output per million tokens. That is an illustrative price comparison, not a guarantee of savings on a real workload.

There is limited outside evidence in the supplied material. Kingy AI reports an early pilot of seven small tasks, each run twice, in which 1017, GPT-6.1 Sol and Claude Sonnet 5.5 each passed all 14 runs. The pilot found 1017 cost about a quarter as much as the other two models, but its small task set and tied results cannot rank the models. It also found that 1017 was slower on one task, taking about 27 seconds compared with about three seconds for Sonnet.

Popular right now

Ninja Luxe Café Mini Espresso and Drip Coffee Machine, Stainless Steel, ES301

AmazonView Latest Price

Dyson Supersonic Nural Hair Dryer

AmazonView Latest Price

TCL 75-Inch QM6K QLED Mini LED 4K HDR Smart Google TV, 144Hz

AmazonView Latest Price

As an affiliate, we earn on qualifying purchases.

A Narrow Benchmark Comparison

DeepSeek’s chart compares 1017 with its September release and four named rivals: Kimi K3, GLM-5.3, GPT-5.6 Sol and Claude Opus 5. In that table, 1017 leads four of 11 rows, while Opus 5 leads six and GPT-5.6 Sol leads one. Opus 5’s listed HLE score is 56.3, against 42.6 for 1017; it also leads the table on ProgramBench and NL2Repo-Bench. On three tool-based vision tests, 1017 leads two, while Opus 5 leads one.

The scope of those comparisons matters. DeepSeek’s HLE figure uses the full set, while its text-only result is 44.0; the table’s GLM-5.3 score of 42.0 is for the text-only subset. The source also says several newer models were not included in DeepSeek’s chosen comparison. Kingy AI cites separate Terminal-Bench results of 57 to 66 for current frontier models, but says those evaluations differ, so they are not a direct comparison with DeepSeek’s chart.

The model’s technical description points to a sparse Mixture of Experts design with a one-million-token context. The technical report, as summarized by Kingy AI, describes a Causal Encoder-Decoder architecture with 40 layers and 384 routed experts, six active per token, plus one shared expert. Reported parameter counts differ by counting method: the source gives a 552-billion-parameter backbone with about 8 billion active during prefill and 16 billion during decode, while Hugging Face lists 763 billion. The weights are described as an FP8 checkpoint with FP4 experts.

““minor update””

— DeepSeek, in its model card

What the Results Cannot Confirm

DeepSeek supplied all 11 benchmark results in its model card, and the source says none has been independently reproduced. The scores therefore show what the company reports under its stated setups, rather than a verified ranking across comparable conditions. The separately reported frontier-model results use different evaluations and should not be treated as direct head-to-head scores.

DeepSeek has not said what changed in training for 1017. The Decoder reports that the gains come from post-training with reinforcement learning and self-feedback, in which a model assesses its outputs using confidence or its own critiques. That explanation is attributed to The Decoder, not confirmed in the supplied material by a detailed DeepSeek account. It is also unclear how the reported benchmark gains translate to different agent harnesses, coding repositories or production workloads.

The source mentions a second announcement: DeepSeek opened its search engine to developers through two agent tools. However, the supplied account ends before describing the tools, their behavior, availability or pricing. Those details remain unconfirmed in the available material.

Testing the Update in Practice

Developers can evaluate 1017 through DeepSeek’s website, app or API, where the existing deepseek-flash alias now points to the October release. DeepSeek recommends its own DeepSeek Harness version 0.2.1-rc.1 or later, and says it tuned the model for that harness and for Tencent’s CodeBuddy and WorkBuddy. Teams comparing providers can run the same representative tasks in their own environment, record completion rates and latency, and calculate cost from actual token use.

The next useful evidence would be independent benchmark runs with comparable harnesses and task settings, along with more detail from DeepSeek about the training changes and the announced search tools. Until those arrive, the confirmed development is an available model update with unchanged listed rates; the size and reliability of its reported gains remain open questions.

Key Questions

When did DeepSeek release V4.1 Flash 1017?

DeepSeek released V4.1 Flash 1017 on October 17, 2026, as an update to the version released September 10.

Did DeepSeek change its API prices?

The supplied source says the update kept the listed price unchanged. DeepSeek lists input, cached input and output rates that vary by whether usage falls in its weekday peak hours.

Are the benchmark improvements independently verified?

No independent reproduction is identified in the supplied source for DeepSeek’s 11 model-card results. Treat the scores as company-reported results until comparable independent evaluations are available.

Does 1017 outperform every model in DeepSeek’s table?

No. DeepSeek’s table shows 1017 leading four of 11 benchmark rows. Claude Opus 5 leads six, and GPT-5.6 Sol leads one.

What is known about DeepSeek’s new search tools?

The source says DeepSeek opened its search engine to developers as two agent tools, but the supplied account is cut off before it describes their features, access terms or pricing.

Source: ThorstenMeyerAI.com