DeepSeek V4.1 Flash 1017: Comparing Cost And Agent Capability

🔍 Read the full analysis: DeepSeek V4.1 Flash 1017: Comparing Cost And Agent Capability 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 version across its website, app and API while keeping the listed price unchanged. The company reports higher results on all 11 benchmarks in its model card, with larger gains on some multi-step agent tests, but the scores are company-reported and have not been independently reproduced.

The new version is served under the API name deepseek-flash; older names, including deepseek-v4-flash, also route to it, according to the source report. DeepSeek describes the release as a minor update. Its weights are available under the MIT license, and the model accepts image input as well as text.

In DeepSeek’s published comparison, Terminal-Bench 4.0 increased from 31.2 to 42.3, while ProgramBench rose from 20.3 to 29.1. The increase on DeepSWE v1.1 was smaller, from 74.2 to 77.0. These are benchmark-point changes, not percentages of improvement. The company says outputs are shorter, which can reduce completion time and token charges, but the source material does not provide an independent measurement of that effect.

The API price list gives rates per million tokens: $0.15 off-peak and $0.30 at peak for uncached input, with output at $0.60 and $1.20, respectively. Cached input is listed at $0.003 off-peak and $0.006 at peak. The source report says peak periods are specified in UTC on weekdays, excluding Chinese public holidays, and batch jobs outside peak hours cost half. Actual bills depend on token use and the applicable rate.

A small independent pilot by Kingy AI, as described in the source report, tested seven tasks twice. DeepSeek 1017, GPT-6.1 Sol and Claude Sonnet 5.5 each passed all 14 runs, with the pilot reporting 1017 at roughly one-quarter of the other models’ cost. The sample is too small to rank the models, and one task took about 27 seconds on 1017 versus about three seconds on Sonnet, according to the report.

Lower Costs for Long Agent Runs

The release matters most to teams that run coding agents repeatedly, where a model’s cost per completed task can matter more than its headline score. DeepSeek’s chart, as reported by The Decoder, puts 1017 at 77% on DeepSWE for just under $0.20 per task; a lower-effort configuration reportedly scored above 72% at about $0.10. The September model was shown at 74.2% for about $0.50 per task. These figures come from DeepSeek’s evaluation and should not be treated as a universal cost for every workload.

The trade-off is capability. DeepSeek’s table shows 1017 leading in four of 11 rows, while Claude Opus 5 leads six and GPT-5.6 Sol leads one. Opus 5 scores higher on the table’s reasoning, ProgramBench and NL2Repo-Bench tests. For developers, the update is evidence that a lower-priced model may be useful for some agent workflows, not proof that it matches stronger models across tasks. The cost advantage could matter for high-volume use, while reliability, speed and performance on a particular task still need testing.

What DeepSeek Changed in October

DeepSeek released the prior V4.1 Flash version on September 10, 2026. The October 17 update supersedes that version in the company’s consumer products and API, rather than introducing a separately priced service. DeepSeek recommends its own DeepSeek Harness, version 0.2.1-rc.1 or later, and says the model was tuned for that system as well as Tencent’s CodeBuddy and WorkBuddy.

The model card describes a sparse Mixture of Experts architecture and a one-million-token context window. A technical report summarized by Kingy AI describes a Causal Encoder-Decoder design, with 40 layers split between encoder and decoder. The model’s parameter count varies by accounting method: the report describes a 552-billion-parameter backbone, while Hugging Face lists 763 billion. Those figures should not be read as directly comparable without the counting definitions.

DeepSeek has not said what training changes produced the benchmark gains. The Decoder attributes them to post-training with reinforcement learning and self-feedback, but that account is not a detailed release explanation from DeepSeek in the supplied material. DeepSeek’s comparison also uses selected rival models, and the source report says it does not include newer releases such as GPT-6.1 Sol, GPT-6 Astra, Opus 5.5 or Gemini 4 Argon.

“The 1017 version reaches 77.0% on DeepSWE v1.1 at just under $0.20 per task.”

— The Decoder, reporting on DeepSeek’s cost chart

Benchmark Limits and Open Questions

Independent reproduction is lacking for the 11 benchmark results published by DeepSeek. The tests also use different harnesses in at least two cases: Terminal-Bench ran in Claude Code, while DeepSWE used Mini-SWE-V2. Results from separate evaluations can be difficult to compare, especially when prompt settings, tools and scoring procedures differ.

The age and selection of competitors are further limits. DeepSeek’s chart compares 1017 with models including GPT-5.6 Sol and Claude Opus 5, not the newer versions named in the source report. Kingy AI’s separate figures for current frontier models on Terminal-Bench are reported as roughly 57 to 66, but they come from different evaluations and are not a direct comparison.

Several implementation details remain undisclosed or unsettled. DeepSeek has not publicly explained the training changes behind the gains in the material provided. The reported architecture and parameter totals use differing counting methods, and the independent pilot covered only seven small tasks, each run twice. The source also mentions a separate DeepSeek search-tools announcement but cuts off before describing its functions, so its features and availability cannot be confirmed here.

Testing in Real Agent Workflows

The next useful evidence will come from independent, repeatable evaluations that compare 1017 with current models using the same tasks, tools, settings and cost accounting. Larger trials would help establish whether the lower per-task figures persist outside DeepSeek’s chosen benchmark and agent setup, and whether shorter outputs translate into faster completion across different workloads.

Developers can now access the revised model through DeepSeek’s website, app and API, according to the source report. Teams evaluating it should check the current model routing, token rates and harness compatibility, then measure performance on their own tasks. The release does not establish how long the September model will remain available under separate routing, or whether DeepSeek will publish additional training details or independent test results.

Key Questions

What is DeepSeek V4.1 Flash 1017?

It is DeepSeek’s October 17 update to V4.1 Flash. The company routed its website, app and API service to the new version, including calls using the deepseek-flash name, according to the source report.

Did DeepSeek raise the API price?

No price increase is reported. The source says the update replaced the September model at an unchanged listed price. Rates vary by input or output tokens, cache status and peak or off-peak timing.

Which benchmarks improved the most?

In DeepSeek’s published table, Terminal-Bench 4.0 rose from 31.2 to 42.3, and ProgramBench from 20.3 to 29.1. DeepSWE v1.1 increased from 74.2 to 77.0, a smaller gain.

Have independent tests verified DeepSeek’s benchmark claims?

The 11 benchmark results in the model card have not been independently reproduced in the supplied material. Kingy AI reported a separate pilot of seven tasks run twice, but its sample was small and did not distinguish 1017 from two competing models that also passed all runs.

Is 1017 the top model in DeepSeek’s comparison?

No. DeepSeek’s table gives 1017 the lead on four of 11 benchmarks. Claude Opus 5 leads six rows and GPT-5.6 Sol leads one. The selected competitors are not the newest models named in the source report.

Source: ThorstenMeyerAI.com