A Price-and-Performance Guide To DeepSeek V4.1 Flash 1017

🔍 Read the full analysis: A Price-and-Performance Guide To DeepSeek V4.1 Flash 1017 on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get the latest gadgets delivered free with Prime

• Fast, free delivery on millions of items
• Prime Video, Amazon Music and more included
• Member-only deals all year

Start your free Prime trial →Free trial for eligible customers · Cancel anytime

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 listed prices unchanged. The company reports higher scores on all 11 benchmarks it published and lower costs for some completed agent tasks, but the benchmark results have not been independently reproduced.

The release is a minor update to DeepSeek’s V4.1 Flash model, first released on September 10. Calls to deepseek-flash now route to version 1017, as do the older names deepseek-v4-flash and deepseek-v4-flash-vision-exp. DeepSeek says the update produces shorter outputs, which can reduce task time and token costs. Its published API rates remain $0.15 per million input tokens off-peak for uncached input and $0.60 per million output tokens, with higher rates during peak hours.

In DeepSeek’s model-card chart, Terminal-Bench 4.0 increased from 31.2 to 42.3 between the September model and 1017, while ProgramBench rose from 20.3 to 29.1. DeepSWE v1.1, a coding benchmark, moved more modestly, from 74.2 to 77.0. These are company-reported scores; the chart does not provide independent verification. DeepSeek’s chosen competitor set also excludes newer models cited in other evaluations.

The model is available with open weights under the MIT license, according to the source material, and accepts image input. Its technical description lists a one-million-token context window and up to 384,000 tokens of hosted API output. DeepSeek’s report describes a sparse Mixture of Experts design. The company has not explained in detail what changed in training for this update.

Lower Costs for Long Agent Tasks

The update’s clearest practical case is cost per completed task, rather than a claim that it leads every benchmark. The Decoder’s account of DeepSeek’s DeepSWE figures says version 1017 reached 77.0% at just under $0.20 per task, while a lower-effort configuration passed 72% at about $0.10. The September model reached 74.2% at about $0.50 per task in the same reported comparison. Those numbers suggest a potentially useful price-performance option for developers running many coding or agent jobs, but they are not an independent, universal estimate of cost.

DeepSeek lists different rates for peak and off-peak usage. Its stated weekday peak periods are 01:00–04:00 and 06:00–10:00 UTC, excluding Chinese public holidays; batch jobs outside those hours are charged at half the listed rate. For buyers, the effective bill will depend on input and output volume, cache use, scheduling and whether a task needs retries or human review. Price per token alone does not establish price per successful task.

A small pilot reported by Kingy AI offers a limited external check. Across seven small tasks, each run twice, 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 reported that 1017 marked unsupported facts as unknown and ignored instructions injected into input. Because all three models scored perfectly on this small sample, it cannot rank their overall capabilities; one task took 1017 about 27 seconds versus about three seconds for Sonnet.

What DeepSeek’s Chart Measures

DeepSeek released the earlier V4.1 Flash model on September 10, then shipped 1017 on October 17. The new version automatically replaced it on DeepSeek’s services, so users of the default model name do not need to select a separate release. The company recommends its own DeepSeek Harness version 0.2.1-rc.1 or later and says it tuned the model for that harness and Tencent’s CodeBuddy and WorkBuddy.

The published results cover different tasks, including reasoning, coding, automation and vision. DeepSeek’s chart shows 1017 leading four of 11 rows; Claude Opus 5 leads six and GPT-5.6 Sol leads one. For HLE, DeepSeek’s figures use the full set, while some comparison figures use a text-only subset. Terminal-Bench was run in the Claude Code harness and DeepSWE in Mini-SWE-V2. Scores across different tasks and setups should not be treated as one overall ranking.

The source material says the model’s weights are available as an FP8 checkpoint with FP4 experts on Hugging Face and ModelScope. It also describes a 552-billion-parameter backbone, with about 8 billion parameters active during prefill and 16 billion during decoding; Hugging Face lists 763 billion using a different counting method. Those architecture details describe the model, but do not independently establish its measured performance.

“Terminal-Bench 4.0: 42.3 for 1017, compared with 31.2 for the September version.”

— DeepSeek’s published benchmark chart

Benchmark Limits and Open Questions

No independent reproduction of DeepSeek’s 11 benchmark results is provided, and the company has not disclosed what training changes produced the reported gains. The competitor list in its chart does not include some newer models named in the source material. Separate Terminal-Bench figures for current frontier models are reported between 57 and 66 by Kingy AI, but those evaluations use different setups and cannot be directly compared with DeepSeek’s chart.

Task-cost figures also need qualification: the reported DeepSWE results come from a specific benchmark and setup. They do not show how costs compare across a broad mix of production workloads, where prompt length, caching, retries, task success and review requirements can change the total. The seven-task pilot is small, and its tied scores do not establish a general performance ranking.

Details of the training method remain unsettled in the available accounts. The Decoder reports that the gains may come from post-training with reinforcement learning and model self-feedback, but DeepSeek’s release material, as summarized here, does not specify what changed. The model’s stated parameter totals also differ by counting method, so those figures should not be read as directly equivalent.

What Developers Should Track

The next useful evidence would be independent evaluations using the same harnesses, prompts and scoring rules for 1017 and current competing models. Repeated tests on larger task sets could show whether the reported gains hold outside DeepSeek’s own charts and whether lower token use translates into reliably lower spending per successful outcome.

Developers can also check the live API rates and model documentation before deploying, since peak pricing, cache hits and output volume affect bills. DeepSeek has replaced the earlier model under the default endpoint, but the source material gives no date for a further update or a detailed account of the training changes. For now, the confirmed development is a model refresh with unchanged listed prices and company-reported performance gains.

Key Questions

When did DeepSeek release V4.1 Flash 1017?

DeepSeek shipped V4.1 Flash 1017 on October 17, 2026. It replaces the September 10 version on the company’s website, app and API.

Did DeepSeek change the API price?

The source material says listed prices were unchanged. Actual charges depend on peak or off-peak timing, whether input is cached, and the number of input and output tokens.

Are the reported benchmark improvements independently verified?

No independent reproduction of DeepSeek’s 11 published benchmark results is provided in the source material. The scores should be treated as company-reported results.

Is 1017 the top-scoring model in DeepSeek’s chart?

No. DeepSeek’s chart shows 1017 leading four of 11 benchmarks; Claude Opus 5 leads six and GPT-5.6 Sol leads one. The chart compares a selected set of models and does not include some newer competitors.

Can developers download the model weights?

The source material says the weights are available under the MIT license on Hugging Face and ModelScope. It describes an FP8 checkpoint with FP4 experts.

Source: ThorstenMeyerAI.com