DeepSeek V4.1 Flash 1017: Agent Capabilities At One-Tenth The Price

🔍 Read the full analysis: DeepSeek V4.1 Flash 1017: Agent Capabilities At One-Tenth The Price on ThorstenMeyerAI.com

DeepSeek released DeepSeek-V4.1-Flash-1017 on October 17, replacing the September version across its website, app and API while keeping prices unchanged. The company says the update improves results on all 11 benchmarks it published, with its largest gains on some multi-step agent tasks; those figures have not been independently reproduced.

The new model is served under the deepseek-flash API name, and DeepSeek says older names, including deepseek-v4-flash and deepseek-v4-flash-vision-exp, also route to it. The weights are available under the MIT license. The model supports image input, and DeepSeek lists a context window of one million tokens, with up to 384,000 output tokens on its hosted API.

In the company’s model-card table, Terminal-Bench 4.0 rises to 42.3 from 31.2 for the September version, while ProgramBench increases to 29.1 from 20.3. Other gains are smaller: DeepSWE v1.1 moves from 74.2 to 77.0. DeepSeek says the update also produces shorter outputs, which can reduce completion time and token charges, but it does not provide independent measurements of that effect.

DeepSeek’s listed API rates are $0.15 per million uncached input tokens off-peak and $0.30 at peak; cached input is $0.003 and $0.006, respectively. Output costs $0.60 off-peak and $1.20 at peak. The company defines peak weekday hours as 01:00–04:00 and 06:00–10:00 UTC, excluding Chinese public holidays. Rates and task costs depend on usage, cache status and scheduling, so a low token price does not by itself establish the cost of a finished job.

DeepSeek’s chart does not show 1017 leading every comparison. It leads four of the 11 listed rows; Claude Opus 5 leads six, and GPT-5.6 Sol leads one. The table includes selected competitor versions and benchmark setups, rather than a complete or independently verified comparison of current models.

Lower API Prices, Narrower Benchmark Gains

The update’s immediate importance is the combination of low listed token rates and improved scores on certain agent benchmarks. For developers running repeated coding or workflow tasks, lower input and output charges may make experimentation and routine use less expensive. Actual savings depend on prompt length, output size, caching, peak pricing and whether the model completes a task successfully.

Performance claims need a separate reading. DeepSeek’s largest reported improvements are in Terminal-Bench and ProgramBench, while its DeepSWE score rises by 2.8 points. The chart also shows the model trailing on several tasks, including HLE reasoning and ProgramBench relative to Claude Opus 5. That makes the release more relevant as a cost-and-capability trade-off than as evidence that it leads the field overall.

An early Kingy AI pilot, described in the source material, tested seven small tasks twice. It reports 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 sample is small and cannot establish broad reliability or rank the three models. The pilot also found 1017 was not always the fastest.

What Changed From September’s Model

DeepSeek released the preceding V4.1 Flash model on September 10, according to the source account. The October release is described as a minor update, and DeepSeek has replaced the older hosted version rather than requiring users to select a new API name. The model’s weights are also offered as an FP8 checkpoint with FP4 experts, according to the source material.

The technical details and training explanation require attribution. A technical report, as summarised by Kingy AI, describes a sparse Mixture of Experts design with a Causal Encoder-Decoder structure, 40 layers and 384 routed experts, six active per token, plus a shared expert. The report gives a 552-billion-parameter backbone with different active-parameter counts during prefill and decoding; Hugging Face lists 763 billion under a different counting method. DeepSeek has not publicly specified what changed in training for version 1017. The Decoder attributes the gains to post-training with reinforcement learning and self-feedback, but that explanation is not confirmed here by a detailed DeepSeek account.

DeepSeek says it tuned 1017 for DeepSeek Harness and for Tencent’s CodeBuddy and WorkBuddy. Benchmark results can depend on the harness and evaluation setup: the source says Terminal-Bench used Claude Code, while DeepSWE ran in Mini-SWE-V2. The selected comparison table also omits newer rival versions named in the source, limiting what it can establish about relative standing.

Independent Checks Remain Limited

The benchmark figures in the model card are DeepSeek’s own results; the source material says they have not been independently reproduced. It does not provide enough detail to verify every benchmark run, including repeat counts, variability or whether all competitor scores were produced under directly comparable conditions. DeepSeek’s chart uses selected rival models, and the source notes that some newer versions are absent.

The cost comparisons also need qualification. DeepSeek’s DeepSWE chart is reported to show 1017 reaching 77.0% at just under $0.20 per task and a lower-effort setting exceeding 72% at about $0.10. Those are results attributed to the company’s chart, not a general price guarantee. Per-task spending changes with the workload and evaluation method. Kingy AI’s pilot offers limited outside evidence, but 14 runs across seven small tasks are not enough to establish how the model performs on larger or more varied work.

The supplied source ends while describing a separate DeepSeek search announcement, after saying the company opened its search engine to developers through two agent tools. It does not include the tools’ names, features or operating details, so those specifics cannot be reported here. The material also does not confirm whether the reported model architecture or training changes were independently audited.

Developers Can Test the Hosted Update

Developers using DeepSeek’s hosted deepseek-flash endpoint now receive version 1017, according to the release account. They can compare it with their prior results using their own workloads, noting token use, completion quality, latency and peak versus off-peak rates. DeepSeek recommends DeepSeek Harness version 0.2.1-rc.1 or later for its own agent setup, according to the source material.

The next useful evidence would be independent, repeatable evaluations that publish their prompts, harnesses, task mix and cost calculations, alongside comparisons with newer competing models. Further technical details from DeepSeek about training changes and the announced search tools would also help clarify what changed beyond benchmark scores. Until those details are available, the update’s gains and cost advantages should be treated as reported results rather than settled independent findings.

Key Questions

What is DeepSeek-V4.1-Flash-1017?

It is DeepSeek’s October 17, 2026 update to its V4.1 Flash model. DeepSeek says it now serves the updated version through its website, app and API in place of the September release.

Did DeepSeek change the API price?

The release account says the model update came at an unchanged price. DeepSeek lists different rates for cached and uncached input, output and peak or off-peak periods, so total costs depend on usage.

Does 1017 outperform every model in DeepSeek’s chart?

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

Have the benchmark gains been independently confirmed?

The supplied source says DeepSeek’s benchmark results have not been independently reproduced. It describes a small Kingy AI pilot, but that test is too limited to verify broad performance claims.

What should developers check before switching?

Test the model on representative tasks and track quality, token use, latency and task completion costs. Benchmark results can vary with the harness and workload, and peak pricing differs from off-peak rates.

Source: ThorstenMeyerAI.com