DeepSeek V4.1 Flash 1017: Near-Frontier Agents At A Tenth Of The Price. Read The Table Before You Switch.

🔍 Read the full analysis: DeepSeek V4.1 Flash 1017: Near-Frontier Agents At A Tenth Of The Price. Read The Table Before You Switch. 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 DeepSeek-V4.1-Flash-1017 on October 17, replacing the September version on its website, app and API while keeping the listed token prices unchanged. The company says the update improves scores across all 11 benchmarks it published, with its largest gains on some multi-step agent tasks; the results are DeepSeek’s own and have not been independently reproduced.

The new version is served through the API name deepseek-flash; the older names deepseek-v4-flash and deepseek-v4-flash-vision-exp are also routed to it, according to the source report. DeepSeek describes the release as a minor update. Its model weights are available under the MIT license on Hugging Face and ModelScope.

In DeepSeek’s published table, Terminal-Bench 4.0 rose from 31.2 to 42.3, while ProgramBench increased from 20.3 to 29.1. Apex Bench rose from 64.6 to 72.1. The improvement was smaller on DeepSWE v1.1, up from 74.2 to 77.0. These are benchmark score changes, not independently verified results; the source does not specify that the gains apply to every agent task in ordinary use.

The model is described as a sparse Mixture of Experts system with image input and a one-million-token context window. A technical-report summary cited by the source says it uses 40 layers and 384 routed experts, with six active per token and one shared expert. Reported model size varies by counting method: the source gives a 552-billion-parameter backbone and notes that Hugging Face lists 763 billion. DeepSeek has not detailed what changed in training. The Decoder attributed the improvement to post-training with reinforcement learning and self-feedback, but that explanation is not confirmed by DeepSeek in the material provided.

Lower Costs for Agent Runs

The commercial case for 1017 is less about leading every benchmark than about the reported cost per completed task. The source says DeepSeek’s DeepSWE chart puts the model at 77% for just under $0.20 per task, while a lower-effort setting reaches more than 72% at about $0.10. These are figures reported from DeepSeek’s evaluation, not a general guarantee of cost or performance across applications.

DeepSeek’s listed API rates are $0.15 per million uncached input tokens off-peak and $0.60 per million output tokens, rising to $0.30 and $1.20 during peak hours. Cached input is listed at $0.003 off-peak and $0.006 at peak. The source calculates that a request using one million uncached input tokens and 200,000 output tokens would cost $0.27 off-peak at those rates. Actual bills will depend on token use, caching, scheduling and the work required to complete a task.

For teams running many coding or research agents, a lower price can make experimentation and repeated runs more affordable. But raw benchmark scores and price-per-task charts do not settle whether the model will be reliable, fast or economical in a particular workflow. A small independent pilot cited by the source found promising results, while also showing that 1017 was not always the fastest option.

What DeepSeek Compared

DeepSeek’s model card compares 1017 with its September release and four selected rivals: Kimi K3, GLM-5.3, GPT-5.6 Sol and Claude Opus 5. In that table, 1017 leads four of the 11 listed rows. Claude Opus 5 leads six, including HLE, ProgramBench, NL2Repo-Bench and two vision evaluations; GPT-5.6 Sol leads the remaining row. The table therefore shows gains over the previous Flash version, not an overall lead over every model included.

The source notes that DeepSeek did not include newer systems such as GPT-6.1 Sol, GPT-6 Astra, Opus 5.5 or Gemini 4 Argon. It also cites separate Terminal-Bench results for current frontier models in a range of 57 to 66, but says those evaluations are not directly comparable. Benchmark results can vary with the model version, harness, task set and scoring method. DeepSeek’s Terminal-Bench evaluation used the Claude Code harness, while DeepSWE used Mini-SWE-V2.

The source also describes an early Kingy AI pilot across seven small tasks, with each task run twice. It reports that 1017, GPT-6.1 Sol and Claude Sonnet 5.5 each passed all 14 runs, at roughly a quarter of the cost for the latter models. Because all three achieved a perfect score on this limited sample, the pilot could not rank them. One task took 1017 about 27 seconds, compared with about three seconds for Sonnet, according to that account.

Independent Results Still Pending

No independent reproduction of DeepSeek’s 11 benchmark results is reported in the source material. The company’s table also uses a selected set of competitors and does not include several newer models, so it cannot establish how 1017 ranks against the full current field. Separate results cited for newer systems are not direct comparisons because evaluation conditions differ.

Several implementation details are also unsettled in the supplied account. DeepSeek has not said what changed in training, while the explanation involving self-feedback is attributed to The Decoder. The report gives different total parameter counts depending on the counting method, and the supplied source cuts off while describing a separate announcement about search tools. Their full features, availability and terms cannot be established from the material provided.

The small Kingy AI pilot is an early, limited check rather than a broad evaluation: it involved seven tasks run twice, and three models tied on the reported pass count. More testing would be needed to compare reliability, speed and costs across larger workloads and different agent setups.

More Testing and Tool Details

Developers can access 1017 through DeepSeek’s website, app and API, and download the weights under the MIT license, according to the source. The company recommends its DeepSeek Harness version 0.2.1-rc.1 or later and says the model was tuned for that harness as well as Tencent’s CodeBuddy and WorkBuddy. Teams evaluating it will need to test their own workloads, including whether the model’s output length, speed and success rate produce the advertised savings.

The next useful evidence would be independent benchmark runs under clearly described conditions, including comparisons with current models using the same harness and task set. Further detail from DeepSeek about training changes and the reported cost-per-task methodology would help clarify what accounts for the gains. The source also references a DeepSeek search announcement, but its account is incomplete; additional details about the tools and their release are not established here.

Key Questions

What changed in DeepSeek V4.1 Flash 1017?

DeepSeek says the October 17 version improves scores across 11 published benchmarks, with larger gains on Terminal-Bench 4.0 and ProgramBench than on DeepSWE. The company has not detailed the training changes behind the update in the supplied material.

Does the update cost more to use?

No price increase is reported: the source says the model replaced the September version at unchanged listed token prices. API costs vary with input and output tokens, caching and whether usage falls in peak or off-peak hours.

Has DeepSeek’s benchmark performance been independently confirmed?

Not according to the source material. The 11 benchmark results come from DeepSeek, and the reported Kingy AI pilot is small: seven tasks were run twice, with three models tying on the pass count.

Is 1017 the top-performing model in DeepSeek’s comparison?

No. In DeepSeek’s selected comparison table, 1017 leads four of 11 rows. Claude Opus 5 leads six and GPT-5.6 Sol leads one. The table does not include several newer models named in the source.

Can developers use the model’s weights?

The source says the weights are available on Hugging Face and ModelScope under the MIT license. DeepSeek also serves the model through its website, app and API.

Source: ThorstenMeyerAI.com