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

The updated model is served as deepseek-flash, and DeepSeek also routes the older names deepseek-v4-flash and deepseek-v4-flash-vision-exp to it. The model’s weights are available under the MIT license. DeepSeek describes the release as a minor update, but does not specify what changed in training.

In DeepSeek’s model-card figures, 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 gain on DeepSWE, a coding benchmark, was smaller: 74.2 to 77.0. Those results point to uneven improvements rather than a uniform jump across tasks.

DeepSeek’s published prices remain $0.15 per million uncached input tokens off-peak and $0.30 at peak, with output priced at $0.60 and $1.20, respectively. Cached input costs less. The company says shorter outputs can also reduce per-task spending. Its example of one million uncached input tokens and 200,000 output tokens costs $0.27 off-peak, compared with $4 at list prices of $2 per million input tokens and $10 per million output tokens.

Lower Costs for Agent Tasks

The update’s main commercial argument is cost per completed task, not top position on a benchmark table. In a DeepSWE cost comparison reported by The Decoder, 1017 reached a 77% score for just under $0.20 per task; a lower-effort setting scored above 72% at about $0.10. DeepSeek’s September model cost about $0.50 per task for a 74.2% score. These are reported figures, not a direct measure of what every user will pay in production.

If the price and performance pattern holds for real workloads, developers running repeated coding and other agent tasks could spend less without moving to a more expensive model. But a benchmark score does not establish reliability across different software projects, tools or task lengths. Buyers will need to test the model against their own workloads, including the cost of retries, human review and failures.

The release also makes comparison difficult for customers: DeepSeek’s results suggest a low-cost option, while its table does not compare against several newer models named in the source material. That limits what readers can conclude about 1017’s place among current frontier systems.

What the Benchmarks Compare

DeepSeek’s table places 1017 beside its September model and selected models from Kimi, GLM, OpenAI and Anthropic. In that table, 1017 leads four of the 11 listed benchmark rows; Claude Opus 5 leads six, and GPT-5.6 Sol leads one. Opus 5 records higher scores on HLE, ProgramBench, NL2Repo-Bench and two of the three listed vision tests.

The comparison has important limits. DeepSeek selected the rivals and reports all the scores itself; the table does not include newer models cited in the source material, such as GPT-6.1 Sol, GPT-6 Astra, Opus 5.5 or Gemini 4 Argon. Separately reported Terminal-Bench results for newer frontier models fall between 57 and 66, according to Kingy AI, but those evaluations are not directly comparable to DeepSeek’s figures.

The model is described as a sparse Mixture of Experts system with native image input and a one-million-token context. A technical report summarized by Kingy AI describes a 40-layer Causal Encoder-Decoder design with 384 routed experts, six active per token, plus a shared expert. DeepSeek’s report also focuses on KV-cache compression. These architecture details describe the system, but do not independently validate benchmark or cost claims.

““minor update””

— DeepSeek, describing the release

Independent Checks Still Missing

No independent reproduction of DeepSeek’s benchmark table is cited in the source material. The company has not explained what training changes produced the update’s reported gains. DeepSeek also compared 1017 with a selected group of rivals that does not include several newer models, while results from separate evaluations cannot be treated as like-for-like comparisons.

Kingy AI’s early pilot offers a limited outside check: it reports that 1017, GPT-6.1 Sol and Claude Sonnet 5.5 each passed 14 runs across seven small tasks. The pilot is too small to establish which model is better, and it found that 1017 was not always fastest: one task took about 27 seconds, compared with about three seconds for Sonnet. The source material does not provide broader independent results, production failure rates or a full cost comparison across diverse workloads.

Actual bills will depend on input and output volume, cache use, timing and task completion. DeepSeek’s published prices specify peak and off-peak periods, but the benchmark cost estimates do not by themselves show how those rates translate to every user’s workflow.

Developer Testing and Follow-Up

Developers can access the replacement model through DeepSeek’s website, app and API, or download its MIT-licensed weights from Hugging Face and ModelScope. DeepSeek recommends its DeepSeek Harness version 0.2.1-rc.1 or later and says it tuned 1017 for that setup, as well as Tencent’s CodeBuddy and WorkBuddy.

The next useful evidence will be independent tests using larger and more varied task sets, with comparable harnesses and transparent reporting of failures, speed and total cost. It is also not clear from the source material whether DeepSeek will publish further technical details or add newer rival models to its comparisons. For now, users can test the current API version at the same listed token prices, while treating the headline benchmark gains as claims awaiting wider verification.

Key Questions

What changed with DeepSeek V4.1 Flash 1017?

DeepSeek released V4.1 Flash 1017 on October 17, 2026. It replaced the September version on the company’s website, app and API, and DeepSeek reports higher scores on all 11 benchmarks in its model card.

Did DeepSeek raise its prices?

No price increase is reported. DeepSeek says token prices are unchanged; actual costs depend on input and output volume, cache use and whether requests fall in peak or off-peak hours.

Are the benchmark gains independently verified?

Not in the source material. The benchmark scores come from DeepSeek’s own model card, and no independent reproduction of the full table is cited. A separate small pilot by Kingy AI is not large enough to confirm broad performance rankings.

Does 1017 outperform the models in DeepSeek’s table?

Not across the table as a whole. DeepSeek’s figures show 1017 leading four of 11 rows, while Claude Opus 5 leads six and GPT-5.6 Sol leads one. The comparison also omits several newer rivals named in the source material.

What should developers watch for when testing it?

Tests should track completed-task quality alongside total cost, speed, retries and failures. The reported pilot found a case where 1017 took about 27 seconds versus about three seconds for Claude Sonnet 5.5, showing that low price does not necessarily mean faster completion.

Source: ThorstenMeyerAI.com