DeepSeek V4.1 Flash 1017 And The Economics Of Near-Frontier Agents

🔍 Read the full analysis: DeepSeek V4.1 Flash 1017 And The Economics Of Near-Frontier Agents 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 across its website, app and API while keeping prices unchanged. The company says the update improves all 11 benchmarks in its model card, with the largest reported gains on long, multi-step agent tasks; independent evidence remains limited.

The model is served as deepseek-flash, and DeepSeek says older aliases, including deepseek-v4-flash, now route to 1017. Its weights are available under the MIT license. DeepSeek describes the release as a minor update, but has not detailed what changed in training. The Decoder reported that the gains may come from post-training with reinforcement learning and self-feedback; that account is not a detailed training disclosure from DeepSeek.

In DeepSeek’s benchmark table, Terminal-Bench 4.0 rises from 31.2 for the September model to 42.3 for 1017, while ProgramBench increases from 20.3 to 29.1. DeepSWE, a coding benchmark highlighted by the company, moves from 74.2 to 77.0. Those figures are points on the benchmarks as reported by DeepSeek, not independently reproduced results.

The company also says 1017 produces shorter outputs, which can reduce completion time and token charges. The model is tuned for DeepSeek Harness, with version 0.2.1-rc.1 or later recommended, and for Tencent’s CodeBuddy and WorkBuddy. Its listed API rates are $0.15 per million uncached input tokens off-peak and $0.30 at peak; output costs $0.60 off-peak and $1.20 at peak. Cached input is cheaper.

Lower Costs for Long Agent Runs

The release makes the economics of agent work a central part of DeepSeek’s pitch. Agents that use tools may issue repeated requests and produce lengthy outputs before a task is complete. If 1017 can complete similar work with fewer tokens, its lower listed rates could make repeated or large-scale runs less expensive for developers.

DeepSeek’s reported cost chart, described by The Decoder, puts 1017 at just under $0.20 per DeepSWE task at a 77.0% score. A low-effort setting is reported to clear 72% at about $0.10 per task; the September model is shown at about $0.50 for a 74.2% score. These are figures from a company benchmark comparison reported by a secondary source, not a broad measure of costs across real deployments.

Kingy AI reported an early pilot covering seven small tasks, each run twice. In that test, 1017, GPT-6.1 Sol and Claude Sonnet 5.5 each passed all 14 runs, with 1017 costing roughly a quarter as much as the other two. The pilot also found differences in speed: one 1017 task took about 27 seconds, compared with about three seconds for Sonnet. With every model scoring perfectly on this small set, the results do not establish which performs best overall.

Products Worth Considering

Apple AirPods Pro 3 Wireless Earbuds with Active Noise Cancellation

AmazonView Latest Price

EcoFlow Delta 3 Max Portable Power Station with 220W Solar Panel, 2048Wh

AmazonView Latest Price

LG UltraGear 32GX850A-B 32-inch 4K UHD OLED Gaming Monitor

AmazonView Latest Price

As an affiliate, we earn on qualifying purchases.

A Stronger Model, With Limits

DeepSeek released the earlier V4.1 Flash model on September 10. The October 17 update replaces that version under the same API alias, so users of deepseek-flash receive 1017 automatically. The company calls the release a minor update and has not announced a price increase.

The technical description identifies a sparse Mixture of Experts architecture with native image input and a one-million-token context window. It reports a 552-billion-parameter backbone, with about eight billion active parameters during prefill and 16 billion during decoding. Hugging Face lists 763 billion, reflecting a different counting method. DeepSeek’s technical report is titled “Pushing the Limits of KV Cache Compression”; the supplied summary says the cache is about 890 bytes per token.

DeepSeek’s chart shows gains over its September model, but not a lead across the board. Version 1017 leads four of the 11 rows, while Claude Opus 5 leads six and GPT-5.6 Sol leads one. DeepSeek compared against the models named in its chart, including GPT-5.6 Sol and Opus 5; it did not include newer models such as GPT-6.1 Sol, GPT-6 Astra, Opus 5.5 or Gemini 4 Argon. Separately reported Terminal-Bench results for current frontier models range from 57 to 66, according to Kingy AI, but those evaluations differ and are not directly comparable.

““minor update””

— DeepSeek, describing the release

Independent Results Still Sparse

The 11 benchmark scores in DeepSeek’s model card come from the company, and the supplied material does not include independent reproduction of the full chart. Evaluation settings also vary: Terminal-Bench was run in the Claude Code harness, while DeepSWE used Mini-SWE-V2. The full-set HLE results are 42.6 for 1017 and 36.8 for the September model; on the text-only subset they are 44.0 and 39.1. GLM-5.3’s 42.0 figure is text-only, so the full-set and text-only numbers should not be treated as a direct comparison.

DeepSeek has not explained the precise training changes behind the update. Its chart also compares against a selected group of rivals rather than the newest available systems. The separate frontier Terminal-Bench figures cited by Kingy AI use different evaluations, leaving the size of any like-for-like gap unresolved.

The source material also begins describing a second announcement: developer access to DeepSeek search through two agent tools. It says Search returns real-time results, but the description ends before specifying the second tool, availability, pricing or technical limits. Those details cannot be established from the information provided.

Developer Use Will Test the Claims

Developers can use the updated model through DeepSeek’s website, app or API, where the deepseek-flash alias now points to 1017. DeepSeek recommends its own Harness version 0.2.1-rc.1 or later, and says it has tuned the model for Tencent’s CodeBuddy and WorkBuddy. The practical test will be whether the reported gains persist across different agent setups and tasks.

Further independent evaluations could clarify whether shorter outputs translate into lower total costs without reducing task quality, and how 1017 compares with newer frontier systems under matching conditions. The available information does not specify a date for such testing or for additional technical disclosures. Details of the announced search tools also remain incomplete in the source material.

Key Questions

What changed in DeepSeek V4.1 Flash 1017?

DeepSeek says the October 17 update improves results on all 11 benchmarks in its model card. It reports especially large gains over the September version on Terminal-Bench 4.0 and ProgramBench, and says outputs are shorter.

Does the update cost more to use?

DeepSeek’s listed prices are unchanged, and the existing deepseek-flash API alias now routes to 1017. Charges still depend on token use, caching and whether the request falls in peak or off-peak hours.

Are the benchmark gains independently confirmed?

The supplied material says the model-card results come from DeepSeek and have not been independently reproduced in full. Kingy AI reported a small pilot with seven tasks run twice, which is not enough to establish performance across broader workloads.

Is 1017 the top model in DeepSeek’s comparison?

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

What is known about DeepSeek’s search announcement?

The source says DeepSeek opened its search engine to developers as two agent tools and that Search returns real-time results. The provided description cuts off before explaining the second tool or giving availability and pricing details.

Source: ThorstenMeyerAI.com