What DeepSeek V4.1 Flash 1017 Could Mean For AI Agent Costs

🔍 Read the full analysis: What DeepSeek V4.1 Flash 1017 Could Mean For AI Agent Costs on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get hardware and tech essentials 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 prices unchanged. The company says the update improves scores on all 11 benchmarks it published, with its largest gains on multi-step agent tasks—results that could make the model cheaper to use for some coding and automation work, though most of the evidence comes from DeepSeek itself.

The model is served as deepseek-flash; calls using older names, including deepseek-v4-flash, are also routed to 1017. DeepSeek describes the release as a minor update. Its model card reports Terminal-Bench 4.0 rising from 31.2 to 42.3 and ProgramBench from 20.3 to 29.1, comparing 1017 with the September version. DeepSWE v1.1 rose by a smaller margin, from 74.2 to 77.0.

DeepSeek’s listed API rates vary by time and cache status. Per million tokens, uncached input costs $0.15 off-peak and $0.30 at peak; cached input costs $0.003 and $0.006, respectively. Output costs $0.60 off-peak and $1.20 at peak. DeepSeek defines weekday peak periods as 01:00–04:00 and 06:00–10:00 UTC, excluding Chinese public holidays. On those rates, the supplied example of one million uncached input tokens and 200,000 output tokens costs $0.27 off-peak.

The update’s weights are available under the MIT license, and the model accepts image input. DeepSeek recommends its DeepSeek Harness version 0.2.1-rc.1 or later and says the model was tuned for that harness and for Tencent’s CodeBuddy and WorkBuddy. The company has not disclosed what changed in training. The Decoder reports that the gains came from post-training with reinforcement learning and self-feedback, but that account is not a detailed disclosure from DeepSeek in the supplied material.

Lower Costs for Long Agent Tasks

The commercial case for 1017 is its cost per completed task, not a claim that it is the strongest model overall. The Decoder’s account of DeepSeek’s DeepSWE chart puts 1017 at 77.0% for just under $0.20 per task; a low-effort setting reportedly cleared 72% at about $0.10. The September model was shown at 74.2% for about $0.50. These are figures from a benchmark chart, and performance and cost in a customer’s own workflow can differ.

If the cost advantage carries over to real workloads, teams running frequent coding or other multi-step agent tasks could spend less per attempt. DeepSeek’s published price schedule also makes long inputs and outputs relatively inexpensive at list rates, especially when input is cached or jobs run off-peak. But token prices alone do not establish total operating cost: task success, retries, latency, tool use and engineering oversight all affect what a finished task costs.

An early Kingy AI pilot offers a small independent data point. Across seven small tasks run twice, its report says 1017, GPT-6.1 Sol and Claude Sonnet 5.5 each passed all 14 runs, at roughly a quarter of the cost of the two comparison models. The pilot also reported that 1017 marked unsupported facts as unknown and ignored instructions embedded in the input. Its sample was too small to rank the models, and one task took about 27 seconds with 1017 compared with about three seconds for Sonnet. The result supports further testing; it does not establish broad reliability or typical savings.

What DeepSeek’s Benchmarks Show

DeepSeek’s model card compares 1017 with its September release and four rivals selected by the company. On that table, 1017 leads four of 11 rows, while Claude Opus 5 leads six and GPT-5.6 Sol leads one. Opus 5 scores 56.3 on the full-set HLE test, compared with 42.6 for 1017, and leads on ProgramBench and NL2Repo-Bench. The table therefore shows improvement over DeepSeek’s prior model without showing overall leadership across the selected tests.

The benchmark comparisons need their test conditions attached. DeepSeek says HLE scores use the full set; its text-only scores are 44.0 for 1017 and 39.1 for September, while the listed GLM-5.3 score of 42.0 is text-only. Terminal-Bench ran in the Claude Code harness, and DeepSWE ran in Mini-SWE-V2. The figures are not interchangeable across different test setups.

The comparison set also omits newer models. Kingy AI cites separately reported Terminal-Bench scores of 57 to 66 for current frontier models, but those evaluations are not direct matches to DeepSeek’s test. DeepSeek describes the architecture as a sparse Mixture of Experts with a one-million-token context window. Its technical report, as summarized by Kingy AI, describes a 40-layer Causal Encoder-Decoder design. These details help characterize the release, but they do not independently validate its benchmark or cost claims.

““minor update””

— DeepSeek, describing the release

Independent Results Remain Limited

DeepSeek produced the benchmark table, and the supplied material does not describe independent replication of its 11 results. The Kingy AI pilot is independent but covers only seven small tasks, with each run twice; its perfect results across three models do not show how they compare on larger or more varied workloads.

It is also unclear how consistently the benchmark improvements translate into lower costs in production. DeepSeek says outputs are shorter and tasks finish faster, but the supplied information does not give a broad set of independently measured task times, retry rates or end-to-end costs. Training changes have not been disclosed by DeepSeek, and separately reported frontier scores use different evaluations, limiting direct comparisons.

Testing in Real Agent Workflows

The next useful evidence will come from independent evaluations on larger task sets and from teams measuring success rates, latency and total cost in their own workflows. Developers can access the model through DeepSeek’s API or use the released weights, while taking account of the company’s peak and off-peak rates. Whether 1017’s benchmark gains produce consistent savings outside the reported tests remains to be established.

Key Questions

What changed in DeepSeek V4.1 Flash 1017?

DeepSeek says the October 17 update improves scores on all 11 benchmarks in its model card, with larger gains on some multi-step agent tests. The company has not published details of its training changes.

Did DeepSeek raise API prices for the update?

No price increase is reported. DeepSeek replaced the September model with 1017 on its website, app and API at unchanged prices. Rates still vary by time, cache status and token type.

Is 1017 the top model in DeepSeek’s benchmark table?

No. It leads four of the 11 listed rows. Claude Opus 5 leads six, and GPT-5.6 Sol leads one. The results are DeepSeek’s own comparisons and use specified benchmark conditions.

Are the reported cost savings independently verified?

Not broadly. Kingy AI reported a pilot across seven small tasks, each run twice, in which 1017 passed all 14 runs at roughly a quarter of the cost of two comparison models. That limited sample does not establish typical savings across production workloads.

Source: ThorstenMeyerAI.com