DeepSeek V4.1 Flash 1017 Brings AI Agent Costs Down—What To Compare

🔍 Read the full analysis: DeepSeek V4.1 Flash 1017 Brings AI Agent Costs Down—What To Compare on ThorstenMeyerAI.com

Age 18–24?Offer from Amazon

Prime made for students and young adults

• Fast, free delivery for dorm and study essentials
• Prime Video and Amazon Music included
• Member-only deals

Try Prime for Young Adults →Free trial for eligible 18–24 year olds

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 without changing listed token prices. The company says the update improves scores on all 11 benchmarks it published, with the largest gains on multi-step agent tasks; those results have not been independently reproduced, and DeepSeek’s comparison chart uses rivals that are not the newest available models.

The updated model is served as deepseek-flash; the older names deepseek-v4-flash and deepseek-v4-flash-vision-exp also route to it, according to the source report. DeepSeek has released the weights under the MIT license on Hugging Face and ModelScope. Its model card lists a one-million-token context window, with hosted API output of up to 384,000 tokens.

In DeepSeek’s published table, Terminal-Bench 4.0 rose from 31.2 to 42.3 and ProgramBench from 20.3 to 29.1, compared with the September model. DeepSWE v1.1, a coding benchmark, increased from 74.2 to 77.0. The gains were not equal across tests: DeepSWE rose 2.8 points, while the larger changes appeared in Terminal-Bench and ProgramBench. These are company-reported benchmark scores, not independently verified results.

DeepSeek’s listed prices are $0.15 per million input tokens for uncached requests and $0.60 per million output tokens off-peak; peak prices are $0.30 and $1.20, respectively. Cached input is listed at $0.003 off-peak and $0.006 at peak. The report says an example job using one million uncached input tokens and 200,000 output tokens would cost $0.27 off-peak at DeepSeek’s rates. That estimate depends on the stated token counts and pricing window, and does not represent a universal cost per task.

Why Lower Task Costs Matter

For teams running coding agents or other workflows that make many model calls, the relevant comparison is often cost per completed task, not the price of one token or a single benchmark rank. DeepSeek’s figures, if borne out in independent testing, suggest the update may make longer agent runs more affordable while improving performance over its prior Flash version.

The Decoder reported DeepSeek’s DeepSWE chart puts 1017 at 77% success for just under $0.20 per task, while a lower-effort setting reached more than 72% at about $0.10. The September model was shown at about $0.50 per task for a 74.2% score. These are figures from a company benchmark chart as reported by the publication, not a general guarantee of costs or results on customer workloads.

One small pilot offers a limited external check. Kingy AI reported that 1017, GPT-6.1 Sol and Claude Sonnet 5.5 each passed 14 of 14 attempts across seven small tasks run twice, with 1017 costing roughly a quarter as much. The sample is too small to establish broad superiority, and the perfect scores do not distinguish the models. The pilot also found that 1017 was not always fastest.

How the Update Was Tested

DeepSeek described 1017 as a minor update to V4.1 Flash, which it released on September 10. The source report says DeepSeek has not specified what changed in training. The Decoder attributed the improvement to post-training with reinforcement learning and self-feedback, but that explanation is reporting by the publication, not a detailed account confirmed in the supplied model information.

DeepSeek says it tuned the model for its own DeepSeek Harness and for Tencent’s CodeBuddy and WorkBuddy. Its architecture is described in a technical report as a sparse Mixture of Experts model with image input. The report, as summarized by Kingy AI, describes a Causal Encoder-Decoder design; model-size figures vary by accounting method, with the source citing a 552-billion-parameter backbone and Hugging Face listing 763 billion.

The benchmark table also needs qualification. DeepSeek compared 1017 with its September model and selected systems including Kimi K3, GLM-5.3, GPT-5.6 Sol and Claude Opus 5. In that table, 1017 led four of 11 rows, while Opus 5 led six and GPT-5.6 Sol led one. The source report says newer frontier models were not included. Separately reported Terminal-Bench results for newer models were higher, but they came from different evaluations and are not a direct comparison.

““minor update””

— DeepSeek, as characterized in its model announcement

Limits of the Published Evidence

Independent replication remains absent for DeepSeek’s 11 benchmark results. It is not yet clear how the update performs across a wider range of production workloads, different agent harnesses or repeated tests by independent evaluators. Benchmark scores can depend on task selection, tool access, prompting and evaluation setup.

The source report does not provide a full account of what changed in training, and the self-feedback explanation is attributed to The Decoder. DeepSeek’s chart also does not compare the model with several newer systems named in the report. Results from separate evaluations cannot fill that gap without matched conditions.

Cost figures also require care. Token prices are not the same as total operating costs: actual spending depends on input and output volume, cache use, peak or off-peak timing, retries and whether a task succeeds. The reported pilot covered seven small tasks, each run twice, so it cannot establish performance across broad or high-stakes use.

What Buyers Should Compare

Developers evaluating 1017 should run matched tests on their own tasks, using the same prompts, tools, model settings and success criteria for each candidate. They should record completion rate, cost per successful task, latency and retry frequency, rather than relying on token prices or one benchmark score alone.

For a fair comparison, teams will also need current model versions and equivalent harnesses. DeepSeek recommends DeepSeek Harness version 0.2.1-rc.1 or later, according to the source report; results from other setups may differ. The next useful evidence will be independent evaluations that publish their task sets, methods and cost calculations, alongside any further details DeepSeek provides on training and model changes.

Until those comparisons are available, 1017’s lower listed prices and gains over the September Flash version are confirmed aspects of the announcement, while broad claims about its standing against current frontier models remain unsettled.

Key Questions

When did DeepSeek release V4.1 Flash-1017?

DeepSeek released it on October 17, 2026. The source report says it replaced the September version on DeepSeek’s website, app and API.

Did DeepSeek change its API prices?

No price change is reported. DeepSeek’s listed rates remain different for peak and off-peak periods, and for cached versus uncached input. Actual task costs depend on token use and other workload details.

Are the benchmark improvements independently verified?

Not in the supplied reporting. The 11 benchmark results come from DeepSeek, and the source says they have not been independently reproduced. Kingy AI reported a small pilot, but its seven-task sample is not broad validation.

Does 1017 lead DeepSeek’s comparison chart?

It leads four of the 11 rows in DeepSeek’s chart. Claude Opus 5 leads six and GPT-5.6 Sol leads one. The chart does not include several newer rivals identified in the source report.

What should a company compare before switching?

Test models on the same real tasks and tools, then compare successful completions, total cost, latency and retries. Token prices and benchmark results alone do not show how a model will perform on a particular workload.

Source: ThorstenMeyerAI.com