DeepSeek V4.1 Flash 1017: Is A Tenth Of The Price Worth The Tradeoff?

🔍 Read the full analysis: DeepSeek V4.1 Flash 1017: Is A Tenth Of The Price Worth The Tradeoff? on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get audio and creator gear 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 its September Flash model on the company’s website, app and API. The company says the update improves results across all 11 benchmarks it published while producing shorter outputs, with no change to listed API prices; however, those results have not been independently reproduced, and the comparison set omits newer competing models.

The model is served as deepseek-flash, so API users calling that name are routed to 1017 automatically. DeepSeek also routes the older names deepseek-v4-flash and deepseek-v4-flash-vision-exp to the updated model. Its weights are available under the MIT license, according to the source material.

In DeepSeek’s model-card results, Terminal-Bench 4.0 rose from 31.2 to 42.3, while ProgramBench increased from 20.3 to 29.1. DeepSWE, a coding benchmark, moved from 74.2 to 77.0. These are benchmark-point scores reported by DeepSeek, not independently verified findings. The gains vary by test: the DeepSWE increase was smaller than the gains on Terminal-Bench and ProgramBench.

DeepSeek’s chart shows 1017 leading four of 11 benchmark rows. Claude Opus 5 leads six, including HLE and ProgramBench, while GPT-5.6 Sol leads one. The comparison is limited to the models DeepSeek selected, including GPT-5.6 Sol and Opus 5; the chart does not include newer models named in the source material. Some results also use different benchmark subsets or testing harnesses, so not every row is directly comparable.

The Cost Case for Agent Tasks

The update’s clearest potential value is cost per completed task, rather than a claim to lead the field overall. DeepSeek’s published prices are $0.15 per million uncached input tokens and $0.60 per million output tokens off-peak, rising to $0.30 and $1.20 during peak hours. Cache-hit input is listed at $0.003 off-peak and $0.006 at peak. DeepSeek says peak pricing applies from 01:00–04:00 and 06:00–10:00 UTC on weekdays, excluding Chinese public holidays.

Using those rates, the source’s example of one million uncached input tokens and 200,000 output tokens costs $0.27 off-peak. That is a price illustration, not a general estimate for every job: actual cost depends on token use, cache hits and time of use. The reported lower cost could matter to developers running repeated coding or agent workflows, where inference charges accumulate. Shorter outputs may also reduce usage, though DeepSeek has not provided an independently checked measurement of the savings across typical workloads.

DeepSeek’s DeepSWE chart, as reported by The Decoder, puts 1017 at 77.0% for just under $0.20 per task; a lower-effort setting reached more than 72% at about $0.10. The same report says the September version cost about $0.50 per task for a 74.2% result, and rival models on that chart ranged from $2 to more than $10. These figures support a cost-efficiency argument, but they remain tied to a particular benchmark and setup. A low price is useful only if performance, reliability and operating requirements suit the task.

Kingy AI’s early pilot offers limited independent evidence: it tested seven small tasks, each run twice, and reported all three tested models passing all 14 runs. The pilot included 1017, GPT-6.1 Sol and Claude Sonnet 5.5, with 1017 costing roughly a quarter as much as the other two in that test. Because every model passed, the sample could not rank their capability. Kingy AI also found 1017 was slower on one task, taking about 27 seconds compared with roughly three seconds for Sonnet.

How DeepSeek Defines This Update

DeepSeek describes 1017 as a minor update to V4.1 Flash, which it released on September 10. The new version replaces that release through DeepSeek’s own services; existing API calls using the general Flash name do not need a new model identifier. The company has not said what changed in training. The Decoder attributes the gains to post-training with reinforcement learning and self-feedback, but that account is not a detailed disclosure from DeepSeek in the supplied material.

The technical report, as summarised by Kingy AI, describes a sparse Mixture of Experts model with native image input and a one-million-token context window. It identifies a Causal Encoder-Decoder design, with 40 layers split between encoder and decoder, 384 routed experts and six active per token, plus a shared expert. Reported parameter counts differ by counting method: the source describes a 552-billion-parameter backbone and lists about 8 billion active during prefill and 16 billion during decoding, while Hugging Face lists 763 billion.

The report is titled Pushing the Limits of KV Cache Compression and puts the KV cache at about 890 bytes per token. The hosted API is reported to allow up to 384,000 output tokens. DeepSeek says the model has been tuned for its own DeepSeek Harness and for Tencent’s CodeBuddy and WorkBuddy. Those details matter because benchmark results and real-world costs can depend on the software harness and task setup, not just the model weights.

“1017 passed 14 of 14”

— Kingy AI, in its early independent pilot

Evidence Behind the Score Gains

Independent confirmation remains limited. DeepSeek’s published table supplies the main benchmark evidence, and no broad independent reproduction is provided in the source material. Kingy AI’s pilot is small and cannot distinguish between the three models that passed every task. The results therefore do not establish how 1017 performs across a wider range of real-world work.

The benchmark comparisons also have limits. DeepSeek chose the rivals in its chart, and its list does not include several newer models identified in the source material. Kingy AI reports separate Terminal-Bench scores of 57 to 66 for current frontier models, but says those evaluations are not directly matched to DeepSeek’s test. It is not clear from the available information whether the reported performance gap would hold under identical prompts, software, compute and scoring conditions.

Other open points include what changed during training, how the shorter outputs affect accuracy on tasks beyond the reported tests, and how typical users’ bills will compare with DeepSeek’s illustrative costs. The supplied material also mentions a second announcement about developer access to search tools, but does not provide enough detail to establish its features, availability or pricing.

What Users Can Test Next

Developers can access 1017 through DeepSeek’s website, app and API, where the general deepseek-flash name now routes to the new version. The MIT-licensed weights are also listed on Hugging Face and ModelScope, according to the source material. That gives users options to test the hosted service or work with the released checkpoint, though the material does not specify hardware requirements or deployment costs.

The practical next step is to compare 1017 with alternatives on the same tasks, prompts, harnesses and output limits, while recording both pass rates and total cost. Users considering API workloads should also account for DeepSeek’s peak-price window and cache status rather than relying on a single per-task estimate. Further independent evaluations would help establish whether the gains in DeepSeek’s chart carry over to different coding, reasoning and agent workflows.

DeepSeek’s reported search-tool announcement may add another route for developers building agents, but details are incomplete in the available material. Its tools, access terms and relationship to the model update remain to be clarified.

Key Questions

What changed in DeepSeek V4.1 Flash 1017?

DeepSeek says the October 17 update improves its results on all 11 benchmarks in its model card and produces shorter outputs. It replaces the September version across the company’s website, app and API. The company has not detailed what changed in training in the source material.

Did DeepSeek raise the API price?

No price increase is reported. DeepSeek’s listed token rates are unchanged for users routed to 1017, though costs vary by input and output tokens, cache hits and whether requests fall within the stated peak hours.

Does 1017 outperform the other models in DeepSeek’s chart?

Not across the chart as a whole. DeepSeek’s results show 1017 leading four of 11 rows; Claude Opus 5 leads six and GPT-5.6 Sol leads one. The chart does not include several newer rivals named in the source material, and the results have not been broadly independently reproduced.

Is there independent evidence that 1017 is cheaper per task?

Kingy AI reported a small pilot in which 1017 passed 14 runs across seven small tasks and cost roughly a quarter as much as GPT-6.1 Sol and Claude Sonnet 5.5 in that test. All three models passed every run, so the pilot could not rank them, and its limited task set does not establish typical costs across workloads.

Can developers use the model weights themselves?

The source material says the weights are available under the MIT license on Hugging Face and ModelScope. It identifies an FP8 checkpoint with FP4 experts but does not provide hardware requirements or an estimate of the cost of self-hosting.

Source: ThorstenMeyerAI.com