On October 17, 2026, DeepSeek released DeepSeek-V4.1-Flash-1017, an AI model that handles both images and text. It is an improved version of V4.1 Flash, which debuted in September, and the API model deepseek-flash has also been switched to the new version, alongside the official website and app. DeepSeek says it improved performance on code fixing and tool-based tasks and reduced processing time and output tokens, while keeping pricing and the way the API is called unchanged. However, although it beats competing models on some benchmarks, gaps remain in other areas, and results also vary depending on the execution environment. DeepSeek stresses improvements in both performance and efficiency in its official changelog, but judging how much real work can be entrusted to a low-priced model requires considering not only benchmark scores but also the time and cost it takes to finish a task.
Ahead of rivals on code fixing, narrowing the gap with Opus on terminal operation
On DeepSWE v1.1, which measures code-fixing ability, DeepSeek-V4.1-Flash-1017 scored 77.0. That is up from 74.2 for the previous version, and in the comparison table DeepSeek published it exceeds Opus 5 at 74.0 and GPT-5.6 Sol at 73.0. It also leads K3 and GLM-5.3 by wide margins. On Terminal-Bench 4.0, an evaluation of agents that solve tasks by operating a terminal, however, Opus 5 still holds the top spot.
The main results listed in the official model card for the October 17 version are as follows.
| Benchmark | V4.1 Flash (previous) | 1017 version | K3 | GLM-5.3 | GPT-5.6 Sol | Opus 5 |
|---|---|---|---|---|---|---|
| Terminal-Bench 4.0 | 31.2 | 42.3 | 12.6 | 37.9 | 39.9 | 51.8 |
| DeepSWE v1.1 | 74.2 | 77.0 | 67.5 | 66.9 | 73.0 | 74.0 |
| ProgramBench | 20.3 | 29.1 | 17.5 | 19.0 | 23.0 | 37.0 |
| NL2Repo-Bench | 65.4 | 68.2 | 58.0 | 58.0 | 56.8 | 75.3 |
Terminal-Bench uses Pass@1, DeepSWE uses issue resolution rate, ProgramBench uses Almost@1, and NL2Repo-Bench uses Score as their metrics, so figures in different rows cannot be compared as if they were the same kind of accuracy rate. These are also results based on the model versions and evaluation conditions DeepSeek selected, not an independent ranking covering every model or the latest generation from each company.
On Terminal-Bench 4.0, the score rose 11.1 points, from 31.2 for the previous version to 42.3 for the new one. The gap to Opus 5's 51.8 narrowed from 20.6 points to 9.5. The new version also surpasses GPT-5.6 Sol's 39.9 and GLM-5.3's 37.9, showing gains in terminal operation, where it had previously trailed competitors.
However, the improvement cannot be assumed to come entirely from the model itself. The previous version's model card explains that Terminal-Bench evaluation mainly used the DeepSeek Harness, while the footnotes for the new version specify Claude Code. For DeepSWE, the new version used Mini-SWE-V2. In other words, what the published figures confirm is an improvement in evaluation scores, not how much performance would improve if only the model were swapped with the execution environment held constant.
The degree of improvement also differs across other coding benchmarks. On ProgramBench, the score rose from 20.3 to 29.1, surpassing GPT-5.6 Sol's 23.0 but falling short of Opus 5's 37.0. On NL2Repo-Bench, it scored 68.2, ahead of K3, GLM-5.3, and GPT-5.6 Sol, but still trails Opus 5's 75.3. Posting the top score on DeepSWE does not mean it outperforms competitors on every coding-related task.
Strong marks on practical agent work, but challenges remain in vision and reasoning
Benchmarks that evaluate practical work using tools produced even more striking results. On all three of Automation-Bench, Agents' Last Exam, and Apex Bench, the 1017 version outperforms the four competing models in the comparison table. Gains appear not just in simple code fixing but also in the ability to carry out a sequence of tasks using multiple tools.
| Benchmark | V4.1 Flash (previous) | 1017 version | K3 | GLM-5.3 | GPT-5.6 Sol | Opus 5 |
|---|---|---|---|---|---|---|
| Automation-Bench v1.0.6 | 54.8 | 60.5 | 46.7 | 48.8 | 45.8 | 50.3 |
| Agents' Last Exam | 31.8 | 35.4 | 27.6 | 28.5 | 26.7 | 28.6 |
| Apex Bench v1.1 | 64.6 | 72.1 | 47.8 | 56.6 | 54.3 | 65.8 |
The figures come from the official model card for the 1017 version. The Agents' Last Exam development team aims to evaluate long-running professional tasks in a form where outcomes can be verified. The 1017 version scored 35.4 on this benchmark, but that does not mean it can automate 35.4% of typical workplace jobs. Still, unlike tests that only measure knowledge, it offers evidence of the ability to actually carry out work.
On the other hand, gaps with competitors remain in advanced reasoning and visual recognition. On HLE (Humanity's Last Exam), a collection of difficult questions, in the no-tools evaluation, the score improved from 36.8 to 42.6 but fell well short of Opus 5's 56.3. On image-related benchmarks, it also improved over the previous version but does not consistently beat GPT-5.6 Sol or Opus 5.
| Benchmark | V4.1 Flash (previous) | 1017 version | K3 | GLM-5.3 | GPT-5.6 Sol | Opus 5 |
|---|---|---|---|---|---|---|
| HLE, no tools | 36.8 (39.1†) | 42.6 (44.0†) | 43.5 | 42.0† | 44.5 | 56.3 |
| Chartography, with tools | 78.9 | 83.0 | 68.1 | — | 79.9 | 84.0 |
| BabyVision, with tools | 89.6 | 92.5 | 85.7 | — | 88.9 | 94.1 |
| ZeroBench-main, with tools | 49.0 | 51.0 | 41.0 | — | 53.0 | 52.0 |
The figures are based on the 1017 version model card. † denotes the text-only subset of HLE, and "—" indicates items for which the comparison table lists no figure. The 42.6 for HLE and GLM-5.3's 42.0 cannot be compared directly because they cover different evaluation sets. In addition, according to the official changelog, ZeroBench-main uses Pass@5, which evaluates success across five candidates, and must be distinguished from Pass@1, which is measured with a single candidate.
In vision, Chartography improved from 78.9 to 83.0 and BabyVision from 89.6 to 92.5. However, these are results for whole systems combining tools and do not directly indicate the model's standalone image understanding. Whether it understood an image accurately or reached the right answer by using external tools needs to be considered separately. Moreover, the results include no margins of error or confidence intervals, so scoring close to Opus 5 alone does not justify concluding the two models are equivalent in ability.
Scores change with the execution environment, even for the same AI model
One thing to keep in mind when evaluating this performance improvement is the difference in the execution environment that runs the AI agent. In the comparison published in the previous version's model card, even when the same V4.1 Flash was evaluated on DeepSWE v1.1, results differed: 65.6 with Codex, 69.8 with Claude Code, and 74.2 with mini-SWE. The gap between the highest, mini-SWE, and Codex reached 8.6 points. This is not a result for the 1017 version but an example of the same September model evaluated in different execution environments.
An AI agent is not simply a system where you enter a question into a model and receive an answer. It combines tools such as file reading and command execution and carries out multiple steps while managing its work history. How tools are called and what information is returned to the model when errors occur also change its subsequent decisions and results. For this reason, the best score listed in a model card is not necessarily reproducible in your own development environment.
The compute allocated to reasoning also affects results. In the previous version's competitor comparison, the maximum reasoning setting was used, but the footnotes of the new version's comparison table alone do not make clear whether reasoning budgets and the number of processing iterations were uniform across all models. In addition, although the new model card lists recommended sampling settings, that does not guarantee those settings were used for every measurement in the comparison table. When considering adoption, it is necessary to compare not just model names but with the execution environment version, reasoning settings, and task pass criteria aligned.
DeepSeek itself says that, in this update, it worked on optimization for Tencent's CodeBuddy and WorkBuddy in addition to the DeepSeek Harness. It also recommends updating the DeepSeek Harness to 0.2.1-rc.1 or later. Rather than swapping in only the model and expecting to reproduce the published scores, it is important to verify how much work quality and processing time change with the combination of tools you actually use.
API pricing unchanged, but cost to complete a task may fall
API pricing for DeepSeek-V4.1 Flash was left unchanged in this update. Off-peak prices per million tokens are $0.15 for new input, $0.60 for output, and $0.003 for cache-hit input. For AI agents that repeatedly refer to past conversations and work history, costs need to be estimated separately for input that can reuse the cache and for newly processed input and output.
The prices confirmed in the official pricing table on October 17, 2026, are as follows.
| Per million tokens | Flash, off-peak | Flash, peak | V4 Pro, off-peak | V4 Pro, peak |
|---|---|---|---|---|
| Input, cache hit | $0.003 | $0.006 | $0.022 | $0.044 |
| Input, cache miss | $0.15 | $0.30 | $0.66 | $1.32 |
| Output | $0.60 | $1.20 | $1.98 | $3.96 |
Peak hours are 10:00–13:00 and 15:00–19:00 Japan time on weekdays, excluding Chinese public holidays; off-peak pricing applies at all other times. The current Pro model is "V4-Pro-0813," and the end of service announced in September has been withdrawn, so it remains available.
For example, if you use 1 million new input tokens, 1 million cached input tokens, and 100,000 output tokens off-peak, the cost is $0.213 for Flash and $0.88 for V4 Pro. These were calculated as 0.15 + 0.003 + 0.60 × 0.1 for Flash and 0.66 + 0.022 + 1.98 × 0.1 for Pro. The price difference for processing the same token volume is clear, but this is only an illustrative estimate, not the measured cost of completing the same job at equal quality. It also excludes the cost of external services and execution infrastructure the agent needs, such as search and code execution.
Separately from these unit price differences, the efficiency gains could lead to lower costs. DeepSeek says it reduced working time and output tokens while improving benchmark results, aiming to balance performance and cost across multiple execution environments. If output tokens decrease, the output charge shrinks even at the same unit price. And if task failures and rework decrease, the total cost of finishing a job can fall as well. However, the official model card and changelog give no specific reduction rates for processing time or token counts, so it is not possible to calculate by what percentage actual payments will fall as a result of this update.
In addition, the search and page-fetch APIs released the same day have a billing structure separate from the model itself. According to the Search & Fetch guide, both search and page fetching cost $0.75 per 1,000 successful requests, and failed requests, empty search results, and timeouts are not charged. The server-side search and fetch features used through the Harness and similar tools are currently free, but pay-as-you-go billing begins at 1:00 a.m. Japan time on November 1. Therefore, even with the model's API unit prices unchanged, total costs including the features an agent uses may not stay the same.
Technical requirements for running the released model yourself
DeepSeek-V4.1-Flash-1017 is a minor update of V4.1 Flash, supporting a context of up to 1 million tokens and image input. The September version's foundation uses a Mixture of Experts (MoE) architecture with a 552-billion-parameter main body. MoE is a mechanism that selects and runs only the parts needed for each task, with 8 billion parameters active during input processing and 16 billion during output processing. It also has a conditional memory called "Engram," which retrieves information from token sequences, amounting to 196 billion parameters. The September model card describes the main model and Engram separately.
It also adopts a "Causal Encoder-Decoder" that changes the processing structure for input and output, a design that reduces the load when repeatedly reading large conversation histories and execution logs. By sharing the KV cache, which holds past input information, across multiple layers, it cut the capacity needed to handle the full context to about one-quarter that of V4-Flash. However, this is an architectural feature announced for the September version, not a reduction rate newly achieved in the 1017 version. Nor has any explanation been published that would allow one to conclude that the reported decrease in output tokens resulted from a specific structure or training method.
The model weights and repository are released under the MIT license, but a small number of parameters active during inference does not mean the memory needed to store the entire model is also small. Likewise, the explanation that KV cache capacity is about one-quarter does not allow one to conclude that the server's total memory requirement drops by the same proportion. Running it in your own environment requires not only holding the main model and conditional memory but also preparing an inference environment that correctly supports image input and tool-calling formats.
The inference implementation DeepSeek released is also stated to be a reference implementation that prioritizes code readability, not a serving engine optimized for production. No Jinja-format chat template is included; instead, a Python reference implementation and "deepseek-recipe," for converting conversation formats, are provided. To build the released model into an actual service, you must put in place a mechanism for efficient inference and verify its connection and behavior with the tools you use.
With this update, DeepSeek-V4.1 Flash has strengthened its competitiveness in code fixing and practical tool-based work, while gaps with competing models remain in advanced reasoning and some vision tasks. In deciding whether to adopt it, whether through the API or self-hosting, it is important to evaluate existing tasks under the same pass criteria and confirm whether processing time and total cost can be cut while maintaining quality. As the number of jobs that meet those conditions grows, the range of work that can be entrusted to the low-priced Flash may expand from answering one-off questions to long-running tasks involving multiple steps.