DeepSeek V4 Flash API: Pricing, Playground and Examples

From code generation to multi-step development

We believe the most noteworthy coding capability of DeepSeek V4 Flash is not generating a single function, but advancing more complete software development tasks in a tool environment. Improvements in public terminal-operation and software-engineering evaluations for the 0731 official release make it more than just an option for "cheaper code completion". DeepSWE, for example, evaluates original, long-running software engineering tasks and verifies results through software behavior, rather than merely judging whether an answer looks reasonable. We therefore see greater value in development work with clear acceptance criteria, such as adding tests, changing features, and addressing cross-file issues.

Connecting information processing into executable workflows

When judging the tool-calling capabilities of DeepSeek V4 Flash, we look beyond whether it can generate correctly formatted call arguments and place more weight on whether it can maintain the task goal across multiple steps. Its excellent performance in this evaluation of tool-use tasks requiring extended execution in real software development environments is one of our main reasons for paying attention to it. We think the practical value of this capability is that it gives the model an opportunity to connect "reading materials, processing data, checking results, and preparing deliverables", rather than stopping at telling users what they should do.

Reducing information gaps caused by splitting materials

For long-document tasks, we place more weight on whether DeepSeek V4 Flash can connect scattered information. For example, the conditions for using an interface may be near the start of a document, exceptions near the end, and the actual implementation spread across several code files. A million-token context provides room to retain these materials together, an important capability compared with models that have smaller contexts.

Balancing everyday tasks and complex analysis

DeepSeek V4 Flash lists three reasoning levels, low, high, and max, which are settings of the same model. Tasks can be tested separately: field extraction mainly checks format and completeness, while code analysis checks diagnosis, reasoning, and modification suggestions. A level can then be selected based on quality, elapsed time, and token usage.

Our customer feedback includes users who praise its capabilities in complex projects, as well as users who report repeated investigation, detours, and substantial token consumption. Our understanding is that it has the ability to keep exploring a problem, but whether that exploration is effective also depends on task boundaries, tool configuration, and reasoning settings; thinking more does not necessarily mean finishing the work faster.

Making repetitive technical tasks economically feasible

Another aspect we value about DeepSeek V4 Flash is that it prompts developers to reconsider which tasks are worth handing over to a model. In our customer feedback, developers have reported frequent parallel usage and discussed using the model for repair suggestions in continuous integration, test generation, and log analysis. The key point about DeepSeek V4 Flash is that its lower barrier to use expands the range of tasks people are willing to try.

Terminal coding: Terminal-Bench 2.1 rises from 61.8 to 82.7

In the DeepSeek official-release comparison, V4 Flash scores 82.7 on Terminal-Bench 2.1, compared with 61.8 for the preview, an increase of 20.9 points. We think these are the most direct figures for this upgrade: the improvement is not limited to written answers, with a substantial upgrade in advancing tasks through a terminal environment as well. For users who need to handle builds, tests, or development-environment issues, this direction deserves more attention than code-generation scores alone. Our interpretation is that the official V4 Flash release has stronger execution capabilities in tool environments.

Complex software engineering: DeepSWE and NL2Repo both improve

DeepSeek V4 Flash scores 54.4 on DeepSWE in the official comparison, versus 7.3 for the preview; NL2Repo rises from 39.4 to 54.2. Both metrics improve, so we view this as an upgrade in software development capabilities rather than the result of optimizing just one type of terminal operation. We are particularly interested in the nature of DeepSWE tasks: they require models to complete extended engineering changes in real code repositories and verify them through behavior. Such results better support the conclusion that "V4 Flash can participate in complex development work".

Multi-tool execution: Toolathlon-Verified reaches 70.3

On Toolathlon-Verified, V4 Flash has an official score of 70.3, compared with 49.7 for the preview. This progress concerns not a single function call but multi-step tool use closer to real software environments. Our understanding of this result is that it is better able to use information obtained in an earlier step for subsequent actions, making it worth using for workflows that need tools, such as processing materials and analyzing data.

Independent abstract reasoning: ARC-AGI-2 reaches 61.4% at Max

Independently verified results published by ARC Prize show that V4 Flash at Max scores 89.0% on the ARC-AGI-1 semi-private evaluation and 61.4% on ARC-AGI-2. On ARC-AGI-2, Low and High score 46.0% and 56.0%, respectively, showing that reasoning settings noticeably affect the results. I value these results because they offer a perspective beyond official coding evaluations from the developer. ARC-AGI-2 focuses on reasoning problems such as symbolic interpretation and rule composition, so this result supports its ability to reason about abstract rules rather than directly proving completion rates in customer support, writing, or software development.

DeepSeek V4 Flash compared with V4 Pro, GLM-5.2, Claude Opus 4.8, and Kimi K3

The following mainly uses official-release comparisons published by DeepSeek; tool frameworks and reasoning settings may differ between models, so these figures are useful for assessing the direction of their capabilities. If you need to use a model in your business, however, you should rely on tests within that business.

Compared with DeepSeek V4 Pro: Flash surpasses the old preview, but not the Pro release across the board

V4 Flash does surpass the V4 Pro preview in several official agent evaluations, but compared with the official V4 Pro release, Terminal-Bench 2.1 is 82.7 versus 87.9 and DeepSWE is 54.4 versus 62.7, with Pro still ahead. Our judgment is that the progress of V4 Flash is enough to change its value in everyday development, but the capability gap with the official Pro release still needs to be recognized. For frequent work with easily verified results, we place greater value on the efficiency positioning of Flash; for highly complex work where failure leads to costly rework, the higher completion rate of Pro may still make it your first choice.

Compared with GLM-5.2: Flash leads in some execution evaluations, with differences across business scenarios

In the public DeepSeek comparison, V4 Flash scores 70.3 on Toolathlon-Verified, versus 59.9 for GLM-5.2; their DeepSWE scores are 54.4 and 46.2, respectively. We therefore believe that the advantage of V4 Flash in this set of evaluations mainly lies in tool collaboration and software engineering execution, rather than relying only on low cost to remain competitive. In our customer feedback, some users still prefer GLM for complex coding problems, while others report that V4 Flash works well in their codebases. We therefore think their performance may be better or worse in different business scenarios. We recommend testing based on your own business scenarios.

Compared with Claude Opus 4.8: close terminal scores, a more pronounced development gap

V4 Flash and Claude Opus 4.8 score 82.7 and 85.0, respectively, on Terminal-Bench 2.1 in the official comparison; on NL2Repo, their scores are 54.2 and 69.7. For us, the closeness in terminal evaluations supports including V4 Flash among candidates for everyday development tools; the development gap reminds me not to assign V4 Flash to every complex engineering task. It is a model worth serious consideration, but a single score does not establish it as a comprehensive replacement.

Compared with Kimi K3: capability gaps remain in complex engineering and difficult reasoning

In the DeepSeek comparison table, Kimi K3 scores 67.5 on DeepSWE, above the 54.4 of Flash 0731; their HLE scores without tools are 43.5 and 37.8, respectively. These results incline me to believe that Kimi K3 retains an advantage in these complex engineering and difficult reasoning evaluations. At the same time, Kimi officially states that some of its coding results were evaluated using the Kimi Code framework, reminding us that delivered results depend not only on the model but also on the execution framework. Our overall judgment is that V4 Flash cannot replace every stronger model; rather, its response speed and cost make this combination of capability and efficiency especially worth attention for tasks with clear acceptance criteria that require repeated calls.

Usage-based billing and model cost optimization

Charges follow current API Hop rates and actual usage. Top up on demand.

Unified AI models and data APIs, with multi-model access in one place

One API key connects multiple AI models and data services, reducing integration costs and letting you switch with one click. Business applications need both AI and data.

Call logs and application access controls, with clarity on every call

Every API call is recorded in your dashboard. Create separate API keys for different applications.

Evaluate outputs against your task

Use the documented model ID and supported parameters through API Hop. Evaluate outputs against your own acceptance criteria, then inspect usage records and response metadata before production use. A model name or example does not guarantee each result.

Prepare your account and test online

Sign in, confirm your balance, and create an API key.

Try it in the Playground

Explore the model's capabilities in the Playground to find the best way to use it.

Integrate DeepSeek V4 Flash into your business

Follow our developer documentation to integrate DeepSeek V4 Flash into your business.

Future model upgrades

To switch to a newer model, first confirm that its exact ID is returned by the model catalog for your API key. If it supports the same endpoint and request fields, update the model ID, then test response formats, streaming, output limits and billing before sending production traffic. For asynchronous image or video tasks, also verify job submission, status polling and result retrieval.