DeepSeek released the updated V4-Flash through a public API beta on July 30 and formally launched V4-Pro on August 13, completing its two-tier V4 lineup for building AI agents. Developers now have to decide whether demanding work needs Pro’s added capability or whether the faster, cheaper Flash can complete it well enough.

That choice carries a substantial price difference. Pro’s current API rates are three times Flash’s rates for uncached input and output, and both models cost twice as much during DeepSeek’s seven daily peak hours. Teams can lower the bill by using Flash, scheduling flexible work during cheaper hours, or doing both, but jobs that need Pro and must run immediately leave less room to control costs.

DeepSeek is also expanding what developers can build around the models. Pro and Flash support several reasoning levels and widely used API formats, while the open-source DeepSeek Harness lets teams replace or extend an agent’s tools, workflow, storage, execution environment, and interface. The Harness remains a developer preview, with compatibility-breaking updates expected and no published evidence of production reliability or enterprise support.

Software teams, AI product developers, and businesses running automated workflows will need to test complete jobs from their own operations. Flash finished one point behind Pro in Artificial Analysis’s combined independent evaluation, responded faster, and completed the test tasks at a lower average cost. Those results support Flash as a credible lower-cost option, but they do not show whether either model will produce acceptable work consistently in a particular company workflow.

In short, DeepSeek now gives developers a premium model, a cheaper alternative, and an open framework for building the agent system around them. Buyers need to determine whether Flash can handle their work, whether Pro prevents enough failures and rework to justify its higher price, and whether the preview-stage Harness provides enough useful control to outweigh its maintenance demands.

An AI agent model provides the reasoning used to complete a task, while an agent harness is the surrounding software that manages the model’s tools, workflow, sessions, execution environment, and interface.

DeepSeek V4 is a two-tier lineup of text-only AI models that pairs the premium V4-Pro model for demanding agent work with the faster, cheaper V4-Flash alternative.

DeepSeek introduced V4-Pro and V4-Flash in preview form in April, laying out a two-tier approach to building AI agents. The company then released the updated Flash model through a public API beta on July 30 and formally launched V4-Pro on August 13, establishing the current lineup.

Pro is the larger model, with 1.6 trillion parameters, of which 49 billion are active when it handles a request. Flash has 284 billion parameters and uses 13 billion at a time. Parameters are the values a model learns during training, while the active count shows how much of that learned capacity it calls on for a particular request.

Those differences support distinct roles within the lineup. DeepSeek presents Pro as its most capable option and Flash as the smaller, faster, and cheaper choice. The company says Flash’s reasoning comes close to Pro and can match it on “simple Agent tasks,” although it does not explain which tasks qualify as simple. That leaves Flash as a plausible option for less demanding work, while Pro is meant to handle jobs that need more capability.

The two models otherwise share much of the same foundation. Both can work with as many as one million tokens in a single interaction, allowing them to consider a large amount of text at once. They accept and produce text only, however, with no native support for images, audio, or video listed in the supplied independent evaluation. The one-million-token limit therefore describes how much text they can process, rather than a wider range of media they can understand.

Developers can also choose how much reasoning the models apply. DeepSeek offers low effort for simpler tasks, high effort for daily agent work, and maximum effort for more complex jobs, along with separate thinking and non-thinking modes. This gives teams a way to adjust how deeply either model works instead of treating every request the same.

The updates did not require developers to switch API model names, reducing the work required to move to the new versions. Both current models also gained native support for OpenAI’s Responses API, including a one-click Codex setup, while DeepSeek continues to support OpenAI Chat Completions and Anthropic-compatible workflows.

DeepSeek says the latest Pro release brings major agent upgrades and gains in production work, but its announcement includes neither customer results nor an explanation of how those gains were measured. Developers can access Pro through the API or use it in Expert Mode through DeepSeek’s app and website. If both tiers are built for agent work, when does Pro add enough capability to justify moving beyond Flash?

DeepSeek raised API prices for both V4-Pro and V4-Flash as part of the August release, while introducing rates that change according to the time of day. The new structure took effect at 16:00 UTC on August 16 (9:00 a.m. Pacific Daylight Time) and divides usage into peak and off-peak periods. Off-peak rates are 50% lower than the corresponding peak rates.

The company also charges different rates for cached input, uncached input, and output. Cached input is material from a prompt that DeepSeek has already processed and can find in its cache, while uncached input is new material the model must process. During off-peak hours, Flash costs $0.007 per million cached input tokens, $0.22 per million uncached input tokens, and $0.66 per million output tokens. Its peak rates rise to $0.014, $0.44, and $1.32, respectively.

Pro follows the same pricing structure at a higher level. Its off-peak rates are $0.022 per million cached input tokens, $0.66 per million uncached input tokens, and $1.98 per million output tokens. During peak hours, those rates increase to $0.044, $1.32, and $3.96.

The peak windows run from 01:00 to 04:00 UTC and again from 06:00 to 10:00 UTC each day. In Pacific Daylight Time, those periods run from 6:00 to 9:00 p.m. and from 11:00 p.m. to 3:00 a.m. Every other hour receives the off-peak rate. Developers with flexible workloads can therefore lower their API costs by scheduling jobs outside those seven peak hours. Teams that need an agent to respond immediately may have little control over when the expense occurs.

Reuters reported that Pro was about nine times more expensive than Flash for input and 14 times more expensive for output when the model launched on August 13. Those calculations used Flash’s prices at the time: $0.14 per million input tokens and $0.28 per million output tokens. DeepSeek’s new Flash rates took effect three days later, so the earlier ratios no longer describe the price difference between the models under the same current rate period.

At the new rates, Pro costs exactly three times as much as Flash for uncached input and output during both peak and off-peak hours. Its cached-input rate is about 3.14 times higher. Teams can control costs by choosing Flash, moving flexible work into off-peak hours, or doing both. Work that needs Pro’s additional capability and must run during peak periods leaves less room to reduce the bill.

So, are Pro’s stronger results large enough to justify its higher price, or can Flash handle enough everyday agent work to remain the better value?

DeepSeek reports that both current models beat their respective preview versions on every benchmark in its tables. On Terminal Bench 2.1, which tests coding agents working through computer-terminal tasks, Pro rose from 72.1 in preview to 87.9, while Flash increased from 61.8 to 82.7. On DeepSWE, another coding benchmark, Pro jumped from 12.8 to 62.7 and Flash rose from 7.3 to 54.4.

Against competing models, Pro led some tests, while other models posted the highest score on others. Its strongest areas therefore depended on the work being measured. DeepSeek’s table supports the claim that Pro improved substantially over the preview version, but it does not establish an across-the-board lead over every rival.

DeepSeek used its upcoming Harness in minimal mode for the public coding-agent tests. Within that setup, it allowed the models to use maximum reasoning effort. DeepSeek does not explain what its minimal Harness configuration includes.

The company also set top_p to 0.95 and temperature to 1.0, two settings that control how widely a model chooses among possible responses. The agent framework, reasoning level, and response settings can all affect the result. DeepSeek therefore warns that developers may see different scores with another framework. DeepSeek identifies two tests in the table, DSBench-FullStack and DSBench-Hard, as its internal benchmarks, so those results are company evidence rather than independent verification.

Artificial Analysis independently tested the current V4-Pro-0813 and V4-Flash-0731 releases at maximum reasoning effort. It combined nine evaluations covering agent work, tool use, coding, scientific reasoning, knowledge, and long documents into one Intelligence Index score. Pro scored 53 and ranked third among 107 comparable open-weight models. Flash scored 52 and ranked fifth. That one-point difference supports DeepSeek’s claim that Flash can approach Pro’s reasoning ability.

Flash also generated its answers faster. Artificial Analysis measured 107.2 output tokens per second for Flash and 80.2 for Pro. Flash began returning an answer in 1.19 seconds, compared with 1.74 seconds for Pro. That speed can add up when an agent must make repeated model calls to complete one assignment.

Artificial Analysis also compared what the models cost under the same standardized tests. Using DeepSeek’s peak prices and a fixed mix of cached input, new input, and output, Artificial Analysis calculated a blended price of $0.69 per million tokens for Pro and $0.23 for Flash.

The difference narrowed when Artificial Analysis calculated the average cost of completing one test task. Pro cost $0.25 per task and Flash cost $0.11, making Pro about 2.3 times as expensive instead of three times as expensive. Cost per task accounts for both the price of the tokens and how many the model used while completing the same independent tests.

That second factor worked against Flash’s lower price. Across all the tasks included in the Intelligence Index, Flash generated 210 million output tokens, compared with 130 million for Pro. Artificial Analysis does not explain why Flash used 80 million more tokens. Even with that heavier use, Flash’s lower rates kept its average task cost below Pro’s.

Artificial Analysis’s results make Pro’s premium depend on whether its one-point advantage changes the outcome of the task. If both models produce acceptable work, Flash’s faster responses and lower average cost give it the stronger value case.

Because the Intelligence Index combines nine different evaluations, its one-point gap cannot tell a business how the models will compare on its particular work. Failed attempts, retries, and employee review would also change the final expense. Teams will need to test complete tasks from their own workflows to learn whether Flash can produce acceptable work consistently or whether Pro’s additional capability reduces enough rework to justify its higher price.

Once developers choose the appropriate model tier, how much control does DeepSeek give them over the surrounding agent system?

DeepSeek is extending its agent offering beyond Pro and Flash with a separate open-source project called DeepSeek Harness. An agent harness is the software surrounding a model that manages its tools, workflow, sessions, execution environment, and user interface.

DeepSeek built the Harness around one idea: “Everything is a plugin.” Models, tools, and skills are plugins, as are sessions, sandboxes, and filesystems. The loops that govern how an agent works through a task, the orchestration that coordinates its actions, and the interface presented to the user can also be replaced.

Because those parts use the same plugin structure, developers can mix, replace, and extend them without adopting DeepSeek’s complete setup. A team could keep one component, exchange another, and add its own tools or workflow. That gives developers control over more of the agent system than selecting a model through an API alone.

The Harness is powered by the Cordis meta-framework, which supplies the underlying plugin structure. DeepSeek released the code under the MIT license, allowing developers to reuse and modify it within the license terms. They can also start its local web interface with the command npx @deepseek-ai/dsh web. By default, that interface runs on the developer’s own computer at 127.0.0.1:3080.

The current Pro and Flash models have been released, but Harness v0.1 remains a developer preview. DeepSeek says the Harness is changing rapidly and warns that future updates will break compatibility with some earlier versions. Developers who begin experimenting with it now may therefore need to update their code as its plugin interfaces and other components change.

DeepSeek has not released evidence showing how reliably the Harness performs in day-to-day production, whether its interfaces will remain stable, what enterprise support it will provide, or how widely developers have adopted it. Its preview status makes the Harness better suited to testing and experimentation, while the missing reliability and support details leave companies without the evidence they would need to judge a long-term rollout.

Developers can switch between models within the Harness, but DeepSeek does not say it will automatically choose Pro or Flash for each task. DeepSeek has not described automatic model selection based on task difficulty, token price, or whether the request occurs during peak hours. A team that wants that kind of automatic routing would need to add it separately.

Choosing a DeepSeek setup now involves more than selecting Pro or Flash. Teams must determine whether Flash can complete their work or whether Pro performs enough better to justify its higher price. Flexible workloads may also cost less when they run outside peak hours.

Beyond the model choice, teams must determine how much control they need over the tools, workflow, execution environment, and interface surrounding it. They must also decide whether they can accept compatibility-breaking updates and the current lack of documented production reliability and enterprise support.

DeepSeek now offers current models and an open framework for building around them, giving developers a more complete alternative for agent development. Its value will depend on the cost of producing acceptable work and whether the preview-stage Harness provides enough control to justify the additional maintenance it may require.

Q: What did DeepSeek release with V4-Pro and V4-Flash?
A: DeepSeek introduced V4-Pro and V4-Flash in preview form in April, released the updated V4-Flash through a public API beta on July 30, and formally launched V4-Pro on August 13. The releases created a two-tier lineup for AI agent development, with Pro intended for demanding work and Flash serving as the faster, cheaper option.

Q: What is the difference between DeepSeek V4-Pro and V4-Flash?
A: DeepSeek presents V4-Pro as its larger and more capable agent model. V4-Flash is smaller, responds faster, and costs less. DeepSeek says Flash’s reasoning comes close to Pro and can match it on some “simple Agent tasks,” although the company has not defined which tasks qualify as simple. Both models process up to one million tokens of text in one interaction, support several reasoning levels, and accept and produce text only.

Q: How much more does DeepSeek V4-Pro cost than V4-Flash?
A: Under DeepSeek’s current API pricing, V4-Pro costs three times as much as V4-Flash for uncached input and output during the same rate period. Pro’s cached-input rate is about 3.14 times higher. Both models cost twice as much during DeepSeek’s seven daily peak hours as they do during off-peak hours.

Q: When is the DeepSeek API cheaper to use?
A: DeepSeek charges its 50% lower off-peak rates outside 01:00 to 04:00 UTC and 06:00 to 10:00 UTC each day. In Pacific Daylight Time, the peak periods run from 6:00 to 9:00 p.m. and 11:00 p.m. to 3:00 a.m. Teams can reduce API costs by scheduling flexible work outside those periods, while agents that must respond immediately offer less control over when the expense occurs.

Q: Does DeepSeek V4-Pro perform much better than V4-Flash?
A: The difference depends on the work being tested. Artificial Analysis gave V4-Pro an Intelligence Index score of 53 and V4-Flash a score of 52 at maximum reasoning effort. Flash responded faster and completed the standardized test tasks at an average cost of $0.11, compared with $0.25 for Pro. That combined result gives Flash a stronger value case when both models produce acceptable work, but it cannot show whether Pro will prevent enough failures or rework to justify its higher price in a particular business workflow.

Q: What is DeepSeek Harness, and does it automatically choose between Pro and Flash?
A: DeepSeek Harness is an open-source, plugin-based framework for building AI agents. It lets developers replace or extend an agent’s models, tools, skills, workflows, sessions, storage, execution environment, and interface. Developers can switch between Pro and Flash within the Harness, but DeepSeek has not described automatic model selection based on task difficulty, API price, or peak hours. A team that wants automatic routing would need to add it separately.

Q: Is DeepSeek Harness ready for production use?
A: DeepSeek describes Harness v0.1 as a developer preview and warns that upcoming changes may break compatibility with earlier versions. The company has not published evidence about day-to-day production reliability, stable interfaces, enterprise support, or developer adoption. Teams can experiment with the Harness now, but they do not yet have the evidence needed to judge a long-term production rollout.

DeepSeek now offers two current agent models, V4-Pro and V4-Flash, and an open-source framework for building agents around them. This gives developers more control over capability, cost, and infrastructure, while leaving each team to determine which combination can produce reliable work at an acceptable total cost.

Artificial Analysis found only a one-point difference between V4-Pro and V4-Flash in its combined independent evaluation. Flash also responded faster and completed the standardized test tasks at an average cost of $0.11, compared with $0.25 for Pro. Those results make cost per acceptable completed task more useful than either benchmark scores or token prices alone.

Software teams and AI product developers are most directly affected because they must determine which DeepSeek model can complete their agent tasks at an acceptable cost and speed. Businesses running automated workflows must account for variable API rates, retries, and employee review, while infrastructure teams building their own agent systems must decide whether DeepSeek Harness provides enough useful control over tools, workflows, storage, execution environments, and interfaces to justify its upkeep.

DeepSeek’s variable API rates connect operating costs to when work runs. Flexible background jobs can use cheaper hours, while customer-facing agents may incur peak rates whenever users need them. DeepSeek Harness adds another operating demand because its preview status and expected compatibility-breaking updates may require developers to update their code as its interfaces change.

Organizations evaluating DeepSeek should test V4-Pro and V4-Flash on complete tasks from their own workflows. Those tests should measure failed attempts, retries, token use, employee review, response time, and the final cost of acceptable work. Teams considering DeepSeek Harness must separately decide whether its added control justifies expected code changes and the lack of documented production reliability or enterprise support.

In short, V4-Flash offers the stronger value when it completes the work reliably. V4-Pro earns its higher price when it produces meaningfully better results or prevents enough failure and correction to justify the added expense. DeepSeek Harness is useful when its additional control is worth the maintenance its preview-stage interfaces may require.

DeepSeek now offers both the models and the framework for building an agent system. The fuller package comes with a fuller bill: Pro’s premium, peak-hour pricing, and the upkeep of a changing Harness.

Editor’s Note: This article was created by Alicia Shapiro, CMO of AiNews.com, with writing support, AEO/GEO/SEO optimization, image concept development, and editorial structuring support from ChatGPT, an AI assistant. All final editorial decisions, perspectives, and publishing choices were made by Alicia Shapiro.