DeepSeek is widening access to its latest AI model, betting that better autonomous task performance and low pricing will help it win more developers.
Chinese AI startup DeepSeek has launched the public beta of its DeepSeek V4-Flash application programming interface (API), giving developers broader access to the company's latest flagship lightweight model ahead of the full release of V4-Pro.
The public beta follows preview releases introduced in April. DeepSeek said the official version of V4-Pro "will follow soon."
According to the company, the latest V4-Flash update focuses on improving AI agent capabilities — the ability of models to complete complex, multi-step tasks by planning actions and using external tools. DeepSeek said the model delivers "Significantly enhanced agent capabilities, with benchmark results far exceeding V4-Pro-Preview."
The company also said the model now natively supports the Responses API format and is compatible with Codex-based developer workflows, making it easier for developers to integrate into existing AI applications.
DeepSeek said the official V4-Flash release keeps the same architecture and parameter count as the preview version, with improvements coming entirely from post-training rather than changes to the underlying model.
The company reported stronger results across several coding and agent-focused benchmarks, including Terminal Bench 2.1, Cybergym, DeepSWE, NL2Repo, Toolathlon Verified, and its internal DSBench evaluations. In multiple tests, the updated V4-Flash outperformed the earlier V4-Pro preview despite being the smaller model.
For now, the upgrade is limited to the API. DeepSeek's web interface, mobile app, and V4-Pro API remain unchanged.
Pricing strategy heats up AI competition
The launch comes as competition over AI pricing continues to intensify.
DeepSeek said it plans to introduce peak-hour pricing once the official V4 models are released, charging double the standard API rate during periods of highest demand. Even with those higher prices, industry comparisons indicate the company's models would remain significantly cheaper than competing offerings from major US AI providers.
The company has already helped reshape China's AI market through aggressive pricing. Earlier this year, it permanently reduced V4 model prices by 75%, prompting rivals including ByteDance and Tencent to lower their own prices.
What this means for developers and businesses
The latest release reflects a growing shift in the AI industry. Rather than relying solely on ever-larger models, companies are increasingly trying to improve performance through better training while keeping costs under control.
For developers, that could translate into faster and less expensive AI agents for coding, software engineering, automation, and other enterprise workflows without requiring the largest available models.
Businesses evaluating AI platforms may also benefit from broader compatibility. Support for the Responses API format and Codex-based tooling lowers migration barriers for teams already building around established AI development frameworks.
The downside is that the newest capabilities are currently available only through the API, meaning users of DeepSeek's web and mobile products will have to wait until those updates arrive.
Also read: AI detector backlash shows how Substack’s Pangram-powered screening tool is testing writer trust as false-positive concerns create reputational risks.