DeepSeek V4-Flash-0731 Launches: Major Coding Gains, Unchanged Cost
On August 1st, DeepSeek opened public access to its V4-Flash-0731 model. Instead of scaling up parameters, this release focuses on refined post-training techniques and a systematic overhaul of its Agent framework, achieving significant leaps within the existing architecture.
Benchmark Breakthroughs: Coding Prowess Takes a Leap
The most striking improvements are seen in key AI programming evaluations:
- DeepSWE Benchmark: Score surged from 7.3 to 54.4, indicating vastly better comprehension and execution of software engineering tasks.
- Cybergym Evaluation: Score doubled to 76.7, demonstrating enhanced capability in complex, dynamic code generation and security challenges.
- Hard Benchmark Performance: Now approaches the level of top-tier models like Claude Opus 4.8, signaling increased competitiveness in high-stakes arenas.
For developers relying on AI for code assistance, script automation, or debugging complex issues, the new version promises more reliable and precise support.
For Developers: Seamless API Upgrade, Enhanced Value
The update brings tangible benefits for API users. The V4-Flash-0731 model maintains the exact same pricing, meaning no extra cost for the performance boost. It also natively integrates the Responses API, providing better foundation for building interactive AI applications.
Existing users will have their API services automatically transitioned to the new version seamlessly. The current update applies to the Flash API, while the official V4 Pro release remains under active development.
This upgrade underscores a key point: an AI model's value isn't just about parameter count, but the precision of its algorithmic optimization and engineering. Continuously improving core capabilities while keeping access affordable could be a crucial step for broader AI adoption.