Last Updated: July 31, 2026
DeepSeek just pulled off something unprecedented in AI. On July 31, 2026, the company released the official DeepSeek-V4-Flash API into public beta, and its agent benchmarks don't just beat the preview version. They beat DeepSeek's own V4-Pro-Preview, the more expensive, larger model in the same lineup.
According to DeepSeek's official API documentation, V4-Flash-0731 scored 82.7 on Terminal Bench 2.1, compared to V4-Pro-Preview's 72.1. That is a 14.7% win for the budget model over its premium sibling. At $0.14 per million input tokens, this rewrites the value equation for AI agent development.
What Makes DeepSeek V4-Flash a Big Deal
DeepSeek V4-Flash is the lightweight model in DeepSeek's V4 lineup, designed for cost-efficient deployment. The official release (model identifier: deepseek-v4-flash) entered public beta on July 31, 2026, with an API that remains fully backward-compatible. Developers simply set the model name to deepseek-v4-flash and everything works.
The breakthrough here is not a bigger model. It is better training. V4-Flash-0731 uses the exact same architecture and size as the preview version. The only change was re-post-training. Same parameters, radically different agent performance. This suggests DeepSeek found a training methodology that unlocks agentic capability without scaling up compute requirements.
The Benchmark Numbers: V4-Flash vs Everyone
According to DeepSeek's benchmark data released July 31, 2026, here is how V4-Flash-0731 stacks up against its preview, V4-Pro-Preview, GLM-5.2, and Anthropic's Opus-4.8:
| Benchmark | V4-Flash (Official) | V4-Flash (Preview) | V4-Pro-Preview | GLM-5.2 | Opus-4.8 |
|---|---|---|---|---|---|
| Terminal Bench 2.1 | 82.7 | 61.8 | 72.1 | 81.0 | 85.0 |
| NL2Repo | 54.2 | 39.4 | 38.5 | 48.9 | 69.7 |
| Cybergym | 76.7 | 38.7 | 52.7 | - | 83.1 |
| DeepSWE | 54.4 | 7.3 | 12.8 | 46.2 | 58.0 |
| Toolathlon Verified | 70.3 | 49.7 | 55.9 | 59.9 | 76.2 |
| Agents' Last Exam | 25.2 | 15.8 | 16.5 | 23.8 | 25.7 |
| AutomationBench (Public) | 25.1 | 10.8 | 12.8 | 12.9 | 27.2 |
| DSBench-FullStack | 68.7 | 37.0 | 41.8 | 61.8 | 71.6 |
| DSBench-Hard | 59.6 | 25.8 | 31.1 | 54.5 | 71.7 |
The pattern is clear: V4-Flash beats V4-Pro-Preview on every single agent benchmark. On DeepSWE, the improvement from preview to official is staggering: 7.3 to 54.4, a 645% improvement from the same model size through re-post-training alone.
Pricing: The Most Cost-Effective Frontier Agent Model
According to DeepSeek's pricing page, V4-Flash costs:
- Cache-miss input: $0.14 per 1 million tokens
- Cache-hit input: $0.0028 per 1 million tokens
- Output: $0.28 per 1 million tokens
For context, Anthropic's Opus-4.8, which leads on most benchmarks, costs significantly more per million tokens. V4-Flash delivers 82.7 on Terminal Bench 2.1 versus Opus-4.8's 85.0, a gap of just 2.3 points, at a fraction of the price. For teams running agent workloads at scale, that math matters enormously.
One important detail: thinking mode is on by default and bills at the output rate, even though those reasoning tokens are not visible in the final response. Budget accordingly.
Responses API and Codex Support
The official V4-Flash now natively supports the OpenAI Responses API format, meaning developers already building on OpenAI's ecosystem can switch to V4-Flash with minimal code changes. DeepSeek has also specifically adapted the model for Codex integration, with configuration documented on their developer portal.
This is a strategic move. By supporting the Responses API natively, DeepSeek is positioning V4-Flash as a drop-in replacement for OpenAI models in agent pipelines. Combined with the pricing, this makes V4-Flash the strongest "swap-one-line-of-code and save 90%" candidate on the market.
How V4-Flash Compares to V4-Pro: Same Family, Different League
The most remarkable part of this release is the internal story. DeepSeek has two V4 models: Flash (lightweight) and Pro (full-size). The Flash model was supposed to be the cheaper, slightly worse option. Instead, through better post-training, the budget model now outperforms the premium model's preview version across the board.
DeepSeek confirmed in their change log that this update only applies to the V4-Flash API. The V4-Pro API and the models on DeepSeek's app and website remain unchanged. The official V4-Pro release will follow soon, which raises an obvious question: if the budget model is this good now, what will V4-Pro look like when it ships?
What This Means for AI Agent Development
For teams building AI agents, coding assistants, or automated workflows, V4-Flash changes the build-vs-buy economics. At Flowtivity, we have been testing DeepSeek models for agent workloads since the V3 series. The V4-Flash benchmarks suggest you can now get near-frontier agent performance at roughly 1/10th the cost of premium alternatives.
The practical implications:
- Coding agents get a model that scores 82.7 on Terminal Bench 2.1 for pennies per task
- SWE pipelines can run DeepSWE at 54.4, competitive with models costing 10x more
- Tool-using agents get Toolathlon Verified scores of 70.3, behind only Opus-4.8
- Full-stack development tasks score 68.7 on DSBench-FullStack, within 3 points of Opus-4.8
The "good enough and cheap" thesis that defines the current AI market just got its strongest data point yet.
Getting Started with DeepSeek V4-Flash
To use the new model, set your API model parameter to deepseek-v4-flash. The base URL and authentication remain unchanged from previous DeepSeek API versions. Full documentation is available at api-docs.deepseek.com.
For Codex-specific configuration, DeepSeek has published a dedicated integration guide.
"The model that costs less should not beat the model that costs more. DeepSeek just proved that better training beats bigger parameters. That is the whole story."
