Published: July 31, 2026
DeepSeek released V4-Flash-0731 as a public API beta on July 31, 2026, and the smaller model now outscores the company’s own flagship on agent tasks. V4-Flash-0731 scores 82.7 on Terminal-Bench 2.1 against 72.1 for V4-Pro-Preview and 61.8 for the previous V4-Flash-Preview. DeepSeek changed nothing in the architecture: the model keeps its 284B parameter count and redoes only post-training, targeted at coding agents, tool use, and multi-step tasks. The update applies to the V4-Flash API interface alone — the V4-Pro API, the official app, and the web product are unchanged.
What V4-Flash-0731 scores across agent benchmarks
V4-Flash-0731 posts 7 published benchmark results, led by Terminal-Bench 2.1 at 82.7. The full set covers Toolathlon Verified at 70.3, Cybergym at 76.7, DSBench-Hard at 59.6, DeepSWE at 54.4, NL2Repo at 54.2, and Agent Last Exam at 25.2.
Terminal-Bench 2.1 is the score that matters most for business developers. The benchmark measures whether an AI agent can navigate a Linux terminal, run commands, read the output, and finish a multi-step task without human intervention — the exact loop a coding agent or an automation workflow runs in production.
The jump from 61.8 to 82.7 came from post-training alone, such as improved agent-task training data and reinforcement learning tuned to tool-calling. A capability gain of 20.9 points with no architecture change is evidence that agent performance is now a training problem rather than a scale problem.
Responses API support makes V4-Flash a drop-in agent backend
V4-Flash-0731 now supports the Responses API format, the request-response standard OpenAI introduced for agent loops in which a model calls a tool, receives the result, and continues. Development teams that built their agent stack against the Responses API can point it at V4-Flash without rewriting the orchestration layer.
That compatibility is the commercially significant part of the release. A benchmark score changes a shortlist; an interface match changes a migration cost. DeepSeek API access remains among the lowest-priced frontier options globally, though DeepSeek did not publish per-token pricing for V4-Flash-0731 in its release notes — verify current rates on the DeepSeek pricing page before modelling costs.
The agent price-capability race between US and Chinese labs
OpenAI cut GPT-5.6 Luna prices by 80% on July 30, 2026, one day before this release. DeepSeek’s answer arrived as capability rather than a discount, and it targets the same buyer: the team building tool-calling agents who treats cost per completed task as the deciding metric.
The pattern repeats across both labs. A smaller, specifically tuned model beats a larger generalist flagship on agentic benchmarks — GPT-5.6 Luna did this to its bigger siblings, and V4-Flash-0731 now does it to V4-Pro-Preview. Analysts at OfficeChai rate V4-Flash-0731 as delivering Opus 4.8-level performance at a fraction of the price; that comparison is third-party analysis, not a DeepSeek claim.
What this changes for teams choosing AI automation tools
Teams evaluating agent models should benchmark the cheap tier against the premium tier on their own tasks before signing an enterprise contract. The default assumption — that the flagship model is the strongest option for every workload — no longer holds on agentic benchmarks at either DeepSeek or OpenAI. Benchmark comparisons stay task-specific: V4-Flash-0731 leads V4-Pro-Preview on Terminal-Bench 2.1, which is not a claim about reasoning, writing, or long-context retrieval. Our roundup of the 12 best AI agents for business tasks covers which platforms let you swap the underlying model.
Our Take
DeepSeek just made its cheapest model its best one for the work businesses are actually automating this year — coding agents, workflow orchestration, tool calling. The tell is that nothing about the model got bigger. Post-training is now where agent capability comes from, which means the price of a competent agent backend falls faster than the price of raw intelligence. Run your own eval before your next contract renewal.
For Context
WithO2 has tracked DeepSeek’s push toward independent infrastructure — read DeepSeek Is Building Its Own AI Chip to Break Free From Nvidia. The pricing pressure driving this release started with OpenAI’s tiered launch, covered in OpenAI Launches GPT-5.6 Sol, Terra and Luna. For a broader view of the tooling landscape these models plug into, see our guide to the 15 best AI tools for business.
