DeepSeek V4-Flash API Hits Public Beta With Agent Benchmarks That Beat V4-Pro-Preview
Nine agent tasks tested. V4-Flash-0731 won all of them against V4-Pro-Preview - and now slots directly into Claude Code, GitHub Copilot, and OpenCode without a new integration.
DeepSeek V4-Flash API entered public beta today. Announced via @deepseek_ai on X at 11:56 AM on July 31, drawing 8.1 million views in the first hours, the official release ships as version DeepSeek-V4-Flash-0731 and surpasses V4-Pro-Preview across all nine agent benchmark tasks published in the announcement. For developers already using the deepseek-v4-flash model string, no endpoint change is required - DeepSeek updated the model in place.
V4-Flash-0731 Beats V4-Pro-Preview on All Nine Agent Benchmarks
Nine benchmarks, zero losses. V4-Flash-0731 scores 82.7 on Terminal Bench 2.1 - which tests shell commands and file system tasks in a live terminal environment - compared to V4-Pro-Preview's 72.1. DeepSWE shows the sharpest gap: 54.4 for V4-Flash-0731 against 12.8 for V4-Pro-Preview, measuring whether a model can write patches that actually pass existing test suites. On GLM-5.2's benchmark territory - NL2Repo and DSBench-FullStack, where GLM-5.2 scored 48.9 and 61.8 respectively - V4-Flash-0731 hits 54.2 and 68.7, ahead on both.
| Benchmark | V4-Flash-0731 | 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 |
Responses API and Codex Support Make V4-Flash a Backend Swap for Existing Tools
Codex support matters here. V4-Flash now handles both OpenAI and Anthropic API formats natively - developers can point Claude Code, GitHub Copilot, or OpenCode at api.deepseek.com without writing any new integration code, since the same base_url swap that works for OpenAI-compatible tools applies directly. DeepSeek's pricing page still lists Flash as the lower-cost tier relative to V4-Pro, so teams that switched to agent-grade performance today pay inference-tier prices for it. A comparison of Claude Code, GitHub Copilot, and Cursor shows how fragmented the coding agent backend market is - DeepSeek V4-Flash API now slots into all three as a direct model replacement.
Two of the Nine Benchmarks Are DeepSeek's Own - Here Is What That Changes
Two benchmarks lack independent verification. DSBench-FullStack and DSBench-Hard - both showing V4-Flash-0731 scores of 68.7 and 59.6 - are DeepSeek's internal sets, built and evaluated by DeepSeek itself. Publishing proprietary test suites alongside public ones lets a lab report strong numbers on tasks it designed, without third-party replication; labs across the board do it, but it matters more when the proprietary benchmarks are also the ones showing the largest gaps. On fully public tests, V4-Flash-0731 scores 25.2 on Agents' Last Exam against Claude Opus 4.8's 25.7 and 25.1 on AutomationBench Public against Opus 4.8's 27.2 - close, but not level.
Public beta means the performance and pricing are still subject to adjustment before a stable production release. DeepSeek has not announced a timeline for moving V4-Flash out of beta.