DeepSeek took V4 out of preview today. V4-Pro moved to the 0813 build, V4-Flash to 0731, and both are now general availability. We checked our production routing this morning. The unversioned aliases auto-resolve on DeepSeek's endpoint, which means every ClawBox AI session is already hitting the new builds. There is nothing for you to restart and nothing to configure.
Before any numbers, the caveat that matters. Everything below is DeepSeek's own reported figures from their August release notes. Independent evaluators are still running their suites against 0813. We are publishing these because the size of the movement is worth knowing about now, not because anyone outside DeepSeek has confirmed them yet.
The DeepSWE result is the one to look at
DeepSWE measures whether a model can trace dependencies through a real multi-file codebase and actually fix the bug. It is the benchmark that best predicts whether an agent is useful on your repository rather than on a toy problem.
V4-Pro scored 12.8 in preview. The 0813 build scores 62.7. That is close to a fivefold move on the metric that matters most for the work our users actually do.
| Benchmark | V4-Pro preview | V4-Pro 0813 | V4-Flash preview | V4-Flash 0731 |
|---|---|---|---|---|
| DeepSWE | 12.8 | 62.7 | 7.3 | 54.4 |
| Terminal-Bench 2.1 | 72.1 | 87.9 | 61.8 | 82.7 |
| CyberGym | 52.7 | 83.3 | 38.7 | 76.7 |
| DSBench-FullStack | 41.8 | 71.1 | 37.0 | 68.7 |
| DSBench-Hard | 31.1 | 67.2 | n/a | n/a |
| NL2Repo | 38.5 | 61.5 | n/a | n/a |
| Toolathlon-Verified | 55.9 | 74.1 | n/a | n/a |
Where it wins, and where it does not
V4-Pro takes CyberGym at 83.3, just past Fable 5 on 83.1. On Terminal-Bench 2.1 it lands at 87.9, ahead of Opus-4.8 and a whisker behind Fable 5 and Kimi-K3.
It loses DeepSWE. Fable 5 sits at 70.0 and Kimi-K3 at 67.5, both clearly ahead of 62.7. We are showing you that row because you will find it anyway, and because a model that closes this much ground in one release is interesting without needing to win every column.
| Model | Terminal-Bench 2.1 | CyberGym | DeepSWE |
|---|---|---|---|
| DeepSeek V4-Pro | 87.9 | 83.3 | 62.7 |
| Kimi-K3 | 88.3 | 80.0 | 67.5 |
| Fable 5 | 88.0 | 83.1 | 70.0 |
| Opus-4.8 | 85.0 | 78.3 | 58.0 |
| GLM-5.2 | 81.0 | n/a | n/a |
Which model your tier is running
The 62.7 belongs to V4-Pro, and V4-Pro serves the Max tier. On the Pro tier your requests route to V4-Flash-0731. Our plan is called Pro and it does not run the model called Pro. That naming collision is ours, not DeepSeek's, and we would rather put it in writing than let you infer it.
Flash is a serious model in its own right. It moved from 7.3 to 54.4 on DeepSWE and hits 82.7 on Terminal-Bench 2.1, which places it above where V4-Pro sat in preview a few weeks ago.
- Max tier: V4-Pro-0813 by default, 1M token context.
- Pro tier: V4-Flash-0731, 1M token context.
- Free tier: V4-Flash-0731 with a smaller daily budget.
One thing that will bite you quietly
V4 is text-only, and it fails silently on images. The official API replaces an image block with placeholder text rather than returning an error, so the model will answer confidently about a picture it never received.
Keep V4 prompts to code and text. If your workflow leans on screenshots or diagrams, route those to a vision-capable model instead.
Go break something
The upgrade already happened, so there is nothing to claim. Open a session and throw a genuinely messy multi-file refactor at it. Dependency tracing is where the difference shows up fastest.
All benchmark figures on this page are DeepSeek's own published results, August 2026.