GPT-6.2 is getting roasted on X
meanwhile, Astra and Sol just took #1 and #2 on my object detection benchmark
so I have some awkwardly positive results to share
↓ benchmarks and examples
I will not be writing a full review of GPT-6.2 Astra / Sol.
I had a terrible experience testing them. They were essentially unusable.
I do not recommend these models.
It’s pretty wild to take a step back and see how much @roboflow changed in a year. @braddwyer put together a great recap of what we shipped. And we got a lot more cool stuff in the works.
youtube.com/watch?v=GLqemlJa…
playground.roboflow.com/eval…
we have a big update to the evals coming soon. new harder tasks that will move scores from the 80s to the 60s.
let us know if you have specific tasks or types of data you'd want to see inlcuded
Gemini 4 Carbon is now the top overall vision model on our benchmarks
Text/OCR seems to be a consistent high performing area over others
- text localization and recognition: finding text, drawing boxes, and reading it
- short string extraction: reading a specific value from an image
- long text transcription: reading full passages from images
Beyond that, it's #1 in core vision tasks like object detection and reasoning.
Disclaimer: it is the slowest and most expensive of the top models so if you need cost/speed, this one isn't for you
Gemini 4 Carbon is so fucking good at computer vision
- #1 overall
- #2 object detection
- #1 short string extraction
- #4 structured JSON extraction
- #1 long text transcription
- #1 text localization and recognition
very slow and expensive though...
↓ examples
I fine-tuned V-JEPA 2.1 on 25 clips to learn the steps of an F1 pit stop 🏎️
• V-JEPA 2.1 (ViT-B), 384 px, fine-tuned on @roboflow
• Per-frame head: which steps are active + how far each extends
• 64 frames at a time (6.4 s at 10 fps)
• ~2 h to train
A detector can't do this robustly: wheel on and wheel off look the same in a still. The model has to see the motion.
We are releasing a new package from @roboflow, meet reid!
It turns a crop of an object into an embedding, so you can tell objects apart, like this player across clips of a game.
Can't wait to see what you build with it!
github.com/roboflow/re-ID
Re-Identification available in @roboflow Trackers!
Until now, when a player left the frame and came back, the tracker gave them a new ID.
Re-ID fixes that by comparing how players look, and it's now integrated in trackers v2.7.0.
The models come from our new open-source package reid, which also works on its own.
pip install "trackers[reid]" and start using it!
Joining the coalition places Roboflow Labs’ work within a shared effort to advance those capabilities.
We see an opportunity to connect progress in frontier models with the breadth of tasks developers face outside the lab, and to help turn that progress into tools people can build with.
Roboflow Labs draws on what the 2M+ developers and Fortune 500 enterprises building on Roboflow are actually trying to solve.
That real-world task distribution shapes the research, benchmarks like RF100-VL, and data we bring to model builders.
More on Roboflow Labs: labs.roboflow.com/
NVIDIA introduced the coalition in May 2026 to bring model builders, developers, and organizations building physical AI together around open world models.
The effort supports collaboration on models, research, and evaluation methods to advance systems that can understand and simulate the physical world.
More info: roboflow.com/blog/roboflow-j…
Today we’re announcing that we’ve partnered with NVIDIA and joined the Cosmos Coalition, a global initiative among leading AI labs, as a premiere provider of multi-modal data and post-training capabilities that advance frontier world model research.
We believe the next wave of the AI boom will be defined by what intelligence can do in the physical economy, and that vision is the bottleneck to unblock.
GPT-6.1 Terra beats Astra in our vision benchmarks
wtf OpenAI?
had to double-check the results
- text
- object detection
- counting
also ~47% faster and ~92% cheaper
↓ leaderboards and examples
GPT-6.1 Terra: delivering fast and capable computer use at a lower price.
GPT-6.1 Terra understands what’s on screen, takes action across apps, writes naturally, and is a more capable model for your whole workday with the efficiency to take on more of it.
We are hosting the Visual Intelligence Summit on October 22 in SF.
The physical world is AI's biggest opportunity and the best way to advance humanity in our lifetimes. Few people are building for AI beyond the screen.
The Summit is for the people who see it differently and want to be at the frontier of bringing AI into the physical world.
If you’re building for that future, join us.
Ever wanted to know if you're cooked in an IRL game of chess? 💀
So I built Stockfish IRL. A camera watches a real board, tracks every move, and literally yells at you when you blunder.
Built with @roboflow 👇