Astra is the best VLM I've tested. it handles open-vocabulary object detection easily.
Astra is also very expensive. processing one video frame costs about $0.08. naively running every frame of a basketball game through Astra would cost almost $7,000.
nitter.cf/skalskip92/status/2096…
prompt GPT-6 Astra with boxes to generate new detections
really useful when you know the object, but can’t describe it well in words (like shipping container wall dent)
best part is you can do this between different images
- green - positive examples
- red - negative examples
- blue - new detections
↓ GPT-6 Astra box prompting
updated vision model tier list with GPT-6 Astra and Gemini 3.8 Flash
so maybe a good world model / vision model / language model converge to just “a good model”
Astra can do segmentation
this is pure VLM result. no expert models (like SAM) were used
No other VLM even comes close to this quality
- high effort
- avg input tokens / image: 2,052
- avg output tokens / image: 4,685
- avg cost / image: $0.255
- median time / image: 78.3 s
↓ GPT-6 Astra segmentation deep dive
data labeling is dead
GPT-6 Astra easily tells Celtics players from Knicks players
whether they play at home or on the road
GPT-6 Astra is now the top model for vision tasks
sure 3-10x more expensive but 50%+ faster
extremely capable model, OpenAI really pushed the frontier in all directions with this one
This is the kind of vision benchmark we need more of.
Actually testing models on real world tasks instead of just throwing another leaderboard score at us.
Great work by @skalskip92 .
Google has not a single moat left, huh
long context is dominated by OAI/Anthropic and many Chinese models handle it fine and for cheaper. Vision – Anthropic is catching up, Qwen is on par, Astra ahead. Core LLM… lol. Updating 3.x Flash won't cut it guys
Ciekawy wątek, warto rzucić okiem.
Wygląda na to, że Astra w vision właśnie przebija Gemini: zauważa detale, precyzyjne boxy i ogólnie lepiej ogarnia obraz.
อ่าวพี่เจแพ้ ในฟิลด์ที่ถนัดที่สุดซะแล้ว
พี่เจ สู้ๆนะ 😆
Google playing to their strengths!
run Gemini 3.8 Flash and GPT-6 Astra side-by-side
link: playground.roboflow.com/shar…
GPT-6 Astra is so good, the only detection it got "wrong" on this eval is a die that never got labeled.
Vision works remarkably well. This makes Astra a lot more useful for tasks interfacing with the real world
jumps like the one below are a huge deal. i recently worked with cancer drug researchers who sweated over even a few percentage point improvements in cell counting using image processing
So Google lost its only advantage as well