CORe 6.2 is a major step up from CORe 6.1. It keeps everything that made the previous flagship feel grounded. Transparent reasoning, honest answers, careful refusals, and adds three things that fundamentally change what you can build with it.
It understands video
You can now pass a video_url straight into a chat completion and CORe 6.2 will actually watch it. Frame sampling, scene transitions, on-screen text, motion, all of it. Drop a short clip, ask “what happens here?”, and you get an answer that's grounded in the actual pixels, not a guess based on the filename.
Public https:// URLs are fetched server-side, so request bodies stay small. Supported formats include MP4, WebM, MOV, M4V, MPEG, and AVI; up to 80 MB per file and 12 videos per request.
It scores meaningfully higher on the things you can measure
- 92.4 MMLU ; competitive-frontier territory for general knowledge and multi-step reasoning.
- 99.0 HarmBench ; at the top of what's been publicly reported for adversarial-prompt robustness.
It has a personality
This is the part that's harder to put in a benchmark. CORe 6.2 maintains a recognizable voice across coding questions, casual chat, emotional topics, and adversarial prompts. While it'll match your tone, it doesn't dissolve into corporate-assistant blandness when a topic gets serious (most of the time). Users who've spent a few hours with it tell us it feels like one assistant they're talking to, not a different one every session.
What's the same
Everything you already use. The OpenAI-compatible API, the 262K token context window, tool calling, streaming, JSON mode, the COReAlign safety stack, all of it works exactly like it did with CORe 6.1. Existing integrations don't need a single line changed. Just bump the model field to core-6.2.
Try it
CORe 6.2 is the new default model in the chat app and on the API. If you'd rather stay on 6.1 for now, it's still available under core-6.1.