Ever since several people from the Gemini team posted "Gemini" on social media during the day today, I had a feeling it would be today.
Recently, X, OpenAI, and Anthropic have all been rolling out new releases one after another. Google really couldn't hold back anymore and rushed out to draw a pie in the sky.
We've been chewing on that pie for over half a month. On the data, it punches Astra 6 and kicks Fable 5. How it actually turns out over the next few days will soon be clear.
The official post opens by emphasizing Argon's response style: less fluff, more direct answers, no more beating around the bush or piling on boilerplate. So far, Flash 3.8 has received quite a lot of praise for text editing, and I hope Argon can carry that forward.
Second is coding, with the capability focus on complex code engineering and Agent tasks. The Chrome team used it to refactor 6 million lines of legacy code, compressing a months-long schedule into a few weeks.
The changes in product form are even more worth mentioning. Gemini web has been updated, the previous Spark entry point has been removed, and the chat box can now switch between "Chat" and "Task" modes.
The Gemini app is also the same as Web, and even goes further by directly building Antigravity into the client. I actually tested it: after uninstalling Antigravity 2.0, Antigravity in Gemini still worked normally, so it really is built in.
When Antigravity 2.0 is still present locally, Antigravity in Gemini will directly read the data and configuration of the standalone Antigravity 2.0, including projects, conversations, Agents, Skills, etc.
However, the built-in Antigravity cannot use the two third-party models Claude and GPT. I don't know whether it hasn't been added yet, or whether it will just be cut off like X.
Currently, the built-in version is only supported by the macOS client. Windows? Forget it—that's ages away.
The new model is already available in the API, Gemini app, and Antigravity, but the app and Antigravity are currently limited to Ultra users.
On pricing, $2 per million input tokens and $10 per million output tokens. This price is on par with Claude Sonnet 5.5. As long as its actual performance beats Opus 5.5, Google will have won this round.