Oh, Twin Carbon opens for everyone - a day after Fable 5.2.
Where to get:
–. API in AI Studio and Gemini for business - for anyone who pays, at an enterprise introductory price: $2 / $10 per million tokens, really cheap.
- AI Ultra subscribers - in anti-gravity, Gemini app and AI mode.
scientists. Free for 6 months (plan for 10,000 people).
What's inside:
– /Swarm in Antigravity: One prompt becomes a team of agents. Carbon is an architect and reviewer, lighter models write code in parallel isolated environments, separate validation agents check each change. Well, we've already seen that. In normal mode, it will just eat tokens.
- #1 in CursorBench 4.0, WANDR and AutomationBench - and more than 2x cheaper per task than the next external model.
– New MKA benchmark (document agents with drawings, graphs and tables): second place, but 6 times cheaper than the leader.
- No Cyber: Wiz found 0-days from two major SaaS providers in less than 2 hours.
- Science: Helped formally prove the properties of the Signaling Protocol code - millions of lines of Lean in weeks instead of years.
How does it compare to its competitors:.
– אינדקס AA כולל: משל 5.2 – 65, פחמן – 60, אופוס 5.5 – 59, GPT-6 אסטרה – 57
–. FrontierCode 2.0: Fable 49.1%, Astra 42.8%, Carbon 41.1%.
- Computer usage, OSWorld 2.1: 89.3% / 79.5% / 81.5%.
–. AutomationBench: Carbon 53.9% compared to 51.4% for Fable and 41.4% for Astra.
Prices: Fable $8 / $40, Astra $10 / $50. That is, Carbon is now 4-5 times cheaper according to the price list.
זה מצחיק שגוגל עצמה לא ממליצה להשתמש בו: עבור רוב משימות הקוד והמסמכים, ארגון נשאר "הבחירה האידיאלית". פחמן מיועד למשימות על הגבול, והוא באמת אוכל יותר אסימונים: הוא חוקר זמן רב יותר, חושב לעומק, מפעיל מחזורי אימות. במדד ה-AA, המשימה שלו מגיעה ל-7.76 דולר לעומת 6.24 דולר של אסטרה, אם כי האסימון זול פי 2.5.
In short, in terms of raw intelligence it is ahead of Fable 5.2, Google gets something like "almost up at half the price".