Gemini Carbon roundup: coding and long-horizon agent focus, 2–4x slower than Argon
A roundup of rumors about Google's internal Gemini Carbon checkpoint: possibly an Argon upgrade with big SWE-bench and DeepSWE gains over Barium-B, compared by an employee to Opus 5.5, in wider internal A/B tests and possibly on Arena, but 2–4x slower than Argon.
Every sighting and finding about Google’s Gemini Carbon. Carbon is Google's rumored internal model checkpoint focused on complex coding and long-horizon agentic tasks. What's new: - A potential upgrade to Argon, complex multi-step tasks, and coding performance while strengthening its position ahead of Fable 5.5. - Internal reports suggest major SWE-bench and DeepSWE improvements over earlier Barium-B checkpoints, with one employee comparing it to “Opus 5.5.” - Wider internal A/B testing is reportedly underway, with possible sightings on LMSYS Arena. - Carbon appears 2 to 4 times slower than Argon, suggesting heavier reasoning. - Its release path remains unclear. It could replace Argon, become a higher-performance tier, or serve as a stronger final checkpoint. - Despite Flash release rumors, there's no solid evidence linking Carbon to an imminent Gemini 4 Flash launch.