Gemini 4 Carbon checkpoint: leaks, sightings and internal evals
A roundup of Gemini 4 Carbon rumors: an internal experimental branch alongside Argon focused on complex coding and long-running agentic tasks, with reported SWE-bench gains.
Gemini 4 Carbon Checkpoint: Leaks, Sightings and New context. Since my previous post about the Gemini 4 Carbon discovery, more reported details and community sightings have surfaced. Here’s an updated overview of what’s been circulating, alongside my own interpretation. TL;DR: Carbon is reportedly Google’s internal experimental branch being developed alongside the Argon checkpoint, with a strong focus on complex coding, long-running tasks, and agentic workflows. What’s new about this unreleased internal model: Internal evaluations reportedly show substantial gains on SWE-bench and Deep SWE compared with earlier Barium-B checkpoints. One employee described it as feeling like “Opus 5.5,” while some employees had acknowledged that Gemini Argon still fell short in certain areas. Reports suggest that internal A/B testing has expanded across testing environments, primarily AGY, though other environments may also be involved. Carbon’s eventual release format remains unclear. It could replace Argon directly, become a separate performance tier within the Gemini 4 lineup, or serve as a final refinement of Argon that addresses its remaining weaknesses - including concerns from users who consider it “benchmaxxed.” Carbon appears to belong to Gemini 4’s ongoing checkpoint development process, which uses element-based names such as Argon, Barium, and Carbon. Internal documents identified Barium-B as the build chosen for Argon’s public release. That leaves open the possibility that Carbon could become an upgraded checkpoint within Gemini 4.x, potentially serving as a polished, stable final version of Gemini 4 Argon. That’s my interpretation, though, rather than a confirmed plan. Some community members expect Gemini 4 Flash to launch soon, possibly alongside Carbon. There have been hints of a Flash release, but, from what I’ve seen, there isn’t enough evidence to establish either that timing or a connection to Carbon. Carbon reportedly runs two to four times slower than Argon, which could suggest a heavier reasoning model. The latest checkpoint attracted attention a few days ago after apparently strong demonstrations in Arena. Accounts describe the preceding checkpoint as poor, making the latest version’s reported jump in capability especially striking. Perhaps the most notable development: community members report encountering Carbon in LMSYS Arena. If those sightings are accurate, Google may be following a testing approach similar to Argon’s, where a lucky prompt could give someone a chance to interact with the unreleased model. My personal take is that Google may be positioning Carbon to compete for the top spot ahead of Fable 5.5. Another possibility is that a backup checkpoint developed into a much stronger candidate than expected following Argon’s release and successive improvements. If those reported gains hold up, Carbon could address several weaknesses in Google’s current flagship, Gemini 4 Argon: greater reliability, stronger performance on long-running tasks involving multiple steps, and improvements in important benchmark areas where Argon trails. It might also deliver more cost-efficient task completion, although that remains speculative. What do you think? will Carbon appear first as an experimental model in AI Studio, or arrive through another phased rollout? If you’ve encountered it or experimented with it and feel comfortable sharing, drop your observations below.