Gemini 4 Carbon is now available
Find out how Gemini 4 Carbon performs compared to competing models and what that means for your AI projects.
Gemini 4 Carbon is now available through the API and immediately hits the same benchmark score as Anthropic’s Fable 5.2. The user who ran the test noted the parity in raw performance, but also observed that Carbon is slower and more expensive per token. That contrast is the most striking part of the release, especially given Google’s reputation for rapid model rollouts.
The benchmark comparison comes from a single user’s private test suite, which measures latency and cost per generated token. While the score aligns with Fable 5.2, the underlying architecture appears to prioritize output quality over speed, leading to higher compute usage. The user also mentioned that earlier Gemini models (3.1 Pro, 3.6 Flash, 3.6 Thinking) were still in use, suggesting Carbon is a step up in capability but not in efficiency.
Beyond raw numbers, the user found Carbon’s responses surprisingly insightful on a draft paper, surfacing edit suggestions that other top models missed. The model’s ability to generate nuanced feedback indicates a stronger reasoning layer, even if it takes longer to produce results. This aligns with the observation that Gemini’s recent releases have focused on “pelican” style solid outputs rather than raw throughput.
If you’re building software that relies on fast, cheap inference, you may need to stick with faster, cheaper alternatives like Fable 5.2 for now. However, for tasks where depth of analysis matters, such as drafting academic text or detailed code review, Carbon’s richer output could justify the extra cost. Choose the model based on whether latency or insight is the primary driver for your application.
TakeawayUse Gemini 4 Carbon for quality‑focused tasks, but expect higher latency and cost than comparable models.