# Gemini Carbon Opens Up Ahead of a December Price Rise

- Date: 11 Nov 2026 (2026-11-11T17:05:34.000Z)
- Summary: Google widened access to Gemini 4 Carbon, whose introductory token prices double on December 1. Theo’s hands-on Fable 5.2 tests highlight false findings and slow code review despite strong published benchmarks.
- Tags: `digest`, `ai-discourse`, `gemini`, `claude`, `coding-agents`, `model-evaluation`, `pricing`

## Sources

1. [Simon Willison - Release: llm-gemini 0.36](https://simonwillison.net/2026/Nov/10/llm-gemini/) (website)
2. [Theo / t3.gg - Nobody Is Competing With Anthropic](https://www.youtube.com/watch?v=yh5IAUow4Hc) (youtube)

## Executive Summary

- **Google widened access to Gemini 4 Carbon on November 10.** Its introductory token prices double on December 1, so builders should budget trials against the higher rates.
- **Theo’s hands-on Fable 5.2 tests show uneven coding results despite strong published benchmarks.** Repository analysis produced false findings, and PR review was slow; he still prefers Opus 5.5 for everyday coding.

## Gemini Carbon’s Launch Prices Last Until December 1

Gemini 4 Carbon, initially offered to selected cyber defenders, is now more broadly available, according to [Simon Willison’s release note for llm-gemini 0.36](https://simonwillison.net/2026/Nov/10/llm-gemini/).

Launch pricing is **$2 per million input tokens and $10 per million output tokens**, matching the GPT-6.1 Sol and Claude Sonnet 5.5 rates Willison lists. On December 1, Carbon rises to **$4 input and $20 output**, matching Opus 5.5 in that comparison. A fixed token workload will therefore cost twice as much after the introductory period.

Carbon offers low, medium and high thinking settings, with medium as the default. Willison’s updated plugin adds access through his LLM command-line tool.

He also tried Carbon in Google’s Antigravity IDE, asking for a Three.js animation of Yonkers standing up as a giant robot and walking around New York, using OpenStreetMap data. He published the resulting HTML after a follow-up request to package it as a standalone file. That gives developers a concrete output to inspect alongside the release claims.

## Fable 5.2’s Repository Tests Expose Review Costs

In [“Nobody Is Competing With Anthropic,”](https://www.youtube.com/watch?v=yh5IAUow4Hc) Theo describes testing Fable 5.2 on codebase analysis, PR review and coding work. These are creator-run tests on his own projects, with no controlled comparison.

For TS Rust, his TypeScript-to-Rust project, he counted **four incorrect claims among 61 findings**, at a cost of about **$23**. For T3 Code, he counted six false claims among 83. Reviewing eight PRs took roughly **120 minutes of model runtime**.

He found value in Fable’s reviews, including its willingness to let minor nitpicks slide. But its speed and tendency toward overly clever choices made Opus 5.5 his preferred daily driver. He also highlights published coding results where maximum reasoning effort costs more and scores worse than a lower setting.

Test repository understanding and review quality alongside code generation when choosing a coding model. Track false findings, review time and spend, and compare reasoning settings before defaulting to maximum effort.
