ModelsArchitectures & capability
Google makes Gemini 4 Carbon generally available at half price until November 30, and llm-gemini 0.36 adds support
Google opened Gemini 4 Carbon to paying API developers and subscribers on November 10, two weeks after limiting it to cyber defenders. Introductory pricing of $2/$10 per million tokens runs through November 30, and Simon Willison's llm-gemini 0.36 already supports it.
From the Simon Willison article of November 10, 2026 and cited sources: From October 28, Carbon was limited to selected cyber defenders in Google's Fairwind Program. [1] [3] [4] [6]
SOURCES[1] Simon Willison: llm-gemini 0.36 simonwillison.net[3] Google: Gemini 4 Carbon is now available blog.google[4] Google: Gemini 4 Carbon: our next frontier model blog.google
AT Pulse research deskPublished · Updated · 4 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- What happened
- From October 28, Carbon was limited to selected cyber defenders in Google's Fairwind Program. [1] [3] [4] [6] [7]
- Original source
- Simon Willison article, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- November 10, 2026 article published
- Covered
- Nov 11 edition
- Evidence
- 9 sources across 8 sites
- 1Simon Willison: llm-gemini 0.36simonwillison.net
- 2Linked video: llm-gemini 0.36youtube.com
- 3Google: Gemini 4 Carbon is now availableblog.google
- 4Google: Gemini 4 Carbon: our next frontier modelblog.google
- 5Artificial Analysis: Gemini 4 Carbon new #1artificialanalysis.ai
- 6GitHub: llm-gemini 0.36 releasegithub.com
RisingHow it’s measured
YouTube video · 2.5K views · 37 likes
Reach blends public engagement (views, likes, comments, shares, stars) from the platforms where the story or its original source lives: YouTube, Hacker News, GitHub, Substack and similar. A video and the article written from its transcript share one score, taken from the video. Platforms with no data are left out rather than counted as zero, and the age of a story never lowers its score.
From October 28, Carbon was limited to selected cyber defenders in Google's Fairwind Program. On November 10 it opened to paying API developers, Gemini for Business, AI Ultra subscribers, Antigravity and Android Studio. A cyber-specialized version still requires Fairwind approval. On the same day, llm-gemini 0.36 added the gemini-4-carbon model with low, medium and high thinking levels. Medium is the default. [1] [3] [4] [6] [7]
Introductory rates of $2 per million input tokens and $10 per million output tokens end November 30. They then double to $4/$20, the same as Opus 5.5 and half of Fable 5.2's $8/$40. Carbon also uses many tokens. Artificial Analysis measured about 123k output tokens and $7.76 per index task at list price on the high setting. That makes cost per task a better budgeting measure than per-token price. [1] [3] [5] [7]
Rankings conflict. Artificial Analysis scored Carbon 59 and placed it first on October 28, but VentureBeat's November 10 figures show Fable 5.2 at 65 and Carbon at 60. Coding results are also split: Carbon leads on Terminal-Bench 4.0 (65.2% vs 61.6%) but trails on FrontierCode 2.0 (40.6% vs 49.1%). [5] [7]
Per-token rates double on December 1, and independent testing found heavy output-token use. Teams have a short window to evaluate Carbon cheaply. They should budget by cost per task rather than by headline token price.
Executive brief
Google's Gemini 4 Carbon is now generally available at half its list price, and that discount has an end date. Google says introductory pricing of $2 per million input tokens and $10 per million output tokens applies through November 30, after which it rises to $4/$20 (Google). For two weeks the model was limited to vetted cyber defenders. Simon Willison's llm-gemini 0.36 plugin added it the same day it opened up (Simon Willison). Carbon took first place on Artificial Analysis's index in October, but a later comparison puts Claude Fable 5.2 ahead of it.
What changed and event timeline
Carbon announced with a staged rollout
Google limited access to "a subset of cyber defenders" through its Fairwind Program. It cited the U.S. government's voluntary process for testing models before release ().
First on Artificial Analysis index
Carbon scored 59 on Intelligence Index v5, one point ahead of Claude Opus 5.5. It ranked second on the Coding Agent Index ().
General availability
The model opened to all paying API developers, Gemini for Business, AI Ultra subscribers (Gemini app and AI Mode), Antigravity and Android Studio ().
Scheduled price doubling
Rates rise to $4/$20, the same as Opus 5.5 (;).
Capabilities and access
- Model: Gemini 4 Carbon (
gemini-4-carbonin llm-gemini 0.36). - Thinking levels: low, medium and high, with medium as the default (Simon Willison).
- Context and inputs: 1M-token context; text, image, video and speech input (Artificial Analysis).
Read the full section
- Model: Gemini 4 Carbon (
gemini-4-carbonin llm-gemini 0.36). - Thinking levels: low, medium and high, with medium as the default (Simon Willison).
- Context and inputs: 1M-token context; text, image, video and speech input (Artificial Analysis).
- Cached input: 95% discount (Google).
- Cyber-specialized version: still requires Fairwind approval (VentureBeat).
Technical analysis for researchers and developers
- Architecture: Carbon builds on Gemini 4 Argon, described as a mixture-of-experts model trained with reinforcement learning.
- Token use: Google says Carbon spends more tokens than Argon on exploration and verification loops (Google).
- Methodology: Most of the Artificial Analysis index uses private test sets (The Batch). That limits reproducibility but reduces contamination risk.
Read the full section
- Architecture: Carbon builds on Gemini 4 Argon, described as a mixture-of-experts model trained with reinforcement learning. Parameter count, training data and knowledge cutoff are undisclosed (The Batch).
- Token use: Google says Carbon spends more tokens than Argon on exploration and verification loops (Google). Artificial Analysis measured 123k output tokens and $7.76 per index task at the $4/$20 list price, using the "high" setting (Artificial Analysis). Budget on that basis rather than per-token price alone.
- Methodology: Most of the Artificial Analysis index uses private test sets (The Batch). That limits reproducibility but reduces contamination risk.
Claims and evidence
- Cyber benchmarks: 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on CWE-bench. These are Google's figures (Google).
- Leaderboard claims: Google says Carbon is #1 on CursorBench 4.0, WANDR and AutomationBench (Google).
- Independent ranking: Artificial Analysis scored it 59 and ranked it first (Artificial Analysis).
Read the full section
- Cyber benchmarks: 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on CWE-bench. These are Google's figures (Google).
- Leaderboard claims: Google says Carbon is #1 on CursorBench 4.0, WANDR and AutomationBench (Google). VentureBeat found the public CursorBench and WANDR tables lacked Carbon data or were incomplete, leaving those claims unverified. Zapier's AutomationBench does show Carbon ahead: 53.88% versus 51.45% for Fable (VentureBeat).
- Independent ranking: Artificial Analysis scored it 59 and ranked it first (Artificial Analysis).
Context and prior work
Read the full section
- Release pace: Carbon followed Argon by about two weeks (Google).
- Cyber-first gating: Fairwind works with more than 650 partners (Unite.AI).
- Willison's tests: He found Gemini's SVG pelicans already strong, so he also tried a Three.js "Yonkers mech" prompt in Antigravity. It needed a follow-up prompt to produce a standalone HTML file (Simon Willison).
- Google's showcase: Google's pelican-on-a-GBike game was built with Antigravity's /swarm subagents (AI Studio). The linked video's auto-captions contain only music markers (00:00–01:05).
Limitations, safety and contested findings
- Conflicting rankings: Artificial Analysis gave Carbon 59 and first place on October 28.
- Coding results are split: Carbon leads on Terminal-Bench 4.0 (65.2% vs 61.6%) but trails Fable on FrontierCode 2.0 (40.6% vs 49.1%) (VentureBeat).
- Safety figures are Google's own: 0% prompt-injection success on Gray Swan, 99% of simulated cyberattacks blocked, and a 92% agentic-honesty pass rate versus 70% for Argon (Google).
Read the full section
- Conflicting rankings: Artificial Analysis gave Carbon 59 and first place on October 28. VentureBeat's November 10 figures show Fable 5.2 at 65 and Carbon at 60 (VentureBeat).
- Coding results are split: Carbon leads on Terminal-Bench 4.0 (65.2% vs 61.6%) but trails Fable on FrontierCode 2.0 (40.6% vs 49.1%) (VentureBeat).
- Safety figures are Google's own: 0% prompt-injection success on Gray Swan, 99% of simulated cyberattacks blocked, and a 92% agentic-honesty pass rate versus 70% for Argon (Google).
- Hallucination: 11% rate on AA-Omniscience, second-lowest among models tested (Artificial Analysis).
Business and practitioner implications
- Pricing window: Run evaluations before December 1.
- Cost per task: Track cost per task, not per token. Heavy reasoning drives the bill.
- Fast testing: CLI users can compare thinking levels quickly with llm-gemini 0.36.
Read the full section
- Pricing window: Run evaluations before December 1. After that, Carbon costs the same per token as Opus 5.5 and half of Fable 5.2's $8/$40 (VentureBeat).
- Cost per task: Track cost per task, not per token. Heavy reasoning drives the bill.
- Fast testing: CLI users can compare thinking levels quickly with llm-gemini 0.36.
- Security teams: The cyber-specialized version still requires Fairwind approval.
The source trail.
Sources (9)
Simon Willison: llm-gemini 0.36
simonwillison.netLinked video: llm-gemini 0.36
linked video auto captions
www.youtube.com