Gemini 4 Carbon is now available
TL;DR Highlight
Gemini 4 Carbon is a top-tier model that spends more tokens than Argon to reason deeply
Who Should Read
A backend or agent developer choosing which Gemini model to use for complex coding and analysis agents
Core Mechanics
- Carbon sits above Argon
- Argon is still recommended for most software engineering and knowledge work.
- Carbon targets tasks that test model limits, like complex multi-step coding, debugging, and scientific research.
- It uses more tokens than Argon to explore, reason, and run longer verification loops.
- Same price as Argon until Nov 30
- Introductory pricing is $2 per 1M input tokens and $10 per 1M output tokens.
- After Nov 30, 2026 the price becomes $4 input and $20 output.
- All paying developers can use it via the API in Google AI Studio and Gemini for business.
- #1 on three benchmarks at under half the cost
- It ranks #1 on CursorBench 4.0, Perplexity's WANDR, and Zapier's AutomationBench.
- Cost per task is more than 2X lower than the next best external model.
- /swarm preview arrives in Antigravity
- One prompt becomes a team of agents you can track in real time.
- Assign Carbon to deep-reasoning roles like architecture, coordination, and auditing.
- Pair faster, lighter models with parallel isolated workspaces and adversarial verifiers that test every change.
- #2 on the multimodal knowledge agent benchmark
- It ranks #2 on the MKA benchmark built with Artificial Analysis.
- Average cost per task is more than 6X lower than the top model.
- Claimed strongest cybersecurity performance
- Through the Fairwind Program it leads on SEC-bench Pro, ExploitGym, and CWE-bench v1.
- On Google's internal Project Fortify benchmark it hit 100% scan recall and 95% triage precision.
- On open-source repo testing, recall improved over 25% relative to Argon.
- Safety numbers improved over Argon
- Per-attempt breach rate against adaptive red-teaming was 0.8%, down from 3.1% on Argon.
- It passed 92% of the 375-task agentic honesty benchmark versus 70% for Argon.
- Indirect prompt injection success on the Gray Swan IPI benchmark was 0%, though Google says it remains an area to improve.
Community Reactions
- Many paying users can't access it
- One commenter on a Plus plan sees it listed but gets an Ultra-only unavailable message.
- Another says the only ways in are a $99/month plan, the API, or applying as a scientist.
- Reviews are split between impressed and disappointed
- One commenter said it caught points on a paper draft that other top models missed.
- Another ran their own benchmark and found it matched Fable 5.2 but was slower and costlier.
- One commenter who uses Argon for coding and Opus for review hopes to drop Opus.
- Reports of very slow speed at launch
- One commenter saw a p75 TTFT of 60s and 10 TPS and called it unusable.
- Another worries that higher token use will drain their Pro subscription limits faster.
- Argon needed /swarm-style workflows
- One commenter said Argon cut corners when one-shotting complex prompts.
- Adversarial review caught hardcoded hacks and bugs in physics code.
- /teamwork helped but made it take even longer.
How to Apply
- Compare Argon and Carbon on your own prompts today
- Run 5 real work prompts through both models in AI Studio.
- Record answer quality, token usage, and latency in a table.
- Split roles between Carbon and Argon
- Use Carbon only for deep-reasoning steps like architecture, coordination, and auditing.
- Run the other implementation steps on Argon or a lighter model to save tokens.
- Re-estimate costs around the Nov 30 price change
- Recompute monthly cost using the post-Nov 30 rates of $4 input and $20 output.
- Factor in that Carbon uses more tokens than Argon when comparing cost per task.
- Add a verification step to your agent pipeline
- Separate the model that writes code from one that adversarially checks the changes.
- Keep a dedicated review prompt that hunts for hardcoded fixes and bugs.
Try It
# Paste into the Gemini API (AI Studio) or Antigravity — a simplified version of the article's architect/implement/adversarial-verify role split
You will perform 3 roles in order.
1) Architect: read the requirements below, then produce a design and split the work into units.
2) Implementer: write code for each unit.
3) Adversarial verifier: list hardcoded solutions, missing error handling, and bugs in the implementation.
Requirements: {{requirements}}
Codebase context: {{code/files}}
Output each stage separately, fix the issues the verifier finds, then check again.
Next step: run the same prompt on Argon and Carbon and compare the number of issues the verifier finds and the token usage.Terminology
Related Papers
Gemini 4 Neon
Gemini 4 Neon cuts cost per task 32% versus Carbon, at $3 input and $15 output per million tokens
AI is killing All About Berlin
After Google AI Overviews, a Berlin guide site lost 70-75% of its traffic.
Not Everything Needs an LLM
PostHog used small decision models to skip 98.7% of expensive LLM calls
GPT‑6.2 Astra and Sol
GPT-6.2 Astra and Sol cut costs sharply, but over-engineering and early-stop reports are mixed
Show HN: TrnRT – An open-source LLM inference engine for AWS Trainium
One trn2.48xlarge serves GLM-5.3-Flash through an OpenAI-compatible API using NKI megakernels.
Know When to Zoom: How Small Is Too Small for Vision-Language Models?
Splitting large photos into tiles raised Claude and Gemini accuracy by 11.1–14.5 points
Understanding Positional Embeddings in Modern LLMs
A walkthrough of how LLMs encode token position, from GPT-2 to RoPE
Show HN: DeepSeek V3 in ~200 lines of pure Jax
DeepSeek V3's core design in 190 lines of JAX, trained on a MacBook Air in 2.6 hours