Gemini 4 Carbon

Who Should Read

A backend developer building coding agents or security automation on the Gemini API who wants to know the next model's performance, price, and release timing

Core Mechanics

• Carbon is still limited access
  • This week it goes to a subset of cyber defenders first.
  • Access then widens to trusted partners through the Fairwind Program.
  • Developers, businesses, and consumers get it "as soon as possible", with no date given.
• Priced the same as Argon at $2/$10
  • The introductory price is $2 per million input tokens and $10 per million output tokens.
  • Cached input tokens are 95% off the input price.
  • After the introductory period, $4 input and $20 output apply.
• Claims #1 on security benchmarks
  • SEC-bench Pro (finding vulnerabilities) is 88.2%, ExploitGym (exploitation) 54.8%, CWE-bench (patching) 71.0%.
  • The external benchmarks were evaluated without guardrails, per the post.
  • On internal benchmarks, secure code generation success is more than 2x Argon.
• Pitched for multi-day long-horizon work
  • Internal teams ran Carbon for days to root-cause and patch subtle bugs in large codebases.
  • It gains +89 elo on WebDev Arena for #4 (1768).
  • It ranks #1 on Artificial Analysis Intelligence Index v5.0 with 59.
• Fewer shortcuts and fabrications
  • Successful cheating on Terminal-Bench 4.0 is 0.0%, with cheating attempts at 0.4%.
  • High-severity misalignment alerts in internal live coding dropped almost 80%.
  • Unsupported claims are about 2x lower than Argon.
• Less overreach and under-service
  • Overreach is down 26% and under-service down 29%.
  • With ambiguity, it reportedly retrieves sources and surfaces discrepancies to the user before acting.
  • Attack success on Gray Swan's Indirect Prompt Injection benchmark is 0.0% (Argon: 0.7%).

Community Reactions

• Same price but stronger drew praise
  • Some commenters called $2/$10 an incredible deal even as introductory pricing.
  • Others wondered how token-efficient it is.
• Skepticism about 0.0% prompt injection
  • One commenter argued attackers aren't limited to what Google evaluated, so Carbon will be injectable like Argon.
  • The criticism was that AI shops are blind to anything not in a metric.
• Cheating reduction seen as most important
  • One commenter said reducing shortcuts and loophole exploitation is the key to making frontier models usable.
  • They said it was previous Gemini's biggest problem.
• Benchmarks without a release are of little value
  • One commenter said if release comes in a month, stronger competitors will make the numbers look worse.
  • Another complained that a vague "coming soon" leaves people feeling excluded.
• Positive experience with Gemini in customer service agents
  • One user said Gemini's conversation quality since 3 made their customer-service agent harness very well received by users.
  • They also complained that the official harness hurts the model's reputation.
• Many found the naming confusing
  • Commenters said the post doesn't clarify whether Carbon is Argon's successor, a separate model, or a post-train.

How to Apply

• Add a cheating check to your current agent
  • Pull 20 cases from today's coding agent logs where it skipped tests or faked results.
  • Rerun the same tasks with the prompt below and count cheating attempts.
• Add the price change to your cost sheet
  • Compute monthly input and output tokens at both $2/$10 and $4/$20.
  • Apply the 95% cached-input discount and compare call sites with heavy repeated prompts first.
• Defer adopting Carbon until after release
  • There's no general access yet, so plan to compare it against Argon on your own eval set once it's public.
  • For security work, verify with your own codebase's vulnerability cases rather than benchmark numbers like SEC-bench.
• Don't rely only on the model for injection defense
  • For agents reading external documents or tool responses, test with your own attack cases regardless of vendor numbers.
  • State in the system prompt that ambiguous instructions must be asked back, not guessed.

Try It

# Paste into Claude, ChatGPT, or Claude Code (simplified version of the post's idea: fewer shortcuts and fabrications)

You are a coding agent. Follow these rules strictly.
1. Never bypass tests, fake results, or weaken test code.
2. Don't make unsupported claims. Write "not verified" for anything you couldn't confirm.
3. If the instructions are ambiguous or sources conflict, don't guess: find the relevant sources, summarize the discrepancies, and ask me first.
4. Don't touch files or settings outside the request. If needed, only propose them with reasons.
5. For blocked work, report a list of what remains required; don't pretend to fill the gap.

Task: {{task description}}
Constraints: {{project constraints}}

When done, report: (a) what you did, (b) what you couldn't verify, (c) what you deliberately did not do, (d) methodological caveats.

Next step: run the same task with and without these rules and compare counts of shortcuts and unsupported claims.

Terminology

Related Papers

• Dayjob: Tax – evaluating AI agents on 50 professional tax assignments

  On 50 professional tax tasks, the best AI agent passed only 10.8%

• Anthropic S-1

  Anthropic filed an S-1 to list Class A shares on Nasdaq as 'ANT'.

• A 10-year-old CNN outperforms frontier LLMs for skin cancer diagnosis

  A fine-tuned Inception v3 beat GPT-6 Astra and Claude Fable 5.1 on 200 skin lesions

• America has lost 200k of jobs in media, film and the arts. How much is AI?

  US creative-industry jobs fell by over 200k in four years, but AI's share of the blame is hard to prove

• No signs of AI payoff outside the tech sector

  Revenue per worker and margins rise at the Mag 7 but stay flat across the S&P 493

• LLMs are making formal software verification feasible at a much larger scale

  Claude proved no-crash properties for 123 open-source libraries and found bugs along the way.

• PHP2Shell: AI uncovers critical RCE in PHP

  Wiz used AI to turn a decade-old PHP core flaw into a working RCE exploit

• Using LLM for confidence scoring is a bad idea

  An LLM's self-reported confidence is plausible text, not a measurement; build scores from outside rules.

Related Resources

• https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/
• https://i.imgur.com/IDFGuJU.jpeg