Wednesday, October 28, 2026
Tech Beat
Oct 28, 2026, 6:47 PMArtificial Intelligence

Google Unveils Gemini 4 Carbon but Delays Access for Safety Tests

Google unveils Gemini 4 Carbon with coding and cybersecurity gains, but delays access for US testing and leaves wider release and pricing dates unclear.

Listen to this briefingAudio briefing

Summary

Google announced Gemini 4 Carbon on October 28, 2026, two weeks after rolling out Gemini 4 Argon, but the staged release remains unavailable and is not the anticipated Flash counterpart. Google claims gains in coding, agentic tasks, scientific analysis, and cybersecurity vulnerability detection, validation, and remediation. Carbon is twice as likely as Argon to generate vulnerability free code, while Chrome rates its high quality patches 10 times better than those from non-Google models. An 89 point WebDev Arena increase puts Carbon at No. 4.

Google’s benchmarks give Carbon slim leads over Argon and place it above GPT-6 Astra, Claude Opus 5.5, and Claude Fable 5.1 in almost every test, although frontier model scores remain close. In Grey Swan testing, Carbon allowed zero percent of prompt injection attacks, against 0.7 percent for Argon, 1 percent for Opus 5.5, and 8.5 percent for GPT-6 Astra. Google will complete testing under the US voluntary framework before deploying Carbon through Fairwinds, with no broad release date. AI Ultra subscribers and Gemini API developers will probably receive access first at discounted rates of $2 per million input tokens and $10 per million output tokens. Both rates will later double, but Google has not set an expiration date, while Argon’s discount ends in November.

Positives

  • Carbon allowed zero percent of prompt injection attacks in Grey Swan testing, outperforming Argon, Opus 5.5, and GPT-6 Astra.
  • Carbon is twice as likely as Argon to generate code free of vulnerabilities, according to Google.
  • Chrome rates Carbon’s high quality patches 10 times better than patches from non-Google models.
  • An 89 point WebDev Arena increase lifted Carbon to fourth place.
  • Google’s benchmarks place Carbon narrowly ahead of Argon and major OpenAI and Anthropic models in almost every test.

Risks & concerns

  • Gemini 4 Carbon remains unavailable while Google conducts staged testing under the US voluntary framework.
  • Google has provided no date for broad access or deployment through its Fairwinds cyber defense program.
  • Carbon’s benchmark leads are only a few points, with competing frontier models scoring in the same range.
  • Discounted Gemini API prices will double, but Google has not disclosed when the introductory period ends.
  • Carbon is an Argon improvement rather than the lower cost Flash counterpart some users were awaiting.
Primary sourceAI - Ars Technicahttps://arstechnica.com/ai/2026/10/google-announces-gemini-4-carbon-its-latest-frontier-ai-model-that-you-cant-use/
Read full article
Editorial note: Tech Beat summarizes and analyzes third-party reporting. The source link is the authoritative article. This page does not reproduce the full source text.

More From The Wire

Artificial IntelligenceOct 28

Anthropic Launches ROI Tools as Claude Code Spending Jumps 10x

Artificial IntelligenceOct 28

Google AI Mode Dodges 2020 Election Winner in US Search Tests

Artificial IntelligenceOct 28

Google Gemini 4 Carbon Leads Cybersecurity Benchmarks With Long Horizon AI