API ManyНовые модели: Qwen 3.8 Max, Claude Haiku 5.5, Claude Opus 5.5, Claude Sonnet 5.5
GeminiClaudeComparison

Gemini 4 Carbon vs Claude Opus 5.5: Coding Agents, Security and Access

Carbon edges Opus 5.5 on the AA Intelligence Index (59 vs 58), but Opus 5.5 leads the Coding Agent Index 62 to 59. Long tasks, security results, cost and access — Carbon has no public API yet.

29 окт. 2026 г. · API Many

On the headline index Carbon is one point ahead. For agentic coding, the job most teams would actually use either model for, Opus 5.5 is still ahead. This page looks at that split, at long-running work and security, and at how far Carbon has moved from Gemini 4 Argon.

Head-to-head

Artificial AnalysisGemini 4 Carbon (high)Claude Opus 5.5 (max)
Intelligence Index v55958
Coding Agent Index59 (with Antigravity CLI)62 (with Claude Code)
Output tokens per Index task123k173k
Cost per Index task (standard pricing)$7.76$10.04
AA-Omniscience v1.15448
Hallucination rate11%50%
Public APINoYes

Sources:Artificial Analysis: Gemini 4 Carbon is the new #1 (Oct 28, 2026) · Artificial Analysis Intelligence Index leaderboard

Coding agents: Opus 5.5 still leads

The Artificial Analysis Coding Agent Index scores a model together with its agent harness. Gemini 4 Carbon (high) in Antigravity CLI scores 59, three points behind Claude Code with Claude Opus 5.5 (max) at 62. Carbon ranks second on Terminal-Bench 4.0 at 64%. Artificial Analysis calls the improvement notable because Google has historically underperformed on coding agents.

Source:Artificial Analysis: Gemini 4 Carbon is the new #1 (Oct 28, 2026)

We leave FrontierCode out of this comparison: the Carbon article quotes FrontierCode v1.1, while Intelligence Index v5 uses FrontierCode 2.0, so the two sets of numbers are not comparable.

Long tasks

Google says Carbon is designed for long-horizon engineering and that its teams have seen it run for days, persisting through dead ends on multi-step tasks such as root-causing and patching subtle bugs across large codebases. That is Google’s internal experience, not a benchmark.

Source:Google: Gemini 4 Carbon announcement (Oct 28, 2026)

What can be measured is token use: on the Intelligence Index, Carbon used 123k output tokens per task against 173k for Opus 5.5 (max), which is why it costs about 23% less per task ($7.76 vs $10.04) at standard prices.

Source:Artificial Analysis: Gemini 4 Carbon is the new #1 (Oct 28, 2026)

Cybersecurity focus

Security is the reason for Carbon’s gated rollout. Google reports 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on Collinear’s CWE-bench, all evaluated without guardrails, and calls them state of the art. Google’s announcement gives no Opus 5.5 results on these benchmarks, and the Artificial Analysis pages cited here do not cover them, so we cannot put the two models side by side here.

Source:Google: Gemini 4 Carbon announcement (Oct 28, 2026)

Because Google is deliberately holding Carbon back while it tunes guardrails, security teams are the first users, through the Fairwind Program, which is open by application to high-priority defenders such as governments, healthcare providers and telecommunications services.

Sources:Google: Gemini 4 Carbon announcement (Oct 28, 2026) · Google DeepMind: Fairwind Program

Availability

  • Claude Opus 5.5: available now, including on API Many.
  • Gemini 4 Carbon: no public API, no model ID, no date for developers. Artificial Analysis lists it as “No publicly available API”.
  • Price when it opens: $2 / $10 per 1M input / output tokens introductory, then $4 / $20 per Google; Artificial Analysis says the 50% launch discount has no confirmed end date.

Sources:Artificial Analysis Intelligence Index leaderboard · Google: Gemini 4 Carbon announcement (Oct 28, 2026) · Artificial Analysis: Gemini 4 Carbon is the new #1 (Oct 28, 2026)

Context: what changed from Gemini 4 Argon

Gemini 4 Argon (high)Gemini 4 Carbon (high)Source
Intelligence Index v55059Artificial Analysis
AA-Briefcase (Elo)14831609Artificial Analysis
AA-Omniscience v1.14654Artificial Analysis
Hallucination rate4%11%Artificial Analysis
Cost per Index task (standard pricing)$7.68$7.76Artificial Analysis
Gray Swan IPI attack success (Google-reported)0.7%0.0%Google
Both models list at $4 / $20 per 1M tokens standard pricing. Argon’s hallucination rate is lower than Carbon’s: Carbon’s accuracy rose 10 points to 58%, but its hallucination rate also rose, from 4% to 11%.

If you are on Gemini today, Argon is the model you can actually use, and Carbon is a nine-point jump on the same standard price once it opens. On AA-Briefcase, Carbon is still behind leaders such as Kimi K3.1 Pro (1678 Elo).

Source:Artificial Analysis: Gemini 4 Carbon is the new #1 (Oct 28, 2026)

Related

FAQ

Is Gemini 4 Carbon better than Claude Opus 5.5 for coding?

Not on the Artificial Analysis Coding Agent Index: Claude Code with Opus 5.5 (max) scores 62, Carbon in Antigravity CLI scores 59. Carbon is one point ahead on the overall Intelligence Index (59 vs 58).

Is Carbon better at cybersecurity?

Google reports state-of-the-art results on SEC-bench Pro, ExploitGym and CWE-bench, but publishes no Opus 5.5 numbers on those benchmarks, so there is no like-for-like comparison yet.

Should I wait for Carbon?

There is no date for developer access. If you need a model now, Claude Opus 5.5 and Gemini 4 Argon are both available.

Does Gemini 4 Carbon have an API or model ID?

No public API and no published model ID yet. Access is limited to selected cyber defenders and Fairwind Program partners.

Sources

  1. Artificial Analysis: Gemini 4 Carbon is the new #1 (Oct 28, 2026) — Index and Coding Agent Index scores, Terminal-Bench 4.0, tokens and cost per task, AA-Omniscience, hallucination, AA-Briefcase, Argon comparison.
  2. Artificial Analysis Intelligence Index leaderboard — Live leaderboard and the “No publicly available API” label.
  3. Google: Gemini 4 Carbon announcement (Oct 28, 2026) — Long-horizon claims, cybersecurity benchmarks, Gray Swan IPI, access plan and prices.
  4. Google DeepMind: Fairwind Program — Who the Fairwind Program is for.
  5. Google AI for Developers: Gemini API models — Public Gemini API models; Carbon not listed as of October 29, 2026.
Попробуйте с одним API-ключом

Все модели из статьи доступны в API Many. Оплата по факту от $10.

Смотреть модели →

Ещё в блоге