Key Notes
- Google calls Gemini 4 Carbon its strongest model, and Artificial Analysis places it first on Intelligence Index v5 at 59 while Claude Opus 5.5 retains the coding-agent lead.
- Initial access is limited to selected cyber defenders, with a phased Fairwind partner rollout and no announced date for broad availability.
- Introductory API rates are $2 per million input tokens and $10 per million output tokens, rising to standard rates of $4 and $20 after an unspecified launch period.
Google has unveiled Gemini 4 Carbon, calling it its strongest AI model yet and targeting the software engineering and cybersecurity work that increasingly defines the frontier AI race. Carbon enters at the top of Artificial Analysis’s Intelligence Index v5 with a score of 59, while its first rollout will be limited to selected cyber defenders.
The combination is significant: Google is presenting a general-purpose capability advance through a controlled security deployment. Introductory API rates are $2 per million input tokens and $10 per million output tokens, with higher standard rates to follow. A broad release date has not been announced.
Carbon follows the Gemini 4 Argon rollout we covered earlier this month. This release raises the question from whether Google can compete with the leading labs to where its newest model provides the strongest practical advantage.
What Carbon’s First-Place Score Means
The headline result is a lead on a newly revised test. Artificial Analysis introduced Index v5 on October 27, adding harder coding and scientific workflows and updating several existing evaluations. It says 60% of the composite now rests on private test sets or held-out answers.
The index gives equal overall weight to knowledge work, coding, science and general capabilities. Its nine evaluations therefore measure a broader mix than conversational fluency or a single programming test. A score of 59 is an index value, rather than a claim that Carbon correctly completes 59% of every kind of task.
Artificial Analysis’s Carbon analysis credits the model with first place in four of those nine evaluations. That supports a strong overall result while leaving room for competitors to lead individual specialties. Scores from earlier index versions also need their original context because the test mix has changed.
Claude Still Leads Coding Agents, While Muse and Kimi Have Other Strengths
On Artificial Analysis’s separate Coding Agent Index, Carbon scores 59 with the Antigravity CLI, compared with 62 for Anthropic’s Claude Opus 5.5 in Claude Code. Carbon is second in that comparison. The matching number on the two indices does not make them the same measurement.
The knowledge-work picture is similarly competitive. Artificial Analysis reports a substantial gain over Argon on AA-Briefcase, but Carbon remains behind models including Grok 4.7, Meta’s Muse Wonder and Moonshot AI’s Kimi K3.1 Pro. Those results make a blanket claim that Carbon wins every major benchmark difficult to sustain.
Google’s own comparison page also shows a mixed frontier: Carbon trails OpenAI’s GPT-6 Astra on FrontierSWE 2 and Claude Opus 5.5 on the listed OSWorld offline subset. The competitive pattern is more informative than a single winner label: coding, desktop operation and document work can favor different systems.
That is the broader race we examined with Muse Wonder. Carbon strengthens Google’s position, while teams choosing a model still need to match it to the work they actually want completed.
Cybersecurity Has the Clearest Early Results
Carbon’s security capabilities span finding flaws, establishing that they are exploitable and generating fixes. Google reports 88.2% on SEC-bench Pro and 54.8% on ExploitGym. Its evaluation notes identify both results as self-computed using published evaluation setups and competitor figures drawn from their reports.
Those notes provide useful boundaries. The SEC-bench result uses a 183-task JavaScript-engine split, while ExploitGym runs in an isolated setting without internet access and with mitigations disabled. Google’s announcement also says the external cybersecurity capability benchmarks were evaluated without guardrails. These settings describe a capability assessment; they should not be read as the permissions or protections available in an ordinary product session.
The patching result has a separate public reference. The public CWE-bench leaderboard places Carbon first at 71% programmatic pass@1, ahead of the 68% listed for Argon, GPT-6 Astra and Grok 4.7. The benchmark evaluates agents auditing and repairing repositories, with checks intended to ensure the vulnerability is blocked and existing functionality still works.
The distinction matters operationally. A convincing vulnerability report creates work for a security team; a valid patch can remove some of it. Neither number, by itself, measures the review effort, integration problems or previously unseen failure cases an organization will encounter after deployment.
The timing follows the AI-assisted research behind the Chrome 156 fixes. Defensive capability is becoming a concrete test of model usefulness, alongside the risk that similar skills can support unauthorized attacks.
Introductory Pricing Will Double After the Launch Period
Google’s announcement sets the introductory rates at $2 per million input tokens and $10 per million output tokens. After that period, the rates become $4 and $20 respectively. Cached input receives a 95% discount from the applicable input rate.
That makes cached input $0.10 per million tokens at the introductory rate and $0.20 at the standard rate. Google has not given an end date for the introductory period. Budgeting for sustained use therefore requires keeping the standard rates visible rather than treating launch pricing as permanent.
Token prices also tell only part of the cost story. Artificial Analysis’s model profile puts Carbon at $7.76 per Intelligence Index task using standard prices. Its accompanying analysis places that below Claude Opus 5.5’s $10.04 but above GPT-6 Astra’s $6.25 in the compared configurations.
Those are evaluation costs, not quotes for a company’s workload. An agent that produces fewer unsuccessful attempts can justify a higher token rate, while long reasoning runs or repeated tool calls can erase an apparent price advantage. Cost per completed and accepted task is the more useful comparison.
Access Begins With Vetted Cyber Defenders
Google says a subset of cyber defenders will receive Carbon this week, followed by additional trusted partners through Fairwind. This is a phased introduction, rather than general availability for everyone using the Gemini app or developer tools.
The existing Fairwind program prioritizes governments, critical infrastructure operators and core technology platforms. It describes organizational vetting, user-level access controls and restrictions on reselling or redistributing frontier-model access. Its purpose is to give defenders an early opportunity to strengthen systems against capabilities that could also be misused.
Selection for a broader partner program should not be assumed to mean immediate Carbon access. Google has not published a universal launch timetable or promised that every existing partner receives the new model at once. For most readers, the relevant status is announced capability with limited initial deployment.
Long-Running Work Raises the Reliability Bar
Artificial Analysis lists a one-million-token context window and support for text, image, video and speech input, with text output. Google also says its teams have seen Carbon persist through engineering tasks for days. Large context and persistence can help with complex work, but duration is useful only when the model continues to pursue the right objective.
Google reports fewer severe misalignment alerts in internal coding use and a 0% attack-success rate on its evaluated Gray Swan indirect-prompt-injection set. These are results under specified tests and deployments, not a guarantee that every malicious instruction or unauthorized action will be prevented.
For developers, Carbon’s next test will be sustained performance on their own repositories, tools and acceptance criteria. The benchmark lead gives Google a stronger starting position. The restricted rollout will show how much of that capability translates into reliable defensive work when a human team must review, deploy and maintain the result.
Disclaimer: AIstify is an independent media brand owned and operated by NuvexMedia LLC, publishing news, research, and insights on artificial intelligence, emerging technologies, automation, and related industries. NuvexMedia LLC invests in and collaborates with companies across the AI, technology, software, and digital innovation sectors. These relationships do not influence AIstify’s editorial coverage, and the publication maintains full editorial independence to provide accurate, timely, and objective information. © 2026 NuvexMedia LLC. All rights reserved. This content is for informational purposes only and should not be considered legal, tax, investment, financial, or other professional advice.
