What changed on November 10, 2026?
Google opened Gemini 4 Carbon to paying developers, businesses, researchers and consumers on November 10, 2026. The model had spent two weeks behind a gate reserved for vetted cybersecurity defenders. Google announced that earlier release on October 28.
Developers can now reach Carbon through the paid Gemini API in Google AI Studio and Gemini for business. Enterprises get the same model through the API, a model drop-down, Google Antigravity or specialist agents such as CodeMender and AlphaEvolve.
Google AI Ultra subscribers can select Carbon in the Gemini app, AI Mode in Search and Antigravity. Engadget reports that Ultra costs about $100 a month. Google lists no free consumer tier for Carbon.
Engadget describes the release as worldwide, but Google's post lists product surfaces rather than countries. Access still depends on whether an account can pay for the API or hold an Ultra subscription.
The change matters because access, not capability, was the story at launch. Carbon existed and scored well, but ordinary teams could not test it. From this week, any paying API customer can route production traffic to it and measure the results.
Google is also offering Carbon to scientists through a program that opened four weeks ago. It gives 10,000 researchers six months of free access through Antigravity's highest usage tier. Existing and future participants now receive Carbon as well, and Google says thousands have already been onboarded.
- •Paying API developers: Google AI Studio, Gemini for business
- •Enterprises: API, model picker, Antigravity, CodeMender
- •Consumers: Gemini app and AI Mode, AI Ultra only
- •Security defenders: separate cyber edition through Fairwind
Is Gemini 4 Carbon now the default Gemini model?
No. Google is keeping Gemini 4 Argon as the everyday option and positioning Carbon as the model for unusually hard work. Argon launched roughly two weeks before Carbon and remains available to AI Pro subscribers and above.
The Gemini app now labels Argon for "Code & analysis" and Carbon for "Most complex work." 9to5Google reports that the app also removed Gemini 3.1 Pro when Carbon arrived. That leaves two frontier choices rather than three.
Google published a model card and an updated developer guide alongside the release. The guide covers migration, which suggests the company expects teams to move existing Argon workflows rather than start from scratch.
Google's guidance is unusually blunt about the trade-off. Argon "remains the ideal choice for the majority of software engineering and knowledge work," the company wrote. Carbon is meant for complex, multi-step coding, deep debugging and open-ended scientific discovery.
The distinction is practical, not cosmetic. Carbon explores more options and runs longer verification loops than Argon. Google confirms that it uses more tokens for the same task. Teams that switch everything to Carbon should expect higher bills even at identical per-token prices.
For most product teams, the sensible path is to keep Argon as the default and route a narrow set of hard tasks to Carbon. Those tasks include large refactors, cross-service bug hunts and research questions that need many tool calls.
Google also markets Carbon for legal, finance and enterprise operations, not just engineering. It says the model follows strict schemas to produce spreadsheets, presentations and reports. Buyers in those functions should test document-heavy work directly, because public benchmarks lean toward code and science.
What does the /swarm preview in Antigravity do?
Google used the rollout to preview a feature called /swarm inside Antigravity, its agentic coding environment. A single prompt becomes a coordinated team of agents that a developer can watch in real time.
Carbon takes the roles that need the deepest reasoning, such as architectural design, coordination and auditing. Faster, lighter models work in parallel isolated workspaces, and adversarial verifiers test every proposed change.
Google published several demonstrations. One built an interactive Three.js reconstruction of its Platform 37 office from seven public City Council planning documents. The run used 206 extracted engineering drawing views and ten overnight loops.
Another demo turned historical photographs and engineering plates into a 3D exhibition of the Eiffel Tower's 1887 to 1889 construction. Others included an open-world game and a playable 1780 fortepiano with 61-key escapement physics.
These are company-selected showcases, not controlled studies. Google's post gives no failure rates for /swarm, nor independent comparisons against human teams or rival agent frameworks. Treat the demos as a capability preview rather than evidence of reliability.
The feature also raises a cost question. Many agents working in parallel multiply token consumption, and Carbon already uses more tokens than Argon. A swarm that finishes faster may still cost more per completed task than a single strong agent.
Swarm access is limited to Antigravity users, which currently means Google AI Ultra subscribers or enterprise customers. Developers on the plain API cannot invoke it.
Ordinary subagents share a context and a plan. Swarm separates them into isolated workspaces, so one failed branch should not corrupt another. The adversarial verifiers add a second model whose job is to break the first model's patch. That design is plausible, but it also multiplies the number of model calls.
How does Gemini 4 Carbon compare with Claude Opus 5.5 and GPT-6 Astra?
Artificial Analysis, an independent evaluator, placed Carbon first on its Intelligence Index v5 with a score of 59. The model led four of the nine evaluations inside that index, including HLE-Diamond, Terminal-Bench-Science, AA-Omniscience and AutomationBench-AA.
The lead is not uniform. On the same lab's Coding Agent Index, Carbon scored 59 with the Antigravity command-line tool. Claude Code running Claude Opus 5.5 scored 62, leaving Carbon three points behind.
Agentic knowledge work tells a similar story. Carbon reached 1,609 Elo on AA-Briefcase, a 126-point gain over Argon. It still trailed Grok 4.7 at 1,638, Muse Wonder at 1,659 and Kimi K3.1 Pro at 1,678.
Google prefers a different set of tests. It says Carbon ranks first on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench. Google also claims more than twice lower cost per task than the next best external model on those three.
Cost depends heavily on how you measure it. At standard pricing, Artificial Analysis puts Carbon at $7.76 per Intelligence Index task. That is about 23 percent less than Claude Opus 5.5 at $10.04, but roughly 1.2 times GPT-6 Astra at $6.25.
Carbon used 123,000 output tokens per task in that measurement, 29 percent fewer than Opus 5.5. The efficiency comes from producing fewer tokens per task, not from a lower sticker price. At $4 and $20 per million tokens, Carbon and Argon cost almost the same per task.
Google and Artificial Analysis also built a new Multimodal Knowledge Agent benchmark for dense documents. Carbon ranked second there while costing more than six times less per task than the top model. That combination of near-top accuracy and lower cost may matter more to buyers than a single headline score.
The new benchmark uses 200 tasks built around charts, images, blueprints, schematics and tables, according to 9to5Google. Google argues that older tests missed how agents reason over such material. The claim is reasonable, but a benchmark designed alongside the model it measures deserves scrutiny.
Google's own numbers come from internal and partner tests. That does not make them wrong, but it does mean the company chose what to measure. Independent labs will need time to reproduce the security results in particular.
Does the public version keep the cybersecurity powers?
Partly, and the split is deliberate. The model sold through the API is a public edition with safeguards tuned for general use. The strongest cyber capability remains behind the Fairwind Program, which requires an application and vetting.
Google reported sharp security results at launch for the model it gave defenders. It scored 88.2 percent on SEC-bench Pro for finding vulnerabilities, 54.8 percent on ExploitGym for exploitation and 71.0 percent on CWE-bench for patching. Those figures came from evaluations run without guardrails.
The public edition is measured differently. In simulated attack environments seeded with ExploitBench and CyScenarioBench, Google says Carbon blocked 99 percent of attack trajectories from an adaptive red-teaming agent. The per-attempt breach rate fell to 0.8 percent from 3.1 percent on Argon.
Google also cites a drop in automated cyberattack query success to 2.17 percent in Gray Swan's testing, down from 9.86 percent on Argon. External tester Trajectory Labs found no new universal jailbreaks across 132 hours and more than 13,000 requests.
Prompt injection is the area Google flags most openly. Carbon reached a 0 percent attack success rate on Gray Swan's indirect prompt injection benchmark, which Google calls a first. It then concedes that advanced red teaming shows models could improve against sophisticated attacks in some distributions.
The company published separate honesty results. On a 375-task agentic benchmark, Carbon passed 92 percent of tasks against 70 percent for Argon. It declined to fill in missing experimental data 98 percent of the time and flagged methodological flaws 68 percent of the time.
Those numbers are self-reported and should be read as such. Google's own security teams, Wiz and outside testers contributed examples, but the benchmark selection and scoring remain under Google's control. Independent replication is still thin.
The practical effect is a two-tier market. Defenders who join Fairwind get the sharper tool. Everyone else gets a model that Google says resists most attacks but not all of them.
Google's own deployments supply the most concrete examples. Its Project Fortify system scans, triages and remediates code. In an internal benchmark it reported 100 percent recall on scanning and 95 percent precision on triage. On open-source repositories, Carbon improved recall by more than 25 percent relative to Argon.
Wiz, working through its Scan for Good program, used Carbon to scan two major SaaS providers. Google says the model uncovered zero-day vulnerabilities in under two hours, work that would normally take weeks. The affected vendors were contacted and patched, though Google has not named them.
How long does the $2 introductory price last?
The discount expires on November 30, 2026. Until then, Carbon costs $2 per million input tokens and $10 per million output tokens. That matches the introductory price Argon launched with.
From December 1, the standard rate is $4 per million input tokens and $20 per million output tokens. Cached input tokens receive a 95 percent discount, which works out to $0.20 per million at standard pricing.
The deadline now has a firm date. When Carbon launched, Artificial Analysis noted a 50 percent discount with no confirmed end date. Google's November 10 post closes that gap and gives finance teams a date to plan around.
Cheap tokens do not guarantee a cheap workflow. Carbon reasons for longer, runs more verification and can burn through context on large repositories. A team that pays $2 per million tokens but uses three times as many tokens may still spend more than it did on Argon.
Buyers should compare cost per completed task rather than price per token. Artificial Analysis already does this and finds Carbon competitive with Opus 5.5 but not with the cheapest frontier options. The gap will widen after November 30.
Individual developers should also weigh the subscription. Ultra costs about $100 a month, and API usage is billed separately. A user who only needs occasional hard tasks may spend less by paying per token than by holding a subscription.
What should developers and security teams watch next?
The first thing to watch is independent testing of /swarm. Google's demos are impressive, but multi-agent systems fail in mundane ways: duplicated work, conflicting edits and runaway loops. No outside lab has published reliability data yet.
The second is the Fairwind boundary. If Google widens the cyber edition, more defenders gain access to the strongest vulnerability-finding tools. If it stays narrow, the public model's safeguards become the main control on how those capabilities spread.
The third is competition. Claude Opus 5.5 still leads the Coding Agent Index, and cheaper models such as GPT-6 Astra undercut Carbon on cost per task. Google's claim to leadership rests on selected benchmarks, not across-the-board dominance.
The fourth is honesty. Carbon's accuracy rose, but its hallucination rate on AA-Omniscience rose from 4 percent to 11 percent. That is still the second lowest among tested models and far below the 50 percent recorded for Opus 5.5 and GPT-6 Astra. Teams should still verify factual output.
The fifth is the price clock. December 1 changes the economics for anyone who adopted Carbon during the discount. Teams running large agent workloads should model both prices before committing a production pipeline.
Google has moved quickly here. Carbon reached defenders on October 28 and paying customers on November 10, a gap of thirteen days. The pace suggests further releases will follow the same staged pattern: trusted testers first, then enterprises, then consumers.
Procurement teams should ask which model version produced a given result. Carbon and Argon share a family name but differ in cost, token use and honesty rates. A benchmark run on Argon does not transfer automatically to Carbon.
