Gemini 4 Carbon opens paid access as Antigravity previews /swarm
By AgentRiot Editorial
Google's deeper-reasoning model reaches paying API users, while Antigravity adds coordinated agent teams with separate model and plan limits.

Gemini 4 Carbon is moving beyond its initial cybersecurity-defender release. On November 10, Google opened the model to paying API developers and enterprises and began rolling it out to Google AI Ultra subscribers. Standard API pricing starts at $2 per million input tokens and $10 per million output tokens through November 30, then doubles on December 1, according to Google's announcement (opens in new tab) and API pricing table (opens in new tab).
The accompanying Antigravity update is more than a model-picker addition. Its new /swarm preview lets a coordinator divide work among agents, send their findings through a shared board and bring in verifiers to challenge changes. But the two releases have different access rules: Carbon is rolling out in Antigravity for Ultra subscribers (opens in new tab), while the canonical /swarm documentation (opens in new tab) lists the command for both Pro and Ultra.
For teams evaluating the combination, the question is where deeper reasoning earns its extra token use, and whether a task divides cleanly enough to benefit from parallel workers.
What is available, and where
The paid Gemini API release (opens in new tab) includes AI Studio and Android Studio access; enterprises can use Carbon through Gemini for business. Consumer access in the Gemini app and AI Mode in Search is rolling out to Ultra subscribers. That is a broader release than the October 28 Carbon launch (opens in new tab), not a claim that every account has received it simultaneously.
Within Antigravity, Google names the App, CLI and IDE extensions as Carbon surfaces for Ultra subscribers; its model reference also lists Enterprise (opens in new tab). The /swarm preview is narrower (opens in new tab): App and CLI, not every IDE extension. Pro access to the orchestration feature does not establish Pro entitlement to Carbon; workers can only use models available in the user's model picker (opens in new tab).
Scientists can apply for six months of access to Argon and Carbon through Antigravity's highest usage tier, but Google describes rolling onboarding subject to eligibility, capacity and selection (opens in new tab). It is an application program, not an immediate free-access offer for anyone with a research project.
The introductory tariff is not the permanent bill
Carbon's exact API model ID is gemini-4-carbon. It has no Free-tier availability, and output billing includes thinking tokens. Google's current Carbon tariff (opens in new tab) lists these rates in US dollars per million tokens:
| Inference tier | Input through Nov. 30 | Output through Nov. 30 | Input from Dec. 1 | Output from Dec. 1 |
|---|---|---|---|---|
| Standard | $2 | $10 | $4 | $20 |
| Batch | $1 | $5 | $2 | $10 |
| Flex | $1 | $5 | $2 | $10 |
| Priority | $3.60 | $18 | $7.20 | $36 |
The discounts buy different service behavior. Batch (opens in new tab) is asynchronous, with a target turnaround of 24 hours. Flex (opens in new tab) trades latency and best-effort availability for its lower rate. Priority (opens in new tab) prioritizes traffic above Standard and Flex. They are not interchangeable ways to price a latency-sensitive interactive agent.
Implicit cached input (opens in new tab) costs $0.10 per million tokens on Standard during the promotion, then $0.20. Batch and Flex list $0.05, then $0.10; Priority lists $0.18, then $0.36. Explicit caching is unsupported. The same table lists an invite-only Contributor preview at $1 input, $5 output and $0.05 cached input, with content used to improve Google's products. That differs from the table's no-product-improvement treatment for the ordinary paid tiers.
For a synthetic workload of one million uncached input tokens and 200,000 billed output tokens, including thinking, Standard costs $4 during the promotion and $8 afterward. Batch or Flex would be $2 and $4; Priority would be $7.20 and $14.40. These calculations apply the published tariffs (opens in new tab) to fixed token quantities, excluding cache hits, grounding and other tool charges. They do not predict a completed task's bill or assume different models consume identical tokens.
That distinction matters because Google explicitly says Carbon spends more thinking tokens and takes longer per turn than Argon (opens in new tab). Their per-token rates match, but their actual consumption need not. Antigravity subscription allowances are also a separate access surface, not the API tariff: its plans page (opens in new tab) describes rate limits and credit pools rather than a universal per-task price.
One shared context budget, not two million-token allowances
Carbon's API documentation (opens in new tab) specifies a unified 1,048,576-token window shared by input, thinking and output. A large input leaves less space for reasoning and generation. The million-token long-decode headline does not mean a million-token input plus a separate million-token answer.
The default combined thinking/output budget is 65,536 tokens. Google's long-decode guide (opens in new tab) distinguishes non-streaming foreground hops, capped at 32,768 tokens, from streaming or background hops, capped at 65,536. Longer generation uses continuation; Google's recommended path is background execution with an explicitly higher max_output_tokens. Background execution alone retains the default 64K budget. Each continuation processes prior context again as billable input, with implicit caching potentially reducing that charge.
The Carbon guide (opens in new tab) says max_output_tokens covers thinking and output together. It offers low, medium and high thinking levels, with medium the default.
The practical migration risk is an output budget consumed by reasoning before the deliverable is finished. Raising that budget allows more work, but also more spend; lowering the thinking level is a separate control.
How /swarm coordinates the work
Typing /swarm makes the current conversation the coordinator. It splits the assignment, names workers and posts public briefs describing ownership, exclusions and completion criteria. Workers start with clean contexts. A verifier reviews diffs, runs tests and reports problems back to the responsible worker, according to the feature documentation (opens in new tab).
The launch post (opens in new tab) describes an append-only message board and adversarial verification. The important difference from a string of delegated subtasks is that workers exchange questions, evidence and test results on a common board rather than routing every exchange through the coordinator. The App exposes Board, Members and Graph views; the CLI offers /swarm board or Alt+S for the live terminal board.
Workspace isolation needs care. Although the launch post emphasizes isolated git worktrees, the detailed docs say workers share the workspace by default (opens in new tab). The coordinator can give concurrent editors separate working copies. Teams that need that separation should request it explicitly. Separate worktrees keep edits apart; they do not, by themselves, establish an operating-system security boundary or remove the need to control credentials and tool permissions.
Model assignment is similarly explicit. Workers default to the conversation's model. A prompt can request different models for particular roles, or the coordinator can switch a waiting worker later, within the available picker. Google's suggested pattern (opens in new tab) puts Carbon on architecture, coordination or auditing and lighter models on routine implementation, rather than requiring Carbon for every worker.
Idle members wait without consuming tokens, but active parallel agents consume quota much faster than one conversation. A member that exhausts quota pauses; it does not automatically restart when quota refreshes. The docs also preserve pending tool approvals (opens in new tab), with controls to jump to the waiting worker. A verifier's passing test report is not a substitute for a user's approval of a consequential tool action.
A migration that could justify a team
Consider a hypothetical change to a monorepo with a shared event-schema package, an API service and a webhook delivery worker. The task is to add idempotency keys without breaking older producers. This is an illustration of the workflow, not a tested Carbon run.
The initial brief would define compatibility tests, forbid weakened assertions and request separate working copies for concurrent edits. A Carbon coordinator could assign an architecture worker to settle the schema and migration sequence before implementation starts. That worker would post the field definition and compatibility rules to the board, giving the other workers one contract to follow.
One worker could update the schema package, another the API's validation and persistence, and another the webhook worker's retry handling. They could flag conflicts on the board, such as whether a missing key should be generated at ingestion or rejected. The coordinator would resolve that dependency before incompatible assumptions spread across packages.
A verifier would then inspect the combined changes, check duplicate-delivery and old-client cases, and run integration tests after the branches are brought together. A failed retry test would go back to the delivery worker with the failure evidence, not become a reassuring final summary. The human handoff would include the diffs, actual test outcomes and unresolved issues, with deployment left as a separate decision.
That is the useful unit of parallelism: independent edits against a shared contract, followed by integration checks. A change concentrated in one tightly coupled file may gain little from adding more agents.
The fresh benchmarks show a trade-off, not a universal winner
Current operator results give the rollout more substance than Google's broad first-place language alone:
| Evaluation | Carbon result | Relevant comparison |
|---|---|---|
| CursorBench 4.0 (opens in new tab) | High: 60.4%, $9.22 per task | Fable 5.2 Max: 60.0%, $32.79; Argon High: 52.9%, $4.57 |
| Zapier AutomationBench 1.0.6 (opens in new tab) | High: 53.88%, $2.12 per task | Fable 5.2 with default fallbacks, Max: 51.45%, $4.64 |
| Multimodal Knowledge Agent (opens in new tab) | High: 54% Climb weighted satisfaction | Fable 5.2 Max with default fallback: 57%; Opus 5.5 Max with default fallback: 53% |
Cursor warns that small score differences may not be statistically meaningful. Its costs apply published input, cache and output prices to observed token use. Carbon's narrow lead over Fable is therefore different from a decisive quality gap, and Argon's lower score comes with a lower task cost.
Zapier scores strict completion through deterministic checks of the resulting simulated business state, rather than an agent's claim of success. Its methodology uses API-mode tools, and its Claude fallback configurations matter to the comparison. A leading 53.88% still leaves substantial room for failure; it does not justify unattended access to production business systems.
The new MKA benchmark is especially relevant to Google's knowledge-work pitch. Artificial Analysis runs 200 tasks across 10 industry verticals (opens in new tab), using Stirrup, a 250-turn limit and a GPT-6 Luna medium agentic judge. Climb weighted satisfaction measures the harder rubric criteria, not the share of fully completed tasks. Google developed the benchmark with Artificial Analysis; the operator conducts the evaluations independently.
Carbon is the runner-up here. Google reports roughly sixfold lower average task cost than the top scorer (opens in new tab). Reconstructing the operator's cost formula (opens in new tab) from its public token and standard-price inputs gives about $6.29 for Carbon versus $31.98 for Fable, approximately 5.1-fold, excluding judge costs. The difference in pricing or snapshot basis remains unresolved. These are benchmark-specific estimates, not invoices, an Antigravity swarm quote or a guaranteed cost ratio after the promotion.
Public Carbon is not Carbon for cyber
The broader rollout does not remove the distinction between public Carbon and the stronger cybersecurity access offered through the Fairwind Program (opens in new tab). Carbon for cyber remains a rolling application route for defenders. Paying for public API access is not equivalent to joining it.
Google reports a 0.8% per-attempt breach rate in simulated cyberattack testing, down from Argon's 3.1%. Its full safety report (opens in new tab) also reports 27.8% cumulative rollout breaches across 18 adaptive attack rollouts under the allowed retry budget. Individual-attempt and whole-rollout rates are different denominators, and neither estimates normal deployment risk. The earlier zero-success Gray Swan prompt-injection result applies to a held-out replay benchmark; adaptive tests still expose weaknesses. Those results do not justify trusting arbitrary repository instructions or unrestricted tools.
Paid access does not mean one privacy policy
Under the ordinary paid API terms (opens in new tab), content is not used to improve Google's products, but that is not zero retention: abuse monitoring retains prompts, context and output for 55 days (opens in new tab). Contributor Tier and voluntary sharing (opens in new tab) have different data-use terms. Background execution requires stored interactions (opens in new tab); switching off that storage is not the same as removing abuse monitoring.
The consumer Antigravity terms (opens in new tab) allow interactions to improve products and receive human review, with an interaction-sharing preference users can change. Gemini's consumer app has separate Keep Activity controls (opens in new tab). Do not infer either subscription surface's treatment from Developer API billing.
What to test first
For a first evaluation, choose one migration with clear package boundaries, a known test suite and an explicit spend budget. Keep Argon as the routine-work baseline, place Carbon where an expensive reasoning mistake would matter, and compare completed, verified changes against total tokens and elapsed time. That will say more about the rollout's value than the size of the agent roster.

