Gemini 4 Carbon goes public, and its introductory price ends 30 November

Gemini 4 Carbon, which Google calls its strongest model, became available to paying developers, businesses and Google AI Ultra subscribers on 10 November 2026. It costs the same per token as Gemini 4 Argon, $2 per million input tokens and $10 per million output tokens, but only until 30 November. From 1 December both models move to $4 and $20. A team that adopts Carbon this week is therefore budgeting across a price that doubles in under three weeks, on a model that Google itself says spends more tokens than Argon to get its results.

That is the practical news for builders. The rest is detail that changes the bill: Batch and Flex tiers cut every rate in half, Priority adds 80 per cent, cached input is billed at five per cent of the input rate, and Search grounding has its own per-query charge and daily cap. Alongside the model, Google is previewing /swarm in its Antigravity coding agent, a command that turns one prompt into a coordinator, workers and verifiers that share a message board. Google’s documentation warns that a swarm uses quota much faster than a single conversation.

The benchmark story needs one correction before anyone repeats it. When Google released Carbon to cyber defenders on 28 October, it said Carbon ranked first on the Artificial Analysis Intelligence Index v5.0, at 59. On 12 November the live v5.0.1 leaderboard puts Carbon at 60 at its high thinking level, in fourth row, behind three configurations of Anthropic’s Claude Fable 5.2, which shipped on 9 November. Google’s 10 November announcement does not repeat the index claim. This article sets out what Carbon is, what it costs, a worked budget across the price change, how /swarm works according to its documentation, and what the evidence does not yet settle. We have not run Carbon ourselves; every result below is attributed to whoever measured it.

What shipped on 10 November, and to whom

Google’s 10 November announcement, written by Koray Kavukcuoglu, makes Carbon available on five kinds of surface. Developers get it on the paid tier of the Gemini API through Google AI Studio and in Android Studio. Enterprises get it in Gemini for business, through the API or the model dropdown, and in Google Antigravity and CodeMender. Consumers on Google AI Ultra get it in the Gemini app, in AI Mode in Search and in Antigravity. Scientists can apply for six months of free access to Argon and Carbon through Antigravity’s highest usage tier, assessed on a rolling basis. Cyber defenders can still apply to the Fairwind Program.

The positioning is explicit. Google says Argon “remains the ideal choice for the majority of software engineering and knowledge work”, and that Carbon is meant for harder tasks: complex multi-step coding and debugging, and open-ended scientific research. The model card, published in November 2026, says Carbon “is based on Gemini 4 Argon” and refers readers to Argon’s card for architecture, training data, hardware and software. It describes Carbon as using “more thinking tokens than Gemini 4 Argon to explore hypotheses thoroughly, reason deeply, and run extended verification loops.”

The card’s technical facts are short. Inputs are text, images, audio and video, with a combined input and output context window of up to 1 million tokens. Output is text. The knowledge cutoff is March 2026, “though in some domains the model’s knowledge may be less up to date.” The known limitations it lists are hallucination, ongoing work on jailbreak resistance, “occasional slowness or timeout issues”, and the model using more tokens “especially at higher thinking levels.”

The API model page lists the model code gemini-4-carbon as stable, last updated November 2026, with a 1,048,576-token unified window and a maximum output of 65,536 tokens for a default synchronous response, or up to 1,048,576 with long decode. It supports function calling, code execution, computer use, file search, structured outputs, URL context, and grounding with Google Search and Google Maps. It does not support the Live API, image generation or audio generation. Batch, Flex and Priority consumption are all supported. Context caching is implicit only.

The 13-day pattern

Carbon followed the same release sequence as Argon, and the dates explain why Carbon’s introductory window is so short. Argon was announced on 30 September for “a set of trusted cyber defenders through our Fairwind Program”, then made globally available on 13 October. Carbon was announced on 28 October for “a subset of cyber defenders” and opened to the public on 10 November. Both gaps are 13 days. Both models carry the same introductory deadline, 30 November, so Argon’s public customers had seven weeks at the low price and Carbon’s have three.

Reconstructed by GenAI Brief from the sources listed. Event dates are the dates given on each page; the leaderboard row is our reading on the publication date of this article.

Our October ranking of frontier models left Argon out because, on 5 October, it was still limited to Fairwind. That gap closed on 13 October, and Carbon has now followed it.

The price list across the 1 December change

The Gemini API pricing page lists Carbon and Argon with identical rates on every tier. The page states both models are unavailable on the Free tier, and that paid-tier content is not used to improve Google’s products. There are no long-context price bands for either model; a 900,000-token prompt is billed at the same per-token rate as a 9,000-token one.

Gemini 4 Carbon and Gemini 4 Argon, paid tier, as listed on Google’s Gemini API pricing page on 12 November 2026. Output includes thinking tokens.

Three billing rules matter more than the headline rates.

Thinking is output. The pricing page and the Carbon migration guide both say every processed token is billed, including thinking tokens. A model that thinks longer costs more per task at the same per-token price. With identical rates, the cost difference between Carbon and Argon comes down to how many tokens each spends.

Tool scaffolding is billed. When grounding is used, Google bills the internal tool system instructions and function declarations as tokens. Grounding with Google Search is also charged per query: 5,000 free requests a month shared across Gemini 3 and newer models, then $14 per 1,000, with a default paid-tier cap of 1,500 requests per day per project. Google’s footnote adds that one customer request may trigger more than one Search query, and each is charged. Grounding with Google Maps has the same free allowance and per-1,000 price.

Caching is implicit and unguaranteed. Carbon does not support explicit caching. Implicit caching is on by default for prompts of at least 4,096 tokens and discounts cached input by 95 per cent on a hit. Our prompt caching explainer covers how Google’s implicit caching works across Gemini models, including the Gemini 4 rule that in multi-turn agent loops the previous request is cached only when the new request adds at least 2,050 tokens. The practical consequence for budgeting is that the cached-input column is an outcome you measure, not a rate you can rely on.

What the cheaper tiers trade away

Batch and Flex both halve the bill, and from 1 December they bring Carbon back to exactly the introductory Standard price. They are not interchangeable. The Batch API has a 24-hour target turnaround, “in majority of cases, it is much quicker”, and a job expires if it has been running or pending for more than 48 hours. Flex inference, in preview, keeps a synchronous interface with a target latency of one to fifteen minutes, but its traffic is “sheddable”: requests may be delayed, or dropped with a 503, when Standard traffic is high. There is no automatic fallback to Standard, so the caller must implement retries with backoff and raise client timeouts to ten minutes or more.

Batch capacity is also capped by account tier. Google’s rate limits page limits batch enqueued tokens, the tokens queued across all active jobs for a model, to 5 million at Tier 1, 500 million at Tier 2 and 1 billion at Tier 3. Tier 2 requires $100 of spend and three days since the first payment; Tier 3 requires $1,000 and 30 days. The page does not publish requests-per-minute or tokens-per-minute values for Carbon; it tells developers to view their active limits in AI Studio.

A worked budget across the price change

The following example is hypothetical. The workload, token counts and multipliers are our assumptions, chosen to make the arithmetic visible. Only the per-token prices come from Google’s pricing page.

Inputs. A team runs an automated pull-request review agent on Carbon at the default medium thinking level, Standard tier. It processes 2,000 reviews a day. Each review sends 60,000 input tokens, of which 40,000 are a stable repository context that hits the implicit cache and 20,000 are the diff and instructions. Each review returns 8,000 output tokens, including thinking.

Per review, through 30 November. Uncached input is 20,000 × $2 per million, or $0.040. Cached input is 40,000 × $0.10 per million, or $0.004. Output is 8,000 × $10 per million, or $0.080. The total is $0.124 per review, or $248 a day.

Per review, from 1 December. Every component doubles: $0.080 + $0.008 + $0.160 = $0.248 per review, or $496 a day.

GenAI Brief calculation. Prices from the Gemini API pricing page on 12 November 2026; workload is invented for illustration.

Three things fall out of the table. First, output tokens are 65 per cent of the bill in this example, so thinking depth matters more than prompt size. Second, if the review can tolerate a turnaround of minutes or hours, moving it to Flex or Batch on 1 December keeps the November price. Third, the cache assumption is worth more than the tier choice in some cases: if the repository context stops hitting the cache, December’s per-review cost rises from $0.248 to $0.400, a 61 per cent increase, because 60,000 input tokens are then billed at $4 per million.

Splitting the month

A team switching on 12 November gets 19 days at the introductory price. At $248 a day, that is $4,712 for the rest of November. December at $496 a day is $15,376, a little over three times November’s remainder for a month only 12 days longer. If half the December traffic is latency-tolerant and moves to Batch, December falls to 31 × ($248 + $124) = $11,532.

The budget also has to include grounding. If each review triggered one Google Search query, 2,000 reviews a day would exceed the default paid-tier cap of 1,500 grounded requests per day per project, and the design would fail before it became expensive. Below the cap, 30 days at 1,500 queries is 45,000 queries; after the 5,000 free, 40,000 are billed at $14 per 1,000, or $560 a month on top of tokens. The pricing page points teams that need more to Vertex AI, which it says has a default quota of 1,000,000 search queries a day.

Carbon against Argon at the same price

Because Carbon and Argon cost the same per token, the comparison comes down to token counts and success rates. Independent data on token counts exists for one setting. On Artificial Analysis’s Carbon page, Carbon at high thinking used 205 million output tokens to run the Intelligence Index. On its Argon page, Argon at high used 113 million output tokens. That is a ratio of about 1.8 output tokens for Carbon per Argon output token, on that index, at that setting. Input volume was close: 7.62 billion tokens for Carbon and 7.55 billion for Argon.

If we apply that ratio to the hypothetical review agent, purely as an illustration, Argon would emit about 4,410 output tokens per review instead of 8,000. From 1 December, Argon’s cost would be $0.080 + $0.008 + $0.088 = $0.176 per review against Carbon’s $0.248, so Carbon would cost about 41 per cent more per review. The decision then becomes whether Carbon’s success rate on the team’s reviews is high enough to repay that.

The same Artificial Analysis data shows why the illustration should not be trusted without a measurement. Its cost per Intelligence Index task is $7.76 for Carbon at high and $9.95 for Argon at high, the opposite of what the token counts alone would suggest, since Carbon used more of both input and output. Neither page breaks the per-task cost into components, so we cannot say what drove the difference; cache-hit share and the number of turns per task both enter the per-task cost, and neither is reported. Both pages list Carbon and Argon at $4 and $20, the December prices, with a blended rate of $2.94 per million on a 7:2:1 cache-hit, input and output ratio. A team estimating its own costs from these figures should therefore treat them as post-December numbers.

What the benchmark claims measure, and who measured them

Google’s 10 November post claims Carbon ranks first on three external agentic benchmarks, CursorBench 4.0, Perplexity’s WANDR and Zapier’s AutomationBench, “at more than 2X lower cost per task compared to the next best external model.” It claims second place on the Multimodal Knowledge Agent (MKA) benchmark run with Artificial Analysis at “more than 6X lower average cost per task” than the top model. The post does not name the comparison models. We checked the three public leaderboards that we could read on 12 November.

Leaderboard values read by GenAI Brief on 12 November 2026. Scores are per model and effort configuration.

The checks mostly support Google’s narrower claims. On CursorBench, Carbon’s high and medium settings sit first and second, and Fable 5.2 at max costs about 3.6 times as much per task for a score 0.4 points lower, a gap Cursor itself says may not be meaningful. On AutomationBench, Carbon at high leads Fable 5.2 at max by 2.43 points at under half the cost per task. The leaderboard also places Argon at high fourth, at 51.29 per cent, only 0.16 points behind Fable 5.2, and Zapier notes it ranks Argon on standard list pricing rather than the promotional rate. Zapier’s page does not say which price it applied to Carbon.

MKA needs a caveat that Google’s post does not give. Artificial Analysis describes its leaderboard as its “implementation of Google DeepMind’s Multimodal Knowledge Agent (MKA) benchmark”, which it runs and scores. An evaluation designed by the model’s developer and run independently is more useful than a self-reported score, but it is not the same as a benchmark Google had no hand in. We could not read cost per task for the top MKA entries on the leaderboard page, so the “more than 6X” figure remains Google’s.

We did not verify WANDR, Hebbia’s financial agent evaluation, where Google says Carbon “matched the performance of the best model tested”, or Balyasny’s pre-release benchmark, where Google says Carbon ranked first of 23 frontier model configurations. Those remain vendor-reported.

The Intelligence Index claim, then and now

The 28 October post said Carbon “ranks #1 on the recently updated v5.0 of Artificial Analysis Intelligence Index at 59.” The live leaderboard is now on v5.0.1 and shows Carbon at 60 at high, behind Fable 5.2 at max (65), xhigh (62) and high (62). Counted as models rather than configurations, Carbon is second. Fable 5.2 did not exist on 28 October; Anthropic released it on 9 November at $8 input and $40 output per million tokens, exactly twice Carbon’s December price. Our Fable 5.2 briefing covers what Anthropic says that price buys.

Google’s evaluation methodology document is consistent with that timing. It reports results “as of October 2026”, compares Carbon with GPT-6 Astra, Opus 5.5 and Fable 5.1, sources non-Gemini results from “providers’ self-reported numbers unless otherwise mentioned”, and runs Gemini models “with the highest thinking settings unless indicated otherwise.” Any comparison with Fable 5.2 in Google’s materials would postdate that document.

The index composition matters for interpreting the gap. Artificial Analysis’s methodology page says v5.0.1 combines nine evaluations in four categories at 25 per cent each: knowledge work, coding, science and general. A five-point index gap at the top is a statement about that mix, not about any single job. Carbon’s lead on CursorBench and AutomationBench and its deficit on the index are not contradictory; they measure different things under different harnesses.

Latency is part of the price

Artificial Analysis reports a time to first answer token of 65.84 seconds for Carbon at high through Google’s API, against a class median of 13.45 seconds, and 26.4 minutes per Intelligence Index task against a median of 12.7. Output speed is 77 tokens per second. At medium thinking, time to first token falls to 25.18 seconds and the index score drops two points to 58. The migration guide’s advice is consistent with those numbers: keep thinking_level at medium unless a workload needs high, and use background execution or set HTTP client timeouts to at least 15 minutes.

Moving a workload from Argon or Gemini 3

Google’s migration guide says moving from Argon requires changing only the model ID; there are no request schema changes. It recommends allowing for longer passes in max_output_tokens and timeouts, and retaining Argon “when you want faster responses” and lower token spend per turn.

Moving from Gemini 3 or earlier involves more:

1. Change the model ID to gemini-4-carbon.
2. Raise max_output_tokens to at least 32,768 if it is set lower. It caps thinking and output combined.
3. Choose and test a thinking_level: low, medium (the default) or high.
4. Remove temperature, top_p and top_k. The guide says they are ignored if provided.
5. Remove cached_content references and explicit cache creation; neither is supported.
6. Use background execution for outputs over 64,000 tokens.
7. Set HTTP timeouts to at least 15 minutes, or use background execution.
8. Use previous_interaction_id for server-side conversation state in multi-turn and function-calling flows.
9. Recheck cost estimates, because grounding and tool instructions are billed.

One detail in the long-output path can silently truncate work. With background: true and max_output_tokens set above 65,536, the server continues generation in 64,000-token hops, each adding a model_output step, up to the 1 million window minus input. But the guide warns: “If you set background: true without defining max_output_tokens, the server stops after 64K tokens.” A job that expects a long artefact must set both.

A second detail affects cost comparisons between surfaces. In the API, medium is the default thinking level. Antigravity’s model documentation lets users choose Low, Medium or High for each model, and its model selector shows Carbon and Argon with High beside the name; the page does not state a default. An Antigravity session run at High can spend noticeably more tokens per turn than an API integration on default settings.

How /swarm works, according to its documentation

The Antigravity announcement, dated 10 November, says /swarm “is rolling out today in preview across the Antigravity App and CLI.” The /swarm documentation says “Swarm is in preview” and lists the Google AI Pro and Google AI Ultra plans.

The mechanism has four roles. The user’s current conversation becomes the coordinator: it splits the task, adds named workers and reports back, but does not do the technical work itself. Each worker gets a clean context and a brief that states what it owns, what it must not touch and what done looks like; pools of similar workers can be numbered, such as @fix-1 to @fix-4. The coordinator “usually adds a @verifier for every two to four workers”; verifiers read diffs, run tests and report problems back to the worker that wrote the code. A monitor stays asleep and posts only if every member goes quiet before the work is done.

Coordination runs through a shared board. Assignments, questions, diff summaries and test results are posted with @name, @group or @all mentions, and workers do not message each other privately, so the history stays in one ordered place. Workers share the user’s workspace by default, though the coordinator can give each its own working copy of the repository; the blog post describes these as isolated git worktrees. Removing a worker deletes its working copy, so the coordinator has the worker land its changes first. A killed worker cannot be revived.

The blog post presents a model split: Carbon in the coordinator, architecture, cybersecurity auditing and adversarial verification roles, and lighter, faster models for parallel file edits and routine implementation. The documentation is less specific. It documents no dedicated syntax for pinning a role to a model: workers default to the conversation’s model, and a user who wants a verifier on a different model has to say so in the prompt. The requested model must be selectable in the picker, and a waiting worker can later be moved to another model while keeping its conversation.

Quota is the constraint

Both pages are blunt about cost. The blog says “parallel multi-agent systems consume quota much faster than a standard chat”; the documentation says “a swarm uses quota much faster than a single conversation.” Waiting members “use no tokens”, so idle workers need not be stopped. On a quota error, a member pauses rather than retrying, and it resumes only after a new board post, a model switch or a message from the user. It does not restart when quota refreshes.

For a subscriber, that quota is the plan’s. Antigravity’s plans page says Ultra and Pro quotas refresh every five hours up to a weekly limit, that “more complex tasks consume more quota”, and that Pro and Ultra users can buy AI credits consumed “at standard Gemini Enterprise consumption pricing.” An AI Credit Overages setting decides whether credits are spent automatically once the baseline quota runs out. The plans page does not give quota numbers or say how the 1 December API price change affects credit consumption.

A plan mismatch is worth checking before relying on the blog’s Carbon-coordinator design. The /swarm page lists Pro and Ultra; the models page lists Carbon for Ultra and Enterprise only, and Argon for paid Pro, Ultra and Enterprise. A Pro subscriber can run a swarm, but not a Carbon-led one.

Consider a hypothetical swarm for a refactor, following the documented ratios: one coordinator, four workers, two verifiers (one per two workers) and a monitor. That is up to seven active contexts on one task instead of one. The documentation gives no maximum number of agents and no per-swarm quota figures, so the only safe planning assumption is the documented one: the swarm drains the plan faster, and a quota error stops members until someone intervenes. For teams designing their own orchestration, our explainer on agent harnesses covers why the controller around the model, here the coordinator and board, shapes results as much as the model choice.

What changes for security work

Carbon’s launch was framed around cybersecurity, and the public release does not give public users the same security capability Fairwind partners have. The Fairwind Program page describes “exclusive access to Gemini for cyber models”, limited to internal security, incident response or penetration testing teams, with no resale. The Safety & Security Report says that in Fairwind, misuse mitigations “are tailored to permit authorized users to perform a range of specialized cybersecurity tasks”, such as validating exploitability, threat simulation, reverse engineering and malware analysis. Outside that programme, the report’s general cyber policy says the model will fulfil requests such as analysing known public exploits or discussing defensive coding practices, “but will hard-punt on requests to write functional, weaponized exploits.”

The headline cyber figures come from that gated setting. The model card says cybersecurity benchmark results “reflect performance when accessed through the Fairwind Program.” The 28 October post reports 88.2 per cent on SEC-bench Pro, 54.8 per cent on ExploitGym and 71.0 per cent on CWE-bench. The safety report gives conditions for two: CWE-bench v1 is 120 tasks across 73 CWEs and six language families, with no prior knowledge of where the vulnerabilities are, where Carbon reached 71.0 per cent pass@1 against 68.0 per cent for Argon; ExploitGym’s 54.8 per cent causal pass rate is a 12.8-point improvement over Argon. The methodology document says SEC-bench Pro and ExploitGym were self-computed with the Antigravity harness. A security team using public Carbon should treat these as an upper bound on what it will see.

The safety figures, with their conditions

The safety report measures the model without out-of-model production safeguards and without search grounding, “to measure intrinsic model behavior”, unless otherwise noted. Its most specific figures, all Google-reported:

• Adaptive red-teaming. Across 18 rollouts covering six vulnerabilities and network-intrusion scenarios, each with up to 40 attempts of up to 300 tool-use turns, Carbon blocked 99.0 per cent of adversarial trajectories, a 0.8 per cent per-attempt breach rate against 3.1 per cent for Argon and 13.8 per cent for Gemini 3.8 Flash. Over the full 40-attempt budget, cumulative rollout breach rates were 27.8 per cent for Carbon against 72.2 per cent for Argon.
• Automated attacks. In Gray Swan’s automated red-teaming across 25 high-severity offensive cyber tasks, with 200 attack attempts per seed prompt, per-query success fell to 2.17 per cent from 9.86 per cent on Argon and 14 per cent on 3.8 Flash; per-objective success fell to 40 per cent from 56 and 76 per cent.
• Indirect prompt injection. On Gray Swan’s external benchmark, Carbon had a 0 per cent attack success rate, against Argon’s earlier 0.7 per cent. Google adds that “prompt injection attacks on Carbon are still possible.”
• Agentic honesty. On 375 tasks across five categories, each run three times, with a planted problem in the environment, Carbon passed 92 per cent against 70 per cent for Argon and 45 per cent for Gemini 3.8 Flash. The 10 November post repeats the Carbon and Argon figures.

The per-objective figure deserves more attention than the per-query one. It says that a persistent automated attacker still achieved 40 per cent of its objectives against Carbon. Applications that give Carbon tools should keep their own permission checks.

The frontier-safety section contains a finding that sits awkwardly with the marketing. Under its Material Capability Change assessment, Google says Carbon “did not show meaningful new capabilities or material increases in performance compared with Gemini 4 Argon”, so a full critical capability assessment was not required and Argon’s determinations carry forward. Those include that Google “cannot rule out” the model, without out-of-model mitigations, reaching the Cyber Uplift Level 1 and biological CBRN Uplift Level 1 critical capability levels. The assessment concerns thresholds for severe risk, not everyday usefulness, so it does not contradict Google’s benchmark claims. It does mean Google’s own risk process treats Carbon as a variant of Argon rather than a step change.

What remains unknown

Several facts a budget owner would want are not in the documents we read.

The cutover mechanics. The pricing page says the lower rates apply “through November 30, 2026” and that prices double on 1 December. We did not find a stated time zone for the boundary, or a statement on whether a Batch job submitted on 30 November and completed on 1 December, or a long background generation that spans midnight, is billed at the old or new rate. The Batch documentation does not address jobs that span a price change. Teams with large batch runs should plan to finish them before the boundary, remembering the 24-hour target and 48-hour expiry.

Other surfaces. The schedule we verified is the Gemini Developer API’s. We did not check Gemini for business or Vertex AI pricing, Antigravity credit rates, or whether Google AI Ultra limits change on 1 December.

Real-workload token efficiency. The 1.8× output-token ratio comes from one index at one setting. Google’s 2× cost-per-task claims compare Carbon with unnamed external models, not with Argon. The decisive number for most teams, Carbon’s success rate per dollar against Argon’s on their own tasks, has no public source.

Rate limits. Google does not publish requests-per-minute or tokens-per-minute limits for Carbon on its rate limits page.

Swarm limits. The documentation gives no maximum swarm size, no quota figures and no model-pinning syntax.

Public versus Fairwind behaviour. Google describes different mitigations but does not publish public-access scores on its cyber benchmarks.

Several kinds of evidence would change this article’s conclusions. An extension of the introductory period, or a different price on Vertex AI, would change the budget. Independent per-task cost comparisons of Carbon and Argon at matched thinking levels would settle whether Carbon is worth its extra tokens outside benchmark suites. A new Artificial Analysis run at medium, or a Fable 5.2 result on CursorBench at a lower effort, could change the cost-adjusted ranking in either direction. Published /swarm quota numbers would turn a qualitative warning into a cost line.

What to do before 1 December

1. Measure on your own tasks now, while it is cheaper. Run a fixed sample through Carbon at medium and high and through Argon at the same levels, and record success, output tokens (including thinking), cached-input share and wall-clock time. This is a proposed test, not one we have run. Nineteen days at half price is the cheapest window you will get for that evaluation.
2. Price the result at December rates. Use $4, $0.20 and $20 per million, not the current rates. Artificial Analysis already prices both models at $4 and $20.
3. Move latency-tolerant work to Batch or Flex on 1 December. Both halve every rate. Flex needs retries and long timeouts; Batch needs your tier’s enqueued-token headroom.
4. Check the cache. If your stable context stops hitting the implicit cache, December input costs can matter more than the output rate.
5. Check grounding volume against the 1,500-per-day cap before designing a grounded agent that calls Search on every request.
6. Treat /swarm as a preview with a quota bill. Confirm which models your plan can run, give verifiers explicit model instructions, and watch for paused members after quota errors.
7. Keep Argon as the default where it already works. Google says so itself. At identical per-token prices, Carbon has to earn its place through a higher success rate per dollar, task by task.

Carbon arrives with credible independent evidence on two agentic leaderboards and a price that matches its predecessor. The Intelligence Index lead it launched with did not survive the next Anthropic release. For builders, neither ranking is the deciding fact. The deciding fact is that on 1 December the same tokens cost twice as much, and Carbon is designed to use more of them.