ModelsArchitectures & capability
Google opens Gemini 4 Carbon to paying developers and previews agent 'swarm' in Antigravity
Google made Gemini 4 Carbon, its new top-tier model, available to paying API, enterprise and AI Ultra users on November 10, priced at $2/$10 per million tokens until month-end, and previewed an Antigravity feature that coordinates multiple agents from one prompt.
From the The Verge · AI article of November 10, 2026 and cited sources: Cyber defenders had restricted access to Carbon for two weeks. [4] [5] [6]
SOURCES[4] Google: Gemini 4 Carbon: our next frontier model blog.google[5] Google: Gemini 4 Carbon is now available blog.google[6] 9to5Google: Gemini 4 Carbon rolling out to Gemini app 9to5google.com
AT Pulse research deskPublished · Updated · 5 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- Original source
- The Verge · AI article, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- November 10, 2026 article published
- Covered
- Nov 11 edition
- Evidence
- 9 sources across 8 sites
- 1Google releases Gemini 4 Carbon with agent ‘swarm’ featuretheverge.com
- 2Gemini 4 Carbon is now availabledeepmind.google
- 3Artificial Analysis: Gemini 4 Carbon, new number 1artificialanalysis.ai
- 4Google: Gemini 4 Carbon: our next frontier modelblog.google
- 5Google: Gemini 4 Carbon is now availableblog.google
- 69to5Google: Gemini 4 Carbon rolling out to Gemini app9to5google.com
Reach: not measured yet
Cyber defenders had restricted access to Carbon for two weeks. It then arrived in the Gemini API, AI Studio, Gemini for business, the Gemini app and Antigravity. The introductory price of $2 per million input tokens and $10 per million output tokens doubles to $4 and $20 on November 30. [4] [5] [6]
Artificial Analysis, an independent evaluator, ranks Carbon first on its Intelligence Index with a score of 59. On its coding-agent index, though, Carbon scores 59 and trails Claude Code with Opus 5.5 at 62. Carbon also hallucinates 11% of the time, compared with Argon's 4%, though that is far below the rival Claude and GPT-6 Astra models. [3]
The /swarm command in Antigravity turns one prompt into a team of agents. Carbon takes the hard reasoning roles and faster models work in parallel. Google positions Carbon above Gemini 4 Argon rather than as its replacement, and still recommends Argon for most software engineering and knowledge work. [6] [1]
Code written for older Gemini models may break. Requests that use thinking_budget or the MINIMAL thinking level return HTTP 400 errors. Function-call IDs, names and order must match exactly, and long-decode output requires the v1beta1 API. [7]
Independent rankings put Google back at the frontier of a race that had looked like Anthropic versus OpenAI.
Read the full assessment
For buyers, that suggests sending hard reasoning to Carbon, keeping routine work on Argon, budgeting at full price and grounding fact-sensitive outputs.
Executive brief
Google's newest frontier model, Gemini 4 Carbon, ranked first on Artificial Analysis's Intelligence Index v5 with a score of 59. Its introductory API price is $2/$10 per million input/output tokens, about a fifth of OpenAI's GPT-6 Astra and a quarter of Anthropic's Fable 5.2. On November 10, after two weeks of access limited to cybersecurity defenders, Google opened Carbon to paying developers, enterprises and AI Ultra subscribers. It also began previewing /swarm in Antigravity, which turns one prompt into a team of coordinated agents. The ranking has a catch: the same independent evaluator found Carbon's hallucination rate is higher than its sibling model Argon's.
What changed and event timeline
Gemini 4 Argon launches
Google's other frontier model shipped a month before Carbon's wide release. Google says Argon is still the better choice for most software engineering and knowledge work ().
Carbon announced with restricted access
Cyber defenders and trusted partners got early access through Google's Fairwind Program. Google says the phased rollout followed voluntary pre-release testing by the US government ().
Independent index ranks it first
Artificial Analysis gave Carbon (high reasoning) a score of 59 and noted it was not yet publicly available ().
Anthropic ships Fable 5.2
The Verge reports that on benchmarks Google cited, Carbon trails this rival model on most metrics ().
General availability and swarm preview
Carbon arrived in the Gemini API, AI Studio, Gemini for business, the Gemini app, AI Mode and Antigravity. Gemini 3.1 Pro was removed from the app (;).
Introductory pricing ends
Prices double to $4 per million input tokens and $20 per million output tokens ().
Capabilities and access
- Model IDs:
gemini-4-carbon, plusgemini-4-carbon-cyber-permissivefor cybersecurity workloads (Google Cloud docs). - Who gets it: all paying API and enterprise customers, plus Google AI Ultra subscribers.
- Scientists: Engadget reports 10,000 scientists get six months of free access (Engadget).
Read the full section
- Model IDs:
gemini-4-carbon, plusgemini-4-carbon-cyber-permissivefor cybersecurity workloads (Google Cloud docs). - Who gets it: all paying API and enterprise customers, plus Google AI Ultra subscribers. Google hasn't said when, or whether, cheaper tiers or free users will get access (The Verge). On Vertex, the enterprise listing names verified customers in select regions (docs).
- Scientists: Engadget reports 10,000 scientists get six months of free access (Engadget).
- Swarm: Antigravity's
/swarmlets developers give Carbon the hard reasoning roles and pair it with faster models working in parallel (9to5Google).
Technical analysis for researchers and developers
- Context: one 1,048,576-token limit shared by input and output. A single output segment tops out at 65,536 tokens when streaming.
- Reasoning control: thinking levels are
LOW,MEDIUM(default) andHIGH(default on the cyber variant). - Function calling: call and response IDs, names and order must match exactly.
Read the full section
The developer guide documents:
- Context: one 1,048,576-token limit shared by input and output. A single output segment tops out at 65,536 tokens when streaming. "Long decode" continues output across segments and requires the
v1beta1API. - Reasoning control: thinking levels are
LOW,MEDIUM(default) andHIGH(default on the cyber variant). Sendingthinking_budgetorMINIMALreturns an HTTP 400 error, so code written for older Gemini models will break. - Function calling: call and response IDs, names and order must match exactly.
- Caching: only implicit, and only for shared prompt prefixes of at least 4,096 tokens.
- SDKs: minimum versions include
google-genai2.34.0 or later. - Not documented: architecture details and a model card. The guide doesn't describe swarm.
Claims and evidence
- Cyber benchmarks: Google reports SEC-bench Pro 88.2%, ExploitGym 54.8%, CWE-bench 71.0% and 0.0% attack success on the Gray Swan prompt-injection test (Google).
- Independent results: Intelligence Index 59. On the Coding Agent Index Carbon scores 59, behind Claude Code with Opus 5.5 at 62.
- Partner leaderboards: Google cites first place on CursorBench 4.0, WANDR and AutomationBench (Google). No independent corroboration was found.
Read the full section
- Cyber benchmarks: Google reports SEC-bench Pro 88.2%, ExploitGym 54.8%, CWE-bench 71.0% and 0.0% attack success on the Gray Swan prompt-injection test (Google). These are vendor-reported, were run without guardrails, and no third party has reproduced them (NeuralTrust).
- Independent results: Intelligence Index 59. On the Coding Agent Index Carbon scores 59, behind Claude Code with Opus 5.5 at 62. Carbon averages 123k output tokens per task, and an index task costs $7.76 versus $10.04 for Opus 5.5 (Artificial Analysis).
- Partner leaderboards: Google cites first place on CursorBench 4.0, WANDR and AutomationBench (Google). No independent corroboration was found.
Context and prior work
- Two-tier strategy: Carbon sits above Argon rather than replacing it (9to5Google).
- Competitive position: The Verge reads the launch as Google rejoining a race that had looked like Anthropic versus OpenAI (The Verge).
- Gated release: giving cyber defenders capable models before everyone else is now a recurring pattern. Here it came with US government pre-release testing (Google).
Read the full section
- Two-tier strategy: Carbon sits above Argon rather than replacing it (9to5Google).
- Competitive position: The Verge reads the launch as Google rejoining a race that had looked like Anthropic versus OpenAI (The Verge).
- Gated release: giving cyber defenders capable models before everyone else is now a recurring pattern. Here it came with US government pre-release testing (Google).
Limitations, safety and contested findings
- Hallucination is good or bad depending on the comparison. Carbon's AA-Omniscience score is still the highest at 54 (Artificial Analysis).
- NeuralTrust calls Carbon very verbose, citing 200M output tokens for a full index run (NeuralTrust).
- Safety figures are Google's own. They include an 80% cut in high-severity misalignment alerts and 99% of adaptive red-team trajectories blocked (Google).
Read the full section
- Hallucination is good or bad depending on the comparison. Carbon hallucinates 11% of the time versus Argon's 4%, but rival Claude and GPT-6 Astra models sit at 50%. Carbon's AA-Omniscience score is still the highest at 54 (Artificial Analysis).
- Verbosity numbers conflict. NeuralTrust calls Carbon very verbose, citing 200M output tokens for a full index run (NeuralTrust). Artificial Analysis reports 29% fewer output tokens per task than Opus 5.5.
- Safety figures are Google's own. They include an 80% cut in high-severity misalignment alerts and 99% of adaptive red-team trajectories blocked (Google).
- Benchmark selection: companies can pick the results that flatter them (The Verge).
Business and practitioner implications
- Plan on $4/$20 per million tokens, not the introductory rate.
- Send hard, long-horizon reasoning to Carbon and keep routine engineering work on Argon, as Google recommends.
- In fact-sensitive workflows, Carbon's higher hallucination rate than Argon's argues for retrieval grounding and validation.
Read the full section
- Budget for the new price. Plan on $4/$20 per million tokens, not the introductory rate. A low per-token price doesn't guarantee a low per-task cost, because models use different numbers of tokens for the same job (The Verge).
- Route by task. Send hard, long-horizon reasoning to Carbon and keep routine engineering work on Argon, as Google recommends.
- Check facts where it matters. In fact-sensitive workflows, Carbon's higher hallucination rate than Argon's argues for retrieval grounding and validation.
- Migrate carefully. Test thinking-level settings and function-call symmetry before moving production traffic.
The source trail.
Sources (9)
Google releases Gemini 4 Carbon with agent ‘swarm’ feature
Article text retrieved; extracted text may omit tables or interactive elements.
theverge.comGemini 4 Carbon is now available
Related coverage; assess separately
deepmind.google