RELEASED · PAID ACCESS OPEN
Gemini 4 Argon.
Built for complex coding, research, and long-running agents.
Announced September 30, with broader access from October 13. Explore official configuration, prompt workflows, benchmarks, and community feedback alongside the original release log.
Start with the configuration.
These specifications come from the Gemini API documentation. Client quotas, features, and output modes may differ.
- Stable model ID
gemini-4-argon- Shared input / output context
- 1,048,576 tokens
- Default synchronous output limit
- 65,536 tokens
- Long decode output limit
- Up to 1,048,576 tokens, within the shared context limit
- Input → output
- Text, images, video, audio, PDF → text
- Thinking levels
low / medium / high
Supports function calling, structured outputs, code execution, and search grounding. Caching is implicit only. Native image generation, audio generation, and Live API are not supported.
API pricing
USD per million tokens. Introductory pricing through November 30, 2026.
| Period | Input | Cached input | Output |
|---|---|---|---|
| Introductory | $2 | $0.10 | $10 |
| From December 1 | $4 | $0.20 | $20 |
Cached prices are calculated at 95% off input rates. API token pricing is separate from app subscriptions. Check current pricing for tools and consumption modes.
Pricing and end-date source: Google CloudDifferent tests answer different questions.
Checked October 14, 2026. These scores use different tasks and environments; they cannot be added together or treated as one capability ranking.
AA Intelligence Index
Artificial Analysis composite index. Argon output speed is unavailable on the page; the index does not predict every task.
Artificial Analysis Platform evaluation · Devin60.5%FrontierCode 1.1
Cognition reports code quality and mergeability in Devin. Reasoning effort and the complete harness are not disclosed.
Cognition Google-reported · software engineering77.9%DeepSWE v1.1
Reported on the vendor launch page. Tasks, tools, and retry conditions are incomplete; this is not a universal repository success rate.
GoogleSourced workflows and evaluations.
The prompt library includes workflows and configuration guides. Evaluations distinguish vendor reports, independent measurements, and community experience; these are not Tabbit-run tests.
Prompts and workflows
View all 11Gemini 4 Argon: Google Cloud Enterprise Agent Access Workflow
Google Cloud outlines Argon enterprise access and long-running tasks: choose an entry point, fix the model ID, then configure inputs, result verification, and confirmation for irreversible actions.
Gemini 4 Argon: Robotics Reasoning and Manipulation Workflows
Google demonstrates Argon workflows for robot control, Real2Sim, physical understanding, and self-improvement through observation, execution, verification, and correction—not a general hardware safety protocol.
Gemini 4 Argon: Google Gemini API Shared Context and Model Capability Configuration
gemini-4-argon shares 1,048,576 context tokens. Synchronous output is capped at 65,536; Long Decode can use the remaining context. Multimodal inputs produce text output.
Gemini 4 Argon: A Multistep Deliverables Workflow in the Gemini App
Google Gemini’s official posts outline stepwise Argon deliverables: web apps, 3D files, slides, and reports. No complete prompt or output files are supplied; this is a product demonstration, not an independent review.
Gemini 4 Argon: Long Decode continuation protocol and recovery workflow
Argon Long Decode continues in 64K-token background hops. It requires background: true and an output budget above 65,536. Input, thinking, and output share context; every hop is billed.
Gemini 4 Argon: Google's Unattended Workflow for Optimizing a 205,000-Line ADK Codebase
Google reports an 11-hour Argon run in Antigravity /goal on ADK; human review merged only 2 of 6 proposals. Repository rules, benchmarks, and review matter. This is not a general performance guarantee.
Evaluations and latest news
View all 11Gemini 4 Argon: Google’s Official Release, Access Scope, and Capability Boundaries
Google’s launch page records Gemini 4 Argon’s positioning, initially restricted access, and capability benchmarks. Scores are vendor-reported, with incomplete conditions for independent reproduction.
Gemini 4 Argon Model Card: Safety Evaluation Methods, Threat Models, and Limitations
Google's model card covers Argon's automated safety tests, red teaming, and critical capability risks. Relevant risk thresholds cannot be ruled out; deployment mitigations remain necessary. Test sets, prompts, and most threshold details are not public.
Gemini 4 Argon (High): Artificial Analysis Independent Benchmarks, Reliability, and Cost
Artificial Analysis records Argon (High) at 53 on Intelligence Index and 42 on AA-Omniscience. Output speed is unavailable; this snapshot does not establish general latency or all-task performance.
Gemini 4 Argon: Cognition's FrontierCode Coding Test and Devin Environment
Cognition reports 60.5% on FrontierCode 1.1 in Devin. Chart version labels differ, and reasoning effort and the full harness are undisclosed. This is a platform-reported result.
Gemini 4 Argon: Vals AI Professional-Task Benchmarks and Evaluation Methodology
Vals records Index 68.90% ± 0.97 and Finance Agent v2 65.40% ± 0.32 in the October 14 snapshot. Ranking denominators differ; results across tasks cannot be combined into one leaderboard.
Early feedback from X and Reddit
These public posts contain individual cases or opinions with incomplete reproduction conditions, not independent performance conclusions.
Reddit Early Feedback: Gemini 4 Argon Task Experience and Limitations
A first-day Reddit comment discusses speed, high-tier token use, and medium as a compromise. No task, prompt, or billing record is supplied; this is independently unverified personal opinion.
AIHubMix X case: Gemini 4 Argon vs. GPT 6 Astra on cost and elapsed time for the same prompt
AIHubMix self-reports Astra at 25 minutes/$11 and Argon at 3 minutes/$0.50 on one task. No prompt, tier, or billing conditions are disclosed; this is not an independent speed or cost evaluation.
From rumors to broader access
The original rumor and media log is preserved with official access updates added. Early entries without primary links were not independently verified in this update and do not establish specifications.
The “argon” codename surfaces
A post shared output allegedly from a checkpoint codenamed “argon,” claiming a 256K output cap and 2M context. Google never confirmed it, and the official specs later did not match those numbers.
A disguised Arena test theory
An anonymous Arena model labeled “gemini-3.8-flash” performed unusually well, fueling guesses that Gemini 4 was being tested undercover. The label also matches a released model, so it proves nothing.
Official: the codename becomes the product
Google announced the first Gemini 4 frontier model under the Argon name. Initial access was limited to trusted Fairwind Program defenders, with up to 1M-token long decode and $2 / $10 introductory pricing. The earlier 2M-context rumor was not an official specification.
An internal “Carbon” checkpoint and rollout hints
Reports describe an internal checkpoint codenamed “Carbon” being tested on an internal coding platform, with employees rating its coding near Claude Opus 5.5. Around the same time, Argon was spotted in an Antigravity Insider build’s model selector with 256K / 512K / 900K context options.
Broader access: API, Gemini app, and Antigravity
Paid Gemini API and enterprise access are open. The Gemini app and Search AI Mode are rolling out to Pro / Ultra subscribers; Antigravity access is for Ultra. App Pro access does not imply Antigravity Pro access.
Stable API: shared context and long decode
The catalog lists stable gemini-4-argon. Input and output share a 1,048,576-token limit; default synchronous output is 65,536, with long decode extending output up to 1,048,576. Multimodal input is supported, with text output.
Choose an entry point and check access.
As of October 14, 2026, Gemini 4 Argon is listed in the official API catalog. Paid API access, Pro / Ultra in the Gemini app, and Ultra in Antigravity are separate access paths. Check your current Tabbit model picker for availability; inclusion in this library does not establish Tabbit integration.