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.
Suitable for judging: Whether the model is available, which official entry points can be used, which tasks Google claims it covers, and the benchmark results published in the release materials.
Not suitable for extrapolating: General production success rate, actual cost, stability, availability in third-party clients, performance on Chinese-language tasks, or independent conclusions about safety defenses.
Applicable model version: Gemini 4 Argon; the announcement does not provide an API model ID or snapshot string that other providers could verify.
Test environment or client: The October 13 update explicitly mentions the paid Gemini API, Google AI Studio for Google AI Ultra subscribers, and Google Antigravity; no other clients are confirmed by the official pages.
Reasoning level and parameters: Not specified. The announcement does not disclose temperature, reasoning level, tool configuration, system prompt, retry policy, or API snapshot.
The primary source is Google’s official release announcement from September 30, 2026; the October 13 page is used only to verify the related access update. The primary source consists of vendor release information and vendor-reported benchmark results, not an independent evaluation. The page describes only each benchmark’s task direction and score; it does not disclose the complete task set, sample size, number of runs, inputs and outputs, evaluators, random seeds, model snapshot, failure handling, or a unified harness. Claims such as “leading,” “ranked first,” and “tied for first” should therefore remain explicitly vendor claims.
flowchart LR
A[2026-09-30 release announcement] --> B[Fairwind trusted cyber defenders]
B --> C[Gradual access expansion and stronger safeguards]
C --> D[2026-10-13 developer access]
D --> E[Paid Gemini API]
D --> F[Google AI Ultra → AI Studio / Antigravity]| Date | Officially verifiable status | Boundary |
|---|---|---|
| 2026-09-30 | Opened to a group of trusted cyber defenders in the Fairwind Program | Still part of the U.S. government’s voluntary pre-release model-access process; Google said it would continue iterating on safeguards before expanding access |
| 2026-09-30 | Planned expansion to developers, enterprises, and consumers | The announcement says only “as soon as possible” and states that the rollout would start with paid API customers and Google AI Ultra subscribers |
| 2026-10-13 | Developers can start building | Use the paid tier of the Gemini API; Google AI Ultra subscribers can try it in Google AI Studio and Google Antigravity |
Introductory pricing: $2 / 1M tokens for input and $10 / 1M tokens for output; cached input is 95% off the input price, which works out to approximately $0.10 / 1M tokens at the announced introductory price.
Price after the introductory period: The footnote states $4 / 1M tokens for input and $20 / 1M tokens for output. The page does not state when the introductory period ends.
Output limit: 1M tokens, which the announcement says is increased from the previous 64K tokens to an industry-leading level. This is an output-token limit; it should not be treated as evidence of an input-context limit.
| Task or benchmark | Result given on Google’s page | Supported judgment | Not disclosed or cannot be extrapolated |
|---|---|---|---|
| Software engineering / DeepSWE v1.1 | 77.9%, described by the page as state of the art | Supports the vendor-reported score for this version and benchmark setup | Not the success rate for all code repositories; the full harness and sample details are undisclosed |
| Enterprise knowledge work / Vals Index | Described as leading | Supports only Google’s directional claim | No aggregate score, version, weighting, execution details, or raw results |
| Vals Finance Agent v2 | Described as leading | Supports Google’s claimed advantage in multi-step financial research | No score, task set, tools, or evaluators |
| Harvey’s Legal Agent Benchmark | Described as leading | Supports Google’s claimed advantage in legal research and drafting | No score, sample, context, or evaluation protocol |
| AutomationBench | 51.3%, described as ranked first | Supports the end-to-end business-function score reported on the page | Not a general enterprise-automation success rate or a cross-harness ranking |
| LVBench | 91.7%, described as state of the art | Supports the long-video-understanding score reported on the page | Not evidence for all video, language, or length combinations |
| CWE-bench v1 | 68%, described as tied for first | Supports the vendor report for this vulnerability-remediation benchmark | Not a real-world vulnerability-remediation rate or evidence for unsupervised security deployment |
| Google’s internal comprehensive vulnerability benchmark | Said to uncover multiple exposures across codebases in 20 programming languages | Records only the direction of the internal test | No public score, task set, or evaluation protocol |
| Wiz’s internal black-box penetration-testing benchmark | Said to outperform Gemini 3.8 Flash Cyber | Records only Google’s claim about its internal comparison | No public sample, environment, score, or reproduction conditions |
Argon targets complex, long-horizon workflows, with emphasis on real-world software engineering, enterprise knowledge work such as legal and finance, cybersecurity defense, and creative writing.
Google says thousands of Googlers are already using Argon for specialized coding, research, and writing. Examples include quantum-algorithm optimization, data-center memory optimization, and C/C++-to-Rust migrations ranging from tens of thousands of lines to more than 800K lines.
In one libgav1 example, Argon agents replaced 32K lines of SIMD code with safe Rust. Google reports a 2.7× speedup over the Rust port with identical video output; this is a Google internal case study, not a public controlled benchmark.
For cybersecurity defense, Google says Argon can autonomously find, validate, and patch critical software vulnerabilities; the release version for trusted defenders and Google’s internal teams does not include cyber guardrails.
Google says it is continuing to strengthen four areas before broad release: preventing cyber and CBRN misuse, resisting indirect prompt injection, monitoring misalignment, and hardening sandbox environments. The page also says Argon is leading on Gray Swan’s Indirect Prompt Injection benchmark, but provides no score or complete experimental conditions.
These points describe Google’s release strategy and safety claims, not independent safety certification. In particular, “without cyber guardrails” applies only to the trusted defenders and Google internal teams described in the announcement; it should not be interpreted to mean that ordinary developer access disables safety protections by default.
Availability changed clearly: September 30 was a restricted rollout for trusted defenders; October 13 explicitly made the model available to developers through the paid Gemini API, Google AI Ultra’s AI Studio, and Antigravity.
Pricing must be time-qualified: $2/$10 applies during the introductory period; the footnote states $4/$20 afterward. The page gives no changeover date and does not state whether different entry points carry other fees.
Long output is not general long-context capability: The official material specifies only a 1M output-token limit, and that is a specification rather than a recommendation that every task should generate such a long trajectory.
The scores are vendor reports: The 77.9%, 51.3%, 91.7%, and 68% figures have no publicly documented complete methodology; claims of leadership and first place cannot be compared directly across evaluators, versions, or harnesses.
No independent testing: This article does not mix in search snippets, third-party leaderboards, comments, or other sources. A production decision still requires testing with a fixed model snapshot, API entry point, task set, and cost definition.
The official pages are insufficient to reproduce the benchmarks completely. A meaningful follow-up evaluation should at minimum fix the entry point (Gemini API / AI Studio / Antigravity), model snapshot, reasoning and sampling parameters, tool permissions, system prompt, task-set version, stopping conditions, retry policy, and evaluator. It should also record input and output tokens, latency, failure types, human takeovers, and cost separately. Do not substitute a page’s claim of leadership for these measurements.
September 30 release announcement: https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/. The page is titled “Gemini 4 Argon: our next era of frontier intelligence” and covers restricted access, pricing, the 1M output limit, benchmark figures, and safety boundaries.
October 13 developer update: https://blog.google/innovation-and-ai/technology/developers-tools/build-with-gemini-4-argon/. The body explicitly states that developers can use the model through the paid Gemini API, and mentions Google AI Ultra, Google AI Studio, and Google Antigravity.
Developer guide provided in Google’s October 13 update: https://aistudio.google.com/learn/gemini-4-argon-developer-guide. This task records the public link only and does not use its contents as additional evaluation evidence.
Gemini 4 Argon