> ## Content Index
> Fetch the complete content index at: https://www.frontiermodels.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# Google opens Gemini 4 Argon to paid users and developers
- URL: https://www.frontiermodels.com/news/google-opens-gemini-4-argon-to-paid-users-and-developers/
- Published: 2026-10-14T15:15:35.000Z
- Updated: 2026-10-14T15:15:35.000Z
- Description: A request using 200,000 uncached input tokens and 40,000 billed output tokens costs $0.80 on Argon, $1.60 on Opus 5.5 and $4 on Astra. Google’s launch rates end November 30.
- Author: Frontier Models Desk
- Tags: News, models-research, google, gemini, anthropic, claude, openai, gpt

Google [opened Gemini 4 Argon](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-argon?ref=frontiermodels.com) to paid consumer, developer and enterprise channels on October 13 across supported markets. The release moves Google’s flagship model beyond its [September 30 rollout to selected cybersecurity defenders](https://www.frontiermodels.com/news/google-gives-gemini-4-argon-to-selected-cyber-defenders-first). Access still depends on the product and plan, and the discounted API rates expire November 30.

At those introductory rates, a request using 200,000 uncached input tokens and 40,000 billed output tokens costs **$0.80** on Argon, compared with $1.60 on Claude Opus 5.5 and $4 on GPT-6 Astra. Tokens are the units used to meter what a model reads and generates, including billed reasoning. Argon’s price for the same usage doubles to $1.60 on December 1.

Independent testing shows why that comparison is only a starting point. On the [Vals Index](https://www.vals.ai/benchmarks/vals%5Findex?ref=frontiermodels.com), a suite of professional tasks weighted by their sectors’ share of U.S. GDP, Argon leads with 68.90% against Opus 5.5’s 66.97%. On [Terminal-Bench 4.0](https://www.vals.ai/benchmarks/terminal-bench-4?ref=frontiermodels.com), which tests agents working through computer-terminal tasks, the order reverses: Argon scores 57.58% against Opus’s 65.15%.

Argon’s edge reverses on terminal-agent work

Vals puts Argon slightly ahead of Opus 5.5 on its broad index but behind on Terminal-Bench 4.0, showing why models should be compared on the work they will actually perform.

Vals Ind…Terminal…Vals benchmark0.0010.0020.0030.0040.0050.0060.0070.00Score (percent)Gemini 4 ArgonClaude Opus 5… 

__Data for Argon’s edge reverses on terminal-agent work__
| Vals benchmark     | Gemini 4 Argon | Claude Opus 5.5 |
| ------------------ | -------------- | --------------- |
| Vals Index v2.1    | 68.90% \[1\]   | 66.97% \[1\]    |
| Terminal-Bench 4.0 | 57.58% \[2\]   | 65.15% \[2\]    |

**Method:** Plots Vals’ published model scores. Vals Index v2.1 is its GDP-weighted professional-work suite; Terminal-Bench 4.0 uses the Mini-SWE-agent harness. Compare models within each benchmark rather than raw scores across the two suites. Twenty-two of 198 Opus Terminal-Bench attempts used fallback models; Argon had no fallbacks. **As of:** 2026-10-14\. **Comparison:** Gemini 4 Argon and Claude Opus 5.5 evaluated by Vals on the published Vals Index v2.1 and Terminal-Bench 4.0 configurations. (percent score; Leaderboard snapshots retrieved October 14, 2026).

**Sources:**
1. \[1\] [Vals Index v2.1 leaderboard and methodology](https://www.vals.ai/benchmarks/vals%5Findex?ref=frontiermodels.com) (retrieved 2026-10-14)
2. \[2\] [Vals Terminal-Bench 4.0 leaderboard and methodology](https://www.vals.ai/benchmarks/terminal-bench-4?ref=frontiermodels.com) (retrieved 2026-10-14)

Those results have important limits. Most Vals Index component tasks are private, restricting outside reproduction. Some Opus attempts ran on provider fallback models, while Argon had no fallbacks. On Terminal-Bench, 22 of 198 Opus attempts used fallbacks. Counting those as failures lowers its score to 58.08%, much closer to Argon. The published results therefore do not cleanly isolate Opus’s performance.

The release does not provide one subscription that unlocks Argon everywhere. Google sets separate requirements for its assistant, development platform, API and business products.

| Where to use Argon                                       | Launch eligibility                                                                     |
| -------------------------------------------------------- | -------------------------------------------------------------------------------------- |
| Gemini app and AI Mode in Search                         | Google AI Pro or Ultra subscribers; Free and Plus are excluded.                        |
| Google Antigravity, its agentic development platform     | Ultra subscribers and eligible enterprises; Free, Plus and Pro are excluded.           |
| Gemini API, for developers connecting models to software | Paid API tier, billed separately. A Pro subscription does not provide API entitlement. |
| Gemini for business                                      | Enterprise access, with eligibility and pricing dependent on the contract.             |

That split matters to existing Google subscribers. Pro includes Argon in the Gemini app and AI Mode, but not in Antigravity, and it does not cover separately billed API use. Google’s [Antigravity model table](https://antigravity.google/docs/models?ref=frontiermodels.com) and [API documentation](https://ai.google.dev/gemini-api/docs/latest-model?ref=frontiermodels.com) spell out those separate routes. The staged global rollout also did not promise immediate access to every eligible account, and regional restrictions still apply.

For developers, the introductory API rates are $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 on December 1\. Output charges include thinking tokens, the model’s billed reasoning before its visible answer. Reused input can receive a 95% discount when automatic caching finds a match, but user-created caches are unsupported.

The $0.80 example applies identical token consumption across all three services, using [Anthropic’s $4 input and $20 output rates](https://platform.claude.com/docs/en/models/opus-5-5/whats-new-opus-5-5?ref=frontiermodels.com) and [OpenAI’s $10 and $50 rates](https://developers.openai.com/api/docs/models/gpt-6-astra?ref=frontiermodels.com). The request fits each model’s context window and remains below Astra’s long-context pricing threshold. It excludes caching, batch discounts, tools, retries and regional processing charges. Equal token counts do not establish equal quality or completed work.

Vals adds a workload-level view because it records actual token consumption during its tests. At Argon’s post-promotion $4 and $20 rates, it reports $15.68 per Vals Index test for Argon against $32.14 for Opus. On Terminal-Bench, Argon costs $17.64 per test against Opus’s $13.20, alongside its lower published pass rate. At the promotional rates, unchanged Argon token use would cut that Terminal-Bench cost to $8.82\. That is arithmetic, not a new test result, and neither figure includes the human effort needed to inspect or repair the work.

Google also reports broader capability and safety findings in its [model card](https://deepmind.google/models/model-cards/gemini-4-argon?ref=frontiermodels.com). The company says deployment safeguards reduced the assessed residual risk after its evaluations could not rule out that Argon had reached biological and cybersecurity capability thresholds. Those are Google’s assessments, not independent safety findings, and Vals’s task benchmarks do not establish whether the model is safe.

The launch discount makes Argon unusually cheap to test, but not automatically cheap to adopt. A lasting switch needs to make sense at December’s prices after reasoning consumption, failed attempts and review time are included. The strongest case for Argon will be a job it completes well at a lower total cost.