# Gemini 4 Argon comes to Antigravity Pro: access, benchmarks and costs

> Argon is rolling out to paid Google AI Pro subscribers in Antigravity with its own weekly quota. Compare reasoning, benchmarks and costs with Sol, Astra and Opus.


Published: 2026-10-16. Original source: [Google Antigravity · Argon for paid Google AI Pro subscribers](https://x.com/antigravity/status/2110866531932316011).

If you pay for Google AI Pro and have been waiting for Argon in Antigravity, there is now an announcement: **Google began rolling out Gemini 4 Argon to paid Pro subscribers on October 15**. The official [Antigravity account confirmed it on X](https://x.com/antigravity/status/2110866531932316011), along with a dedicated weekly quota for the model.

One distinction helps explain the news. Pro subscribers already had Argon in the Gemini app following [the October 13 launch](/en/guides/gemini-4-argon-available-benchmarks/). This update brings it to the agent in Antigravity, where the previous announcement only included Ultra.

[Google Antigravity’s original announcement · October 15, 2026](https://x.com/antigravity/status/2110866531932316011)

> We’re rolling out Gemini 4 Argon in Antigravity to paid Google AI Pro subscribers today.

## Who can use Argon with Google AI Pro?

You need a **paid Pro subscription outside its trial period**. The [Antigravity model documentation](https://antigravity.google/docs/models) explicitly states this condition. Ultra and Enterprise retain access; Free and Google AI Plus do not include Argon.

Check which application you are opening, too. Argon is available in **the Antigravity app, Antigravity CLI and IDE extensions**. The same documentation explicitly excludes **Antigravity IDE**. An IDE extension and the Antigravity IDE are different ways to use the product.

Google describes a rollout. If you have paid Pro and the model has not appeared, the announcement alone does not confirm that your account is enabled yet. The documentation places the model selector below the conversation prompt box.

## How does Argon's weekly quota work?

The [original post](https://x.com/antigravity/status/2110866531932316011) announces a **separate, weekly-only Argon quota** for Pro, on top of the existing Gemini 3.8 Flash allowance. You can reserve Argon for a demanding task and continue using Flash with its own allowance.

The post gives no fixed number of messages or tasks. It also makes no promise of an Argon Pro reset every five hours. That interval appears on the [general plans page](https://antigravity.google/docs/plans); the specific Argon announcement says weekly-only. Check the usage shown in your account before handing it a long job.

## Argon versus Sol 6.1, Astra and Opus 5.5

We start with **AutomationBench**, which checks whether an agent completes work across applications: finding information, updating records or sending something to the right recipient. Zapier scores the final environment against fixed criteria. Its headline result requires every condition of the task to pass.

Each point in the chart is a published configuration. Use the buttons to switch between cost and reasoning, choose a price and compare an effort level. Tap a point to read its figures. The table preserves every value and its source.

### AutomationBench v1.0.6
Strict score (%), average USD per task. List costs; the promotional column only changes Argon. Opus uses default fallbacks. Sol: OpenAI publication, September 29; others: Zapier board reviewed October 15. Separate publications, not a new joint run.
| Model | Effort | Score | USD/task list | USD/task promotion | Source |
| --- | --- | --- | --- | --- | --- |
| Gemini 4 Argon | Medium | 50.08% | $1.54 | $0.77 | [Zapier](https://zapier.com/benchmarks) |
| Gemini 4 Argon | High | 51.29% | $1.7 | $0.85 | [Zapier](https://zapier.com/benchmarks) |
| GPT-6.1 Sol | Low | 24.7% | $0.157 | $0.157 | [OpenAI](https://openai.com/index/introducing-gpt-6-1-sol/) |
| GPT-6.1 Sol | Medium | 31.7% | $0.1917 | $0.1917 | [OpenAI](https://openai.com/index/introducing-gpt-6-1-sol/) |
| GPT-6.1 Sol | High | 33.2% | $0.2255 | $0.2255 | [OpenAI](https://openai.com/index/introducing-gpt-6-1-sol/) |
| GPT-6.1 Sol | Xhigh | 35.5% | $0.2508 | $0.2508 | [OpenAI](https://openai.com/index/introducing-gpt-6-1-sol/) |
| GPT-6.1 Sol | Max | 36.1% | $0.2989 | $0.2989 | [OpenAI](https://openai.com/index/introducing-gpt-6-1-sol/) |
| GPT-6 Astra | Low | 30.29% | $1.08 | $1.08 | [Zapier](https://zapier.com/benchmarks) |
| GPT-6 Astra | Medium | 34.09% | $1.27 | $1.27 | [Zapier](https://zapier.com/benchmarks) |
| GPT-6 Astra | High | 37.14% | $1.44 | $1.44 | [Zapier](https://zapier.com/benchmarks) |
| GPT-6 Astra | Xhigh | 38.96% | $1.5 | $1.5 | [Zapier](https://zapier.com/benchmarks) |
| GPT-6 Astra | Max | 41.4% | $1.73 | $1.73 | [Zapier](https://zapier.com/benchmarks) |
| Claude Opus 5.5 + fallbacks | Low | 24.2% | $0.51 | $0.51 | [Zapier](https://zapier.com/benchmarks) |
| Claude Opus 5.5 + fallbacks | Medium | 29.53% | $0.65 | $0.65 | [Zapier](https://zapier.com/benchmarks) |
| Claude Opus 5.5 + fallbacks | High | 33.03% | $0.71 | $0.71 | [Zapier](https://zapier.com/benchmarks) |
| Claude Opus 5.5 + fallbacks | Xhigh | 35.77% | $0.89 | $0.89 | [Zapier](https://zapier.com/benchmarks) |
| Claude Opus 5.5 + fallbacks | Max | 42.47% | $1.44 | $1.44 | [Zapier](https://zapier.com/benchmarks) |

Lines connect measured settings. They do not estimate intermediate values; no Argon Low/Xhigh/Max reading is published here.

**Argon High reaches 51.29%** on Zapier's board. Moving up from Medium adds **1.21 percentage points**, while the published list cost per attempt rises from US$1.54 to US$1.70. Promotional costs are US$0.77 and US$0.85. This gives you a concrete view of what the extra reasoning buys within that evaluation.

Sol 6.1 Max scores 36.1% at US$0.2989 per attempt in [OpenAI's publication](https://openai.com/index/introducing-gpt-6-1-sol/). Astra Max reaches 41.4% and Opus 5.5 Max reaches 42.47% on [Zapier's board](https://zapier.com/benchmarks). Opus uses fallback models for refused steps, a condition carried alongside its points. These figures come from separate publications; viewing them together does not establish an identical, simultaneous run of all four models.

For your own decision, look at both coordinates. A lower budget per attempt can help, but you also need to consider how many attempts succeed and how much review follows. Your team's task costs will depend on the tokens, tools and retries each job consumes.

## What changes on other benchmarks?

The order changes with the task. Google reports **77.9% for Argon on DeepSWE v1.1**, a software engineering evaluation. On Terminal-Bench 4.0, **Opus 5.5 reaches 66.4%**, compared with Argon's 57.4%. The table keeps each reading in its own context.

| Benchmark | Gemini 4 Argon | GPT-6.1 Sol | GPT-6 Astra | Claude Opus 5.5 |
| --- | --- | --- | --- | --- |
| DeepSWE v1.1 | **77.9%** | 75.22% (High) | 74.1% | 74.2% |
| Terminal-Bench 4.0 | 57.4% | No data | 58.2% | **66.4%** |
| Terminal-Bench Science 0.1 | 57.6% | 57.02% (Max) | **68.1%** | 63.3% |

Google: maximum reasoning or best available result per benchmark. Sol: High on DeepSWE, Max on Science, per OpenAI. Argon Science uses 6× verifier timeout. Different sources and harnesses; no Argon cost per task published for these tests.
[Google](https://deepmind.google/models/model-cards/gemini-4-argon/) · [OpenAI](https://openai.com/index/introducing-gpt-6-1-sol/) · [Methodology](https://deepmind.google/models/evals-methodology/gemini-4-argon/)

For Science, Argon's methodology increases the verifier timeout sixfold to address timeout issues. That belongs alongside comparisons with other runs. We also leave Sol's Terminal-Bench 4.0 result empty: the sources used here do not publish it.

If you want to build a page, review a repository or research a piece of content, try the models on a task you know and keep the acceptance criteria consistent. These benchmarks evaluate specific workloads. To judge your brand's voice, you will need to read the answers the models produce.

## How much does Gemini 4 Argon cost through the API?

Pro access applies to Antigravity. **The Gemini API has separate billing**: Argon costs US$2 per million input tokens and US$10 per million output tokens through **November 30, 2026**. From December 1, the announced rates are US$4 and US$20. Billed output includes reasoning tokens, according to [Google's pricing page](https://ai.google.dev/gemini-api/docs/pricing).

Compare a specific token volume below using the four models' standard API rates, using short context for OpenAI. Changing the token counts updates the estimate; a token price alone does not determine the cost of completing a task.

| USD per million tokens | Input | Output | Input from Dec 1 | Output from Dec 1 |
| --- | --- | --- | --- | --- |
| [Gemini 4 Argon](https://ai.google.dev/gemini-api/docs/pricing) | $2 | $10 | $4 | $20 |
| [GPT-6.1 Sol](https://developers.openai.com/api/docs/models/gpt-6.1-sol) | $2 | $10 | $2 | $10 |
| [GPT-6 Astra](https://developers.openai.com/api/docs/models/gpt-6-astra) | $10 | $50 | $10 | $50 |
| [Claude Opus 5.5](https://www.anthropic.com/claude-opus-5-5) | $4 | $20 | $4 | $20 |

Illustrative 10,000 input and 2,000 output tokens including reasoning: Argon $0.04 ($0.08 from December 1), Sol $0.04, Astra $0.20, Opus $0.08. Formula: (input × input rate + output × output rate) / 1,000,000. Excludes caching, tools, taxes and fast modes. API billing is separate from Pro. Calculator limited to 272,000 input and 64,000 output; does not model long context.

[Argon supports `low`, `medium` and `high` thinking levels through the API](https://ai.google.dev/gemini-api/docs/latest-model), with `medium` as the default. Increasing the level can change consumption. Google recommends testing all three on your workload; a model card's maximum-effort result does not describe every everyday request.

If you already pay for Pro, start by checking for Argon in a supported application and looking at its quota. Give it a bounded task you can evaluate. That will help you decide whether to reserve it for longer work, keep Flash for quick edits or compare another option using your own materials.

## Plan your next post in HeyMark.

Keep the idea, review the draft with your team, and see how it performed in the accounts you connected.

[Start free](https://app.heymark.ai) · [See how it works](https://heymark.ai/en/#product)

## Developer and agent resources

- [HeyMark MCP documentation](https://heymark.ai/en/mcp/)
- [llms.txt](https://heymark.ai/llms.txt)
- [Full site content for language models](https://heymark.ai/llms-full.txt)
- MCP protocol endpoint: `POST https://mcp.heymark.ai`
- OAuth protected-resource metadata: [/.well-known/oauth-protected-resource](https://mcp.heymark.ai/.well-known/oauth-protected-resource)
- MCP server card: [/server-card](https://mcp.heymark.ai/server-card)