Gemini 4 Argon Benchmarks: Google's Narrow Lead
Google's flagship scores 68.9 percent on the Vals Index and ships today to AI Pro and Ultra subscribers, plus every GitHub Copilot paid tier.
In short
Gemini 4 Argon is Google's new flagship model, topping the company's own benchmark table with 68.9 percent on the Vals Index ahead of Claude Opus 5.5 at 67.0 percent and GPT-6 Astra at 63.1 percent, and it reached GitHub Copilot on October 13, 2026.
At a glance
- Vals Index: 68.9 percent for Argon vs. 67.0 percent for Claude Opus 5.5 and 63.1 percent for GPT-6 Astra, per Google DeepMind.
- DeepSWE v1.1 agentic coding: 77.9 percent for Argon, 74.2 percent for Claude Opus 5.5, 74.1 percent for GPT-6 Astra.
- CWE-bench v1: 68.0 percent, a tie with GPT-6 Astra rather than a lead on cybersecurity.
- Access: Google AI Pro and Ultra; Copilot Pro+, Max, Business and Enterprise, with a gradual rollout per GitHub.
- Copilot surfaces: VS Code, Visual Studio, JetBrains, Xcode, Eclipse, Copilot CLI, coding agent, Copilot app.
Gemini 4 Argon is Google DeepMind's new flagship, and the company's own comparison table puts it in front: 68.9 percent on the Vals Index against 67.0 percent for Claude Opus 5.5 and 63.1 percent for GPT-6 Astra. GitHub switched the model on inside Copilot on October 13, 2026. Every figure here is vendor-reported.
A 1.9-point lead
On the Vals Index the gap to Claude Opus 5.5 is 1.9 points, real but thin. Agentic coding is where Argon pulls away: 77.9 percent on DeepSWE v1.1 versus 74.2 and 74.1 percent, a 3.7-point margin over the nearest rival. On security the table shows no winner at all, with Argon and GPT-6 Astra both at 68.0 percent on CWE-bench v1.
Two further scores arrive with no rival column: 84.2 percent on GraphWalks across the 256k to 1M token range and 91.7 percent on the LVBench video test. The product page lists a window of 1 million tokens.
Turning raw video into simulation
DeepMind's physical-world pitch is robotics: Argon is said to convert raw video into a physics-accurate simulation on its own and to acquire physical skills by watching. On the coding side the company cites a 2.85x speedup on quantum systems, with no method disclosed. Developer forums pass around a Gemini 4 Argon for Robotics variant; the product page lists no such thing and treats Gemini Robotics as a separate model line.
Who can use it today
DeepMind says Argon is live for Google AI Pro and Ultra subscribers, reachable through the Gemini app and Google AI Studio. Security researchers and defenders can apply for six months of free access. The Fairwind Program is opening to several hundred more partners at the same time.
In GitHub Copilot the model covers Pro+, Max, Business and Enterprise across eight surfaces, from VS Code and JetBrains to Xcode and the Copilot CLI. GitHub reports its largest gains in the CLI and describes the rollout as gradual. New models turn on by default unless an administrator disables them in the model policy, and usage is billed at the provider's list pricing.
What we could not confirm
A third report on the launch would not load before publication, so this article rests on the two vendor sources. Nobody has independently reproduced the benchmark table. There is also no API price list and no published premium-request multiplier for Copilot; GitHub points only to the provider's rates.
FAQ
Is Gemini 4 Argon better than Claude Opus 5.5?
By Google's own table, yes, but narrowly: 68.9 versus 67.0 percent on the Vals Index and 77.9 versus 74.2 percent on DeepSWE v1.1. On CWE-bench v1 both Argon and GPT-6 Astra sit at 68.0 percent. No independent re-run exists yet.
How do I get access to Gemini 4 Argon?
A Google AI Pro or Ultra subscription covers it, including the Gemini app and Google AI Studio. Developers also reach the model through GitHub Copilot Pro+, Max, Business and Enterprise.
What does Gemini 4 Argon cost in GitHub Copilot?
GitHub bills usage at the provider's list pricing and has not published a premium-request multiplier. Admins gate access through the model policy, and otherwise new models are enabled by default.
