GitHub Rolls Out Gemini 4 Argon to Tackle Long Agent Coding Runs

GitHub has begun rolling out Gemini 4 Argon, Google’s frontier model, across GitHub Copilot. The model targets long agent sessions, multi-file refactors, and engineering tasks that require planning, execution, testing, and revision. In GitHub’s early tests, Argon completed more tasks than previous Gemini models across every coding benchmark, with its largest gains in Copilot CLI. GitHub also reports that the model checks edge cases before finishing, which could reduce premature completion during extended agent runs.

Who gets Argon, and where

GitHub is gradually enabling Argon for Copilot Pro+, Max, Business, and Enterprise customers. Eligible users can select it in the following products:

• Visual Studio Code and Visual Studio
• JetBrains IDEs, Xcode, and Eclipse
• Copilot CLI
• GitHub Copilot coding agent
• GitHub Copilot app

The staged rollout may take several days to reach every eligible account. Business and Enterprise administrators control access through the model policy in Copilot settings. New models are enabled by default unless an administrator disables them.

Usage determines the bill

GitHub bills Argon at provider list pricing through usage-based billing. Google’s introductory API rates are about $2 per million input tokens and $10 per million output tokens. Output therefore costs five times as much as input, making generated code, tool instructions, and repeated corrections relevant to the total cost of a long agent run.

Google’s numbers, with caveats

Google’s launch materials report 13 outright wins, one tie, and five losses across 19 comparisons with GPT-6 Astra, Claude Fable 5.1, and Claude Opus 5.5. The strongest reported results cover long-horizon engineering, knowledge work, automation, and very large context windows.

On DeepSWE v1.1, Google reports 77.9% for Argon, 74.2% for Claude Opus 5.5, and 74.1% for GPT-6 Astra. The benchmark resembles repository-level agent work more closely than isolated code generation because the model must sustain a plan, modify several files, and validate the result over an extended trajectory. Vendor-reported scores provide directional evidence; repository-level trials remain the more reliable guide for a specific codebase and toolchain.

Coding results split by workload

Google’s coding rows show Argon leading two benchmarks and ranking fourth among four models on FrontierSWE v2 and Terminal-Bench 4.0. Claude Opus 5.5 led Terminal-Bench 4.0 and scored 57.6 on the Artificial Analysis Intelligence Index, compared with Argon’s 52.6. Those results make shell-heavy and systems-oriented work a priority for direct model comparisons.

GitHub’s CLI gains and Google’s long-horizon scores make Argon a strong candidate for:

• Repository-wide refactors spanning many files
• Agent runs that plan, edit, test, and revise repeatedly
• Tasks requiring hundreds of thousands of tokens of context
• Large remediation projects with repeated validation steps

Terminal command sequences, low-level systems work, and tasks represented by FrontierSWE v2 warrant parallel testing with Opus 5.5 or GPT-6 Astra before teams choose a default.

Prompt-injection resistance

Google reports a 0.7% attack-success rate across 15 adaptive attempts on the Gray Swan Indirect Prompt Injection benchmark. Indirect prompt injection occurs when an agent encounters hostile instructions embedded in content it reads, such as an issue, repository file, webpage, or tool response. The result supports evaluating Argon for agents that process untrusted material.

Agents still need restricted credentials, sandboxed execution, explicit tool permissions, audit logs, and human review for sensitive changes. A benchmark score cannot cover every repository, tool integration, or attack path.

Copilot’s model picker gets crowded

Argon follows three frontier-model launches within the same month: GPT-6 Astra on September 3, 2026, Claude Opus 5.5 on September 22, and Gemini 4 Argon on October 13, 2026. Copilot users can now compare models from the leading labs within one interface instead of committing every workflow to a single provider.

GitHub also added Claude Haiku 5.5 the previous week. That release cadence is turning model selection into a workflow decision based on task shape, latency, completion quality, and token cost.

Test the behavior, not the label

Repository-level evaluations should use fixed code snapshots and identical task descriptions so each model faces the same conditions. Teams assessing Argon can track:

1. Task completion and test-pass rates
2. Regressions or unnecessary file changes
3. Human corrections required before merging
4. Recovery after failed commands or tests
5. Wall-clock time and token cost
6. Tool calls made without sufficient justification
7. Resistance to hostile instructions in repository content

These measurements will show whether Argon’s reported self-verification improves actual engineering work, particularly during long Copilot CLI sessions and multi-file changes.