# Gemini 4 Argon Is Live: What Developers Need to Know > Source: > Published: 2026-10-15 10:07:53+00:00 Google’s Gemini 4 Argon went live on October 13 with one specification that reshapes the math for agentic workflows: a unified 1-million-token context window that covers both input and output. That’s a 16x increase over the previous 64K output ceiling, and it’s more important than anything else in the announcement. ## What Gemini 4 Argon Actually Is Argon is Google’s new flagship reasoning model, built for complex software engineering, multi-step autonomous agents, and long-horizon knowledge work. It’s available now via the [Gemini API](https://ai.google.dev/gemini-api/docs/models/gemini-4-argon), Google AI Studio, and Google Cloud. Model ID: `gemini-4-argon`. No free tier — paid API only. Key specs at a glance: 1,048,576-token shared context window, multimodal inputs (text, image, video, audio, PDF), text output only, and tunable thinking modes (low, medium, high). Supported capabilities include code execution, computer use (Preview), function calling, structured outputs, and file search. ## Why 1M Output Changes Agentic Workflows The previous generation’s 64K output limit was the invisible ceiling that broke most serious agentic pipelines. Hit the limit mid-task and you’re either chunking your work — which compounds errors — or restarting with a summarized context, which loses fidelity. Neither is acceptable for production-grade agents. Argon removes that constraint. Google’s own engineering teams used Argon agents to translate the 800,000-line Fuchsia Zircon kernel from C++ to Rust in a single agent run. The agents processed compiler outputs, ran profile-guided iterations, and replaced 32,000 lines of SIMD code with safe Rust that auto-vectorizes. The resulting decoder runs **2.7x faster** than the previous manual Rust implementation. That’s the kind of result that becomes impossible when your agent restarts every 64K tokens. ## Pricing: There’s a Window Here Through November 30, Argon is priced at $2 per million input tokens and $10 per million output tokens — with cached input at $0.10 per million (a 95% discount). That undercuts [Claude Opus 5.5](https://www.anthropic.com/claude/opus) ($4/$20) and GPT-6 Astra ($10/$50) significantly. On December 1, Argon’s pricing doubles to $4/$20, putting it level with Opus 5.5. If you’re evaluating Argon for your stack, do it before December. Implicit caching also rewards agentic workflows that re-read the same context repeatedly — a common pattern in long-horizon tasks. ## Benchmarks: Strong Where It Matters, Weak Where It Doesn’t Argon tops the [Vals Index](https://kingy.ai/blog/gemini-4-argon-specs-benchmarks-pricing/) — covering coding, finance, legal, and tax — and leads 13 of 19 benchmark rows. It hits 84.2% on GraphWalks (long-context retrieval), 77.9% on DeepSWE v1.1, and 88.8% on LABBench 2. It’s not the best pure coding agent. On FrontierSWE v2, Argon scores 55% versus GPT-6 Astra’s 65.5%. On Terminal-bench 4.0, it scores 57.4% against Claude Opus 5.5’s 66.4%. An independent 100-task evaluation by [Gertlabs](https://news.ycombinator.com/item?id=50083864) put it “roughly two months behind the frontier” in head-to-head coding work. The pattern is consistent: Argon wins on knowledge-heavy, long-context, reasoning-intensive tasks. It loses on tight terminal work and frontier coding benchmarks. That’s a deliberate trade-off, not a flaw — just make sure it matches your actual use case before committing. ## One Real Problem: Reasoning Loops Multiple evaluators — including Gertlabs and developers on Hacker News — have flagged Argon for getting stuck in infinite reasoning loops. Gertlabs called it “the first near-frontier model we’ve caught with this problem in a long time.” The model also struggles with reasoning block delineation, which causes higher failure rates in structured reasoning outputs. This matters for autonomous agent deployments. If you’re building agents that run unsupervised terminal tasks, Argon is not the safe default right now. Keep humans in the loop on multi-step work until Google ships a fix. ## How to Get Started Access is through the [Gemini API documentation](https://ai.google.dev/gemini-api/docs/models/gemini-4-argon) or test it directly in [Google AI Studio](https://aistudio.google.com/prompts/new_chat?model=gemini-4-argon). The API reference covers thinking mode configuration, long decode setup, and tool use integration. For agentic use cases — especially long-context migration work, document analysis, and security research — Argon is worth evaluating now, at the introductory price, before December resets the calculus. For pure coding agent work, test against Opus 5.5 and GPT-6 Astra before deciding.