Gemini 4 Argon (2026): 1M Tokens, Price & Benchmarks

Google just raised the bar for what an AI model can do in a single run. Gemini 4 Argon, announced by Google DeepMind on September 30 and released to developers on October 13, 2026, is the company's new frontier model — and its headline number is staggering: one million tokens in a single unified budget.

One important nuance: this isn't a flat 1M output allowance on top of a separate input window. The 1,048,576-token budget is shared between input, thinking, and output. Standard synchronous generation caps at 65,536 tokens, but a new "long decode" continuation protocol lets generation keep going — up to the remaining budget — for truly enormous single runs. That's still up from just 64,000 tokens of output in Google's previous models, and it changes what's possible: when a model can "think" through hundreds of thousands of tokens in one trajectory, it can inspect information, reason, write code, test its work, and iterate — all without starting over. Google calls this the "next era of frontier intelligence."

What Makes Argon Different?

Unlike models built around quick question-and-answer exchanges, Argon is engineered for long-horizon workflows: real-world software engineering, multi-step financial research, legal drafting, cybersecurity defense, and even robotics.

Key specs:

• 1M-token unified budget — 1,048,576 tokens shared by input, thinking, and output; long decode extends generation toward that ceiling
• Three thinking levels — Low, Medium, and High reasoning; on the Artificial Analysis Intelligence Index, Argon High scored 53 points with a 15% hallucination rate on AA-Omniscience, though its answer accuracy ranked below GPT-6 Astra
• Multimodal: accepts text, images, audio, video, and PDFs (text output)
• Knowledge cutoff: March 2026
• No open weights — cloud-only via the paid Gemini API, AI Studio, Android Studio, and the Gemini platform; also in Google Antigravity for Ultra subscribers, and for Pro/Ultra subscribers in the Gemini app and AI Mode in Search

Thinking Deeper: How Argon Reasons

Argon's three official reasoning levels (Low, Medium, High) let developers trade cost and speed against quality — a familiar idea taken further. At High reasoning, independent evaluator Artificial Analysis recorded a respectable 53 on its Intelligence Index and found the model notably cautious: a 15% hallucination rate on AA-Omniscience suggests it's relatively willing to avoid unsupported answers. But the same evaluation placed Argon's answer accuracy below GPT-6 Astra, a reminder that "thinks longer" and "answers better" are not the same thing.

One feature matters enormously for agentic systems: cached input at a 95% discount. Prompts that reuse the same repository context, policy documents, or enterprise knowledge base get nearly free re-processing of that prefix — though only through implicit caching, not explicit control. For companies running long-horizon agents that re-send huge contexts on every step, this discount could dwarf the headline per-token price.

Google's other claims from the October 13 launch: #1 on the Vals Index (an independent ranking, at 68.90%), a 35%+ increase in agentic code submissions with nearly one-third fewer rollbacks among Google's own teams using Argon in Antigravity, and a 0.7% indirect prompt-injection attack success rate for its cybersecurity testing. Google is also offering 10,000 researchers the top Antigravity usage tier free for six months.

Pricing and Availability

Argon is rolling out to Google AI Ultra subscribers and paid API customers. The introductory API price — $2 per million input tokens, $10 per million output — runs through November 30, 2026, then doubles to $4/$20. That deadline matters: a workflow that burns a million output tokens will cost $10 today and $20 in December. And because the budget is shared, long runs also burn input budget — plan for both sides of the meter.

Notably, GitHub added Gemini 4 Argon to Copilot on October 13, so developers can try it right inside their editor without touching the API directly.

Who Should Actually Use Argon?

Argon makes sense if you're doing long, multi-step work: migrating a codebase, analyzing a pile of financial filings, drafting a legal memo from research, or running a security remediation agent. It does not make sense for quick chat-style Q&A — you'll pay premium prices (and wait longer) for capacity you never use. Flash-class models like Gemini 3.8 Flash remain the right tool for short tasks. The honest rule: buy the reasoning budget you actually burn.

Benchmarks: Leading in Knowledge Work, Mixed in Coding

Google reports strong results across domain-specific evaluations:

| Benchmark | Argon score | Context | |---|---|---| | DeepSWE v1.1 (coding) | 77.9% | Beats Claude Opus 5.5 (74.2%), GPT-6 Astra (74.1%) | | AutomationBench (business workflows) | 51.3% | Ranked #1 | | LVBench (long video understanding) | 91.7% | State of the art | | CWE-bench v1 (security fixes) | 68% | Ties for first place | | Harvey Legal benchmark | 19.6% | vs GPT-6 Astra's 5.4% | | Embodied Reasoning (robotics) | 87% | Beats all frontier models |

Honest caveats: Argon trails on FrontierSWE v2 (55%, last of four models in Google's own table) and Terminal-Bench 4.0, where Anthropic's Claude Opus 5.5 leads. No model wins everywhere — and these are largely Google-reported numbers.

Pricing and Availability

Argon is rolling out to Google AI Ultra subscribers and paid API customers. The introductory API price — $2 per million input tokens, $10 per million output — runs through November 30, 2026, then doubles to $4/$20. Developers should lock in early usage if they plan to build on it.

Notably, GitHub added Gemini 4 Argon to Copilot on October 13, so developers can try it right inside their editor.

The Robotics Surprise

Argon's most futuristic trick has nothing to do with chat. Using "Video-to-Sim-to-Real," the model watches a video of a human doing a physical task, builds a full physics simulation, and trains a robot in simulation before the real attempt. Google demonstrated:

• Apptronik's Apollo humanoid sinking seven consecutive cornhole shots after watching a human throw — zero prior task-specific training
• Boston Dynamics' Spot stepping on a pedal bin, grabbing a soda can, and dropping it in — after Argon first tested its balance in simulation
• Digital twins from a single photo: one snapshot of a table generated 4,500+ training episodes, lifting task success from 56% to 75%

The Bottom Line

Gemini 4 Argon is Google's answer to the agentic AI era: a model that doesn't just answer, but works — through code, documents, security flaws, and even physical robots. Its million-token unified context is a genuine technical leap. The open questions are pricing after November, and whether independent testing confirms Google's benchmark story. For developers and enterprises doing long, complex work, it's the model to watch this quarter.

Sources & Further Reading

• Gemini 4 Argon Model Card — Google DeepMind (official)
• Gemini App 4 Argon — 9to5Google
• Gemini 4 Argon in GitHub Copilot — GitHub Changelog