Google Argon hits GA with 1M-token output and $2/$10 API price

Google DeepMind pushed Gemini 4 Argon into general availability on October 13, 2026, with the largest output ceiling of any frontier model — up to 1 million tokens per response — and an introductory API price of $2 per million input tokens and $10 per million output tokens. The release sets up a fresh round of competition with OpenAI's GPT-6 Astra and Anthropic's Claude Opus 5.5, both of which cap output at 128,000 tokens.

What is Gemini 4 Argon?

Argon is a sparse mixture-of-experts transformer. That means a router activates only a subset of the model's expert sub-networks per token rather than the full model. Google built it for long, multi-step tasks in coding, finance, legal, science, and cybersecurity. It accepts text, images, audio, and video, and outputs text only.

Google trained Argon on its TPUs (Tensor Processing Units, Google's custom AI accelerators) using JAX (a Python numerical computing library) and ML Pathways. The knowledge cutoff sits at March 2026. Google has not disclosed total or active parameter counts, which leaves buyers with no independent way to size the model. A caveat follows: parameter opacity is now standard practice, but it limits reproducibility for outside researchers.

How does the 1M token output limit work?

The new ceiling rises from 64,000 tokens on prior Gemini models. Google documents a "long decode" protocol that lets developers chain output beyond the normal limit while preserving the shared 1M-token input-plus-output window. A long prompt eats into the room left for the reply.

At introductory pricing, a single 1M-token response costs $10 in output charges alone. After the promotion ends on November 30, 2026, the rate doubles to $4 input and $20 output per million tokens. That post-promo price matches Claude Opus 5.5 and undercuts GPT-6 Astra's $10 and $50. Cached input tokens get a 95% discount during the promo.

Where does Argon win on benchmarks?

Google's internal table puts Argon first on most rows it measures. On the Vals Index — an independent ranking of model quality on real-world tasks — Argon posts 68.90%, ahead of Claude Sonnet 5.5 at 67.04% and ranked #1 by Vals AI. Argon also leads on:

• DeepSWE v1.1 (a software-engineering benchmark): 77.9%, ahead of Opus 5.5 (74.2%) and GPT-6 Astra (74.1%)
• AutomationBench: 51.3%, versus 42.5% for Opus 5.5
• GraphWalks (256K to 1M tokens): 84.2%, versus 71.8% for GPT-6 Astra
• LVBench (long-video understanding): 91.7%, the top score
• CWE-bench v1 (cybersecurity vulnerability detection): 68.0%, tied with GPT-6 Astra
• ERQA embodied reasoning: 87%

One outlier stands out. On Harvey's Legal Agent Benchmark, Argon scores 19.6% — nearly three times the next-best model, yet still an absolute low. Buyers in regulated legal workflows should read that number as a warning flag, not a victory lap.

Where does Argon fall short?

Argon finishes last of four models on five rows in Google's own table. The pattern matters because three of those losses come on agentic-coding benchmarks — tests that measure a model's ability to act on a developer's behalf, not just answer.

• FrontierSWE v2: 55.0%, versus 65.5% for GPT-6 Astra
• Terminal-Bench 4.0: 57.4%, versus 66.4% for Opus 5.5
• Terminal-Bench Science 0.1: 57.6%, versus 68.1% for GPT-6 Astra
• PostTrainBench: 45.3%, versus 49.3% for Opus 5.5
• OSWorld-2.0 (offline subset): 69.2%, versus 72.6% for GPT-6 Astra

Argon's DeepSWE score is self-computed using a mini-swe-agent harness (an automated software-engineering evaluation tool). The rival DeepSWE numbers come from system cards and leaderboards. Google flags the model's own hallucinations and occasional slowness or timeouts in the model card. Most rival scores in the table are self-reported or pulled from public leaderboards, which means the head-to-head is not apples-to-apples.

A separate note from Google's Frontier Safety Framework: the company says it cannot rule out the biological and cybersecurity critical-capability levels, and reports deploying safeguards it assesses as lowering the risk. The model card warns users explicitly.

How does Argon stack up against GPT-6 Astra, Claude Opus 5.5, and Claude Fable 5.1?

Google positions Argon against three rivals. The argument is mostly about output length and price.

• Argon: 1M-token max output, $2/$10 intro ($4/$20 after Nov 30), March 2026 knowledge cutoff, sparse MoE
• GPT-6 Astra: 128K output, $10/$50, April 2026 cutoff, 1,050,000 context window
• Claude Opus 5.5: 128K output, $4/$20, June 2026 cutoff
• Claude Fable 5.1: 128K output, $10/$50, June 2026 cutoff

Argon wins on input modalities — it takes audio and video natively — and ties or leads on long-context tasks. It trails on agentic coding and terminal-driven workflows, where Opus 5.5 holds a 9-point edge on Terminal-Bench 4.0.

What did Google use Argon for before launch?

Google shared internal results in its launch post. Argon in Antigravity — Google's agentic coding tool — raised agentic code submissions by over 35% and cut rollbacks by nearly a third. The model delivered a 2.85x end-to-end vLLM (an open-source inference engine) speedup for Gemma 4 E2B on an A100 GPU. In Google's open-source Agent Development Kit, Argon cut session prep time by 40% and made cold imports 21% faster. On Gray Swan's adaptive prompt-injection benchmark — a test of resistance to adversarial inputs — attack success fell to 0.7% across 15 attempts.

These are vendor benchmarks. The promised 2.85x inference speedup, in particular, has not been reproduced by outside parties. Treat the numbers as evidence of intent, not proof of parity.

How much does Argon cost, and where can developers use it?

Introductory pricing is $2 input and $10 output per million tokens through November 30, 2026. Cached input tokens cost 95% less. After the promo, prices double to $4 and $20, matching Opus 5.5.

• Developers: Gemini API paid tier in AI Studio and Android Studio
• Enterprises: Gemini platform, Gemini for business, and Google Antigravity
• Consumers: Gemini app and AI Mode in Search for Google AI Pro and Ultra subscribers
• Cyber defenders: the Fairwind Program, with access to Argon without cyber guardrails on a rolling basis

Google first announced Argon on September 30, 2026, for a small group of cyber defenders under Fairwind. The general-availability release broadens that access while keeping the unguarded version restricted to vetted partners. The dual-track rollout suggests Google wants commercial breadth without expanding the attack surface of a model its own safety framework flags as borderline on biological and cybersecurity capabilities.

The story here is not just another model launch. Argon reframes the API comparison around output length and price — two variables that affect unit economics for any workflow that needs to generate more than a short reply. Whether the 1M-token output ceiling becomes a real productivity lever or an expensive toy will depend on how reliably Argon finishes long outputs — and on whether OpenAI and Anthropic choose to match it.