Gemini 4 Argon brings Google back to the frontier, but it does not lead, and its price edge is temporary

What we checked

• Does the cluster reveal a broader market or architecture transition that a single release note can't capture? partly confirmed. Argon is a new sparse mixture-of-experts pretrain, not a fine-tune of an earlier model. It replaced the cancelled Gemini 3.5 Pro, and its release was staged: cyber defenders in the Fairwind program got it on September 30, and general availability followed on October 13. That staged, gated rollout and the 1M-token output window are real shifts in how the model is shipped. The underlying architecture (MoE on TPUs) continues Google's earlier approach. [7][8][2][6]
• (Added) Is Argon the clear frontier leader, as Google's benchmark table suggests? refuted. On the Artificial Analysis Intelligence Index, Argon scores 53. That ties GPT-6 Astra and sits about 5 points behind Claude Opus 5.5 (58). Google's own model card shows Argon trailing on FrontierSWE v2, Terminal-bench 4.0 and OSWorld-2.0. [4][7][6]
• (Added) Does Argon's cost advantage depend on introductory pricing? confirmed. Artificial Analysis measured $1.99 per index task at the $2/$10 introductory price and $3.98 at the standard $4/$20, against $3.26 for GPT-6 Astra. Argon also used about 62k output tokens per task, compared with about 27k for Astra. Google says the introductory price runs through November 30, 2026. [4][3][2]

What we found

• Release terms: the model ID is gemini-4-argon and it is marked stable. It shares a 1,048,576-token window between input and output. Output defaults to 65,536 tokens and can use the full window via long decode. Inputs can be text, image, video, audio and PDF, and output is text only. Thinking has low, medium and high levels. The model supports function calling, computer use (in preview), search and Maps grounding, and batch and flex inference. It does not support the Live API or explicit caching. [1]
• Pricing: the introductory rate is $2 input / $10 output per 1M tokens through November 30, 2026, with a 95% discount on cached input. After that it rises to $4 / $20. When Artificial Analysis published on September 30, Google had not confirmed when the promotion would end. The November 30 date appeared in Google's October 13–14 posts. [2][3][4]
• Independent results: Artificial Analysis gives Argon (high) a score of 53. That ties GPT-6 Astra (max), puts it 1 point ahead of GPT-6.1 Sol and 23 points ahead of Gemini 3.1 Pro Preview, and leaves it below Claude Opus 5.5 (58). Its hallucination rate on AA-Omniscience is 15%, against 51% for Astra. It placed first on AutomationBench-AA, but on Terminal Bench 4 it scored 57%, behind Sonnet 5.5, Opus 5.5 and Astra. [4][5]
• Argon uses a lot of tokens. It averaged about 62k output tokens per index task, compared with about 27k for Astra. Its lower cost per task, $1.99 against Astra's $3.26, comes from cheaper token prices, not efficiency. At standard prices the figure rises to $3.98. [4][6]
• Google's own model card is mixed. Argon leads on the Vals Index (68.9% against 67.0% for Opus 5.5). It trails on FrontierSWE v2 (55.0% against 65.5% for Astra) and Terminal-bench 4.0 (57.4% against 66.4% for Opus 5.5). R&D World also notes that Google ran some science benchmarks itself while taking competitors' scores from public leaderboards. [7][6]
• Safety and release gating: under Google's Frontier Safety Framework, the model card says the cyber critical capability level (CCL) and the biological CCL could not be ruled out, so Google built safety cases for both. Google released the model in phases and took part in the US government's voluntary pre-release access process. Fairwind defenders can apply for a version without cyber guardrails, while the general release blocks offensive cyber tasks. Google reports a 0.7% attack success rate on Gray Swan's indirect prompt injection benchmark. [7][8][3]
• Google's internal-impact claims are self-reported and not independently verified. They include agentic code submissions in Antigravity up more than 35%, a 2.85x vLLM inference speedup for Gemma, and lab-confirmed enzyme designs. The card lists a knowledge cutoff of March 2026. [2][7]

What it means

For most builders, Argon is now a real third option alongside GPT-6 Astra and Claude, and is no longer a fallback. It looks strongest on knowledge work, long-context retrieval and tasks where hallucinations are expensive: its 15% hallucination rate on AA-Omniscience is the best independent result in this launch. It is weaker on terminal-style agentic coding, where Claude and Astra still lead on both Google's numbers and Artificial Analysis's. If coding agents are your main workload, run Argon on your own tasks before switching. Don't rely on the Vals Index lead.

Calculate costs at $4/$20, not the promotional $2/$10. Budget for Argon's long outputs, roughly twice Astra's tokens per task on Artificial Analysis's suite, and use the low or medium thinking levels where you can. The 1M-token output window is useful for long generation jobs, but it can also run up costs quickly. Security-sensitive teams should know that the general model blocks offensive cyber tasks, while an unguarded version exists for vetted defenders. Google may handle future frontier releases with the same staged, gated approach, so plan for gaps between announcement and broad access.

Still open

• Will Google extend the introductory pricing past November 30, 2026, and what will cached-input pricing be at standard rates?
• Why do Terminal-Bench 4 scores for competitors differ between Google's model card and Artificial Analysis? The harness or settings may differ.
• When will smaller or cheaper Gemini 4 variants ship, and will the Flash line move to Gemini 4?
• What were the details of the Frontier Safety Framework safety cases for cyber and bio, beyond the model card summary?