ModelsArchitectures & capability
Google releases Gemini 4 Argon to paying subscribers and API customers, with introductory pricing that doubles after 30 November
Google's Gemini 4 Argon reached paying Gemini subscribers and API customers on 13 October. Third-party tests put it level with GPT-6 Astra and behind Claude Opus 5.5 overall. Its introductory price of $2/$10 per million tokens undercuts rivals but doubles after 30 November.
From the The Verge · AI article of October 13, 2026 and cited sources: Argon rolled out on 13 October to Pro and Ultra subscribers in the Gemini app and Search, and to paid Gemini API, AI Studio, Antigravity, Google Cloud and Gemini for Business users. [1] [2] [5]
SOURCES[1] Google finally releases Gemini 4 Argon theverge.com[2] Google — Gemini 4 Argon: Built to solve your hardest… blog.google[5] Google — Gemini 4 Argon: our next era of frontier… blog.google
AT Pulse research deskPublished · Updated · 5 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- What happened
- Argon rolled out on 13 October to Pro and Ultra subscribers in the Gemini app and Search, and to paid Gemini API, AI Studio, Antigravity, Google Cloud and Gemini for Business users. [1] [2] [5]
- Original source
- The Verge · AI article, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- October 13, 2026 article published
- Covered
- Oct 14 edition
- Evidence
- 9 sources across 7 sites
- 1Google finally releases Gemini 4 Argontheverge.com
- 2Google — Gemini 4 Argon: Built to solve your hardest…blog.google
- 3Gemini 4 Argon: Built to solve your hardest problemsdeepmind.google
- 4R&D World — Google's overdue Gemini 4 Argon reaches the…rdworldonline.com
- 5Google — Gemini 4 Argon: our next era of frontier…blog.google
- 6The Hacker News — Google Rolls Out Gemini 4 Argon to…thehackernews.com
Reach: not measured yet
Argon rolled out on 13 October to Pro and Ultra subscribers in the Gemini app and Search, and to paid Gemini API, AI Studio, Antigravity, Google Cloud and Gemini for Business users. No date has been set for Plus or free users. The API allows 1M tokens of input and output combined. [1] [2] [5]
Third-party results are mixed. Argon scored 53 on Artificial Analysis's Intelligence Index, tying GPT-6 Astra and trailing Opus 5.5 by five points. It tops the Vals Index and Zapier's automation benchmark and has a 15% hallucination rate on AA Omniscience. It trails Astra on FrontierSWE v2 and Terminal-Bench Science. Google says Argon beats its rivals across domains, but every result in its launch table came from Google's own runs, and none had been reproduced by others at launch. [1] [4] [7] [8]
Independent scores show Argon is competitive but not dominant, with strengths in finance, legal and automation tasks and a low hallucination rate.
Read the full assessment
Model choice should therefore rest on cost per task and testing against your own workloads, especially once standard pricing applies.
Executive brief
Google released Gemini 4 Argon to paying subscribers and API customers on 13 October 2026. The biggest surprise is that it was not clearly the strongest model. On Artificial Analysis's Intelligence Index it scored 53, the same as GPT‑6 Astra and five points behind Claude Opus 5.5, as R&D World reported. Google's own claim, as reported by The Verge, is that Argon beats those rivals. Where Argon clearly competes is price: an introductory $2/$10 per million input/output tokens, about a fifth of what its rivals charge. That price doubles after 30 November.
What changed and event timeline
Gemini 3 ships
This was Google's last major frontier release before Argon. In the months that followed, Anthropic and OpenAI models left Google trailing, and internal shake-ups raised doubts about its ambitions ().
Argon announced, cyber defenders first
Google gave the model to vetted defenders through its Fairwind Program. It also went through a voluntary US government review before release ().
Plan for a version without cyber guardrails
Google said trusted defenders and its internal teams would get Argon with its usual cyber guardrails removed. It also said Argon found an undisclosed critical flaw in healthcare software used by hospitals ().
First third-party scores land
Results were mixed: Argon tied Astra on Artificial Analysis's index, trailed Opus 5.5, and ranked eighth on Arena's Agent Arena ().
Meta launches Muse Wonder Full
Meta's own frontier model arrived one day before Argon's public rollout ().
General rollout
Argon reached the Gemini app, AI Mode in Search, the Gemini API, AI Studio and Antigravity ().
Capabilities and access
- Model: Gemini 4 Argon. Third-party tests used the "High" reasoning setting (MindStudio).
- Consumer access: Pro and Ultra subscribers in the Gemini app and Search. In Antigravity and AI Studio, Ultra subscribers only.
- Developer and enterprise access: paid Gemini API tier, Android Studio, Google Cloud and Gemini for Business (Google).
Read the full section
- Model: Gemini 4 Argon. Third-party tests used the "High" reasoning setting (MindStudio).
- Consumer access: Pro and Ultra subscribers in the Gemini app and Search. In Antigravity and AI Studio, Ultra subscribers only. Google has given no date for Plus or free users (The Verge).
- Developer and enterprise access: paid Gemini API tier, Android Studio, Google Cloud and Gemini for Business (Google).
- Price: $2/$10 per million input/output tokens until 30 November, then $4/$20. Cached inputs are 95% cheaper (Google).
- Limits: a combined 1M-token input-plus-output limit through the API. The output limit rises from the previous 64K tokens (Google).
Technical analysis for researchers and developers
- Architecture: none of the reviewed sources describe it. Google documents only a "continuation protocol" for long reasoning tasks and the 1M-token limit (Google).
- Coding results depend on the test. Argon scores 77.9% on DeepSWE v1.1 against Astra's 74.1%, but 55.0% on FrontierSWE v2 against Astra's 65.5% (R&D World).
- Argon leads LABBench2, which tests literature analysis and troubleshooting experiments (88.8%).
Read the full section
- Architecture: none of the reviewed sources describe it. Google documents only a "continuation protocol" for long reasoning tasks and the 1M-token limit (Google).
- Coding results depend on the test. Argon scores 77.9% on DeepSWE v1.1 against Astra's 74.1%, but 55.0% on FrontierSWE v2 against Astra's 65.5% (R&D World).
- Science results split the same way. Argon leads LABBench2, which tests literature analysis and troubleshooting experiments (88.8%). It trails on Terminal-Bench Science, which checks whether computational research tasks actually get completed (57.6% against Astra's 68.1%) (R&D World).
- Token use: Argon reportedly produces about 2.3 times as many output tokens per task as Astra. Teams should measure cost per task, not just the per-token price.
Claims and evidence
- Google says Argon beats GPT‑6 Astra, Claude Fable 5.1 and Opus 5.5 across coding, cyber, science, finance and law (The Verge).
- Third-party scores agree only in part:
- Argon tops the Vals Index at 68.9%, ahead of roughly 67% for the Claude 5.5 models.
Read the full section
- Google says Argon beats GPT‑6 Astra, Claude Fable 5.1 and Opus 5.5 across coding, cyber, science, finance and law (The Verge). One analysis found that all 19 launch-table results were run by Google itself, and none had been reproduced by a third party at launch (Emergent).
- Third-party scores agree only in part:
- Argon tops the Vals Index at 68.9%, ahead of roughly 67% for the Claude 5.5 models.
- It tops Zapier's automation benchmark at 51.29%.
- It has a 15% hallucination rate on AA Omniscience, against 51% for Astra.
- Opus 5.5 still leads Artificial Analysis's overall index (MindStudio).
- Cyber claims are Google's alone and cannot be checked yet. These include the healthcare vulnerability and the Wiz "Scan for Good" findings. Google has not named the affected software (The Hacker News).
Context and prior work
- Cyber-first rollouts are now standard. Releasing first to "trusted cyber defenders" follows recent frontier launches by other labs (The Verge).
- Comparison point: Google measures Argon's vulnerability-discovery gains against its earlier Gemini 3.8 Flash Cyber model (The Hacker News).
- Delays: R&D World links the late launch to a Gemini 3.5 Pro that it says was originally promised.
Read the full section
- Cyber-first rollouts are now standard. Releasing first to "trusted cyber defenders" follows recent frontier launches by other labs (The Verge).
- Comparison point: Google measures Argon's vulnerability-discovery gains against its earlier Gemini 3.8 Flash Cyber model (The Hacker News).
- Delays: R&D World links the late launch to a Gemini 3.5 Pro that it says was originally promised. This is reported framing, not something Google has confirmed (R&D World).
Limitations, safety and contested findings
- Safety figures are self-reported.
- Unrestricted cyber access is a deliberate trade-off. Giving vetted outside parties a version without guardrails increases dual-use exposure (The Hacker News).
- Coding ability is disputed. Some Google employees reportedly questioned Argon's practical coding skills.
Read the full section
- Safety figures are self-reported. Google cites a 0.7% attack success rate on Gray Swan's prompt-injection test, a 2% shortcut rate on coding tasks, and fewer than 1% of bio red-team sub-tasks unlocked (Google).
- Unrestricted cyber access is a deliberate trade-off. Giving vetted outside parties a version without guardrails increases dual-use exposure (The Hacker News).
- Coding ability is disputed. Some Google employees reportedly questioned Argon's practical coding skills. Google rejects that characterization (R&D World).
- Ranking and naming are inconsistent. Argon is #1 on Arena's overall leaderboard but #8 on Agent Arena (MindStudio, R&D World). One analysis lists the rival as "GPT 6.1 Astra" rather than GPT‑6 Astra.
Business and practitioner implications
- Cost: at introductory rates, R&D World puts Argon at $1.99 per task against $3.26 for Astra (R&D World).
- Where it looks strongest: finance, legal and tax tasks, automation, and keeping hallucinations low.
- Coding agents: compare it against Astra and Opus on your own repositories before switching.
Read the full section
- Cost: at introductory rates, R&D World puts Argon at $1.99 per task against $3.26 for Astra (R&D World). That advantage narrows considerably at the standard rate (MindStudio). Contracts should be priced on the post-30 November rate.
- Where it looks strongest: finance, legal and tax tasks, automation, and keeping hallucinations low.
- Coding agents: compare it against Astra and Opus on your own repositories before switching.
- Competition: with launches from Meta and Google in the same week, there are clearly more than two serious contenders at the frontier (The Verge).
The source trail.
Sources (9)
Google finally releases Gemini 4 Argon
Article text retrieved; extracted text may omit tables or interactive elements.
theverge.comGoogle — Gemini 4 Argon: Built to solve your hardest problems
blog.googleGemini 4 Argon: Built to solve your hardest problems
Related coverage; assess separately
deepmind.google