Gemini 4 Argon goes on sale with a paywall and a 1M-token output cap

What Google shipped, and who gets it

Google's new flagship model, Gemini 4 Argon, did not reach everyone at once. It went out in stages, and some users are still shut out. Google first sent the model to a small group of trusted cybersecurity defenders through its Fairwind Program. The company said it wanted to finish safety work and testing before a wider launch.11 On October 13, 2026, about two weeks later, Argon became available to developers and enterprises through the paid Gemini API, AI Studio and Google Cloud. Google AI Pro and Ultra subscribers got it in the Gemini app and in AI Mode in Search.4

The limits are as much a part of this launch as the model. People on Google's free plan and on the $4.99 AI Plus plan can't use Argon at all.1 Using Argon in Antigravity, Google's agentic coding environment, requires an Ultra subscription.20 In Search, Argon in AI Mode is open only to signed-in adults with personal Google accounts in supported countries, and it has daily caps that Google has not disclosed.18 SiliconANGLE reported that Ultra users get higher caps than Pro users, but no outlet has published the actual numbers.18

Coverage mostly agrees on the timeline, with small differences in dates. 9to5Google and others put the announcement on September 30.1416 Other reports say the Fairwind preview itself opened on October 1.8 At least two trackers had not caught up with the launch: one still listed Argon as restricted to phased trusted access5, and another, published October 18, still showed general availability as undated.2 The trackers that do record it agree on October 13 as the date of broad release.419

The headline feature: a million tokens of output

Google is leading with Argon's output limit. The company says the model can now produce up to 1 million output tokens, up from 64,000. Its pitch is that this gives the model enough room to reason through long, multi-step problems in a single pass.11 Google says the model is aimed at coding, financial research, legal drafting and autonomous patching of security vulnerabilities.11

Reports differ on how that million-token figure works, and the differences matter for developers. One detailed write-up says Argon has a shared budget of 1,048,576 tokens that covers input, internal reasoning, function calls and output together. It adds that a separate long-decode mode allows up to 1 million output tokens minus whatever the input used.19 Another account says the output comes from a continuation mechanism. Each API call stops at an internal limit, often between 32,000 and 65,000 tokens, and the server returns a token that lets the model pick up mid-word. Google's servers can also chain these segments together automatically.20 The same report says Argon does not use that room unless asked. Testers had to tell it explicitly not to summarize or stop early, and one documented million-token run took more than three hours across 16 segments.20

Taken together, the reports suggest the 1-million-token figure is better read as a maximum the API can reach with some setup than as how the model behaves by default. That is still a real difference from rivals. One comparison notes that neither Anthropic nor OpenAI has published a separate output limit for its current flagship.16 Teams expecting to get book-length answers without changing their prompts should plan for extra work.

On benchmarks, Google's own reported figures look strong but are narrow. Argon tied for first place on CWE-bench v1, which tests whether a model can fix security vulnerabilities, with a score of 68%.14 Other reports say it posted state-of-the-art results on the Vals Index of professional tasks, Automation Bench and the Deep Sweep coding benchmark.20 One rankings roundup credited Argon with a 15% hallucination rate, the lowest among top models. The same roundup still put Anthropic's Claude Opus 5.5 ahead on overall intelligence and coding.10

Why the cyber-first rollout matters

The Fairwind Program is the most unusual part of this release. Google trained Argon to be strong at cyber defense. During the preview, it made the model available to trusted defenders and internal teams without cyber guardrails, so they could use its full security capabilities.14 Early access was limited to government partners, operators of critical infrastructure and vetted security firms.19

This gives security defenders the strongest version of the model first and keeps the general public behind a safety review. Labs have been releasing more and more capable models into the security field, and Google's approach is a working example of tiered access based on how risky a capability is. Anthropic is doing something similar: its Mythos 5.1 tier is still restricted while the related Fable model is broadly available.5 The industry may be moving toward a norm where the strongest cyber capabilities go first, or only, to vetted users.

Google has also avoided a public setback that hit OpenAI. One October roundup, citing the Wall Street Journal, reported that OpenAI cancelled GPT-6.1 Astra after safety tests, and that the company has not put a replacement on its calendar.2 Next to that, Google's slower, gated rollout looks more cautious than restrictive.

Pricing: cheap now, double later

Google launched Argon at an introductory API price of $2 per million input tokens and $10 per million output tokens. Cached input costs 95% less than regular input.14 When the introductory period ends, the price rises to $4 for input and $20 for output.14 Several reports say the discount runs through November 30, so the higher rate starts in December.1816

At the introductory price, Argon costs about the same per token as Claude Sonnet 5.5 and GPT-6.1 Sol Pro, both of which list at $2 and $10.6 This is a deliberate move to win developers early. Anyone building budgets around today's price, though, faces a doubling within weeks. One analysis warned that any budget built now assumes a price that will double.2 Teams that come to depend on Argon's long-output workloads will feel the full increase, because output tokens are where the million-token feature uses the most.

A paywall strategy, not just a model launch

Argon arrived alongside changes to Google's consumer plans that point the same way. Starting October 9, free Gemini app users were limited to Gemini 3.5 Flash-Lite and lost access to 3.6 Flash and 3.1 Pro. AI Plus subscribers lost access to Pro.12 Meanwhile, the $19.99 AI Pro plan gained Deep Think, which had previously been available only on more expensive plans.12 The changes follow Google's May move to usage limits based on compute. Those limits reset every five hours up to a weekly cap, and AI Pro users get four times the standard allowance.17

The reporting also shows a pricing gap: one outlet describes Ultra at about $249.99 a month18, while 9to5Google refers to $99.99 and $199.99 plans above Pro.12 The direction is clearer than the exact prices. Several reports say Google is using Argon to widen the gap between free and paid tiers and push people toward subscriptions.1519 That is a gamble when competitors are still giving away capable models. Anthropic's Claude Sonnet 5.5 is still free on claude.ai, and free accounts can choose Claude Haiku 5.5.1

Where Argon sits in a crowded autumn

Argon arrived during an unusually busy few weeks of releases. Anthropic updated its entire public lineup to Opus 5.5, Sonnet 5.5 and Haiku 5.5 between September 22 and October 7.8 Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens.1 Anthropic also made a fast mode for Opus 5.5 generally available on October 15, at a reported $8 and $40.8 On October 21 it opened a waitlisted preview of Claude Duet, a background agent app.4

OpenAI brought GPT-6 to everyday ChatGPT users with what it calls Intelligent UI, meaning answers that can include charts and small interactive tools.9 On October 16 it added parallel agents to GPT-6 Pro.4 Open-weight models also kept pace. Meta released Muse Spark Open, a mixture-of-experts model with 816 billion total parameters, 42 billion active per token and a 1-million-token context.7 Meta called it the most capable US open-weight model so far, while keeping its top Muse Wonder models closed.4 Mistral has said the weights for Large 4 will ship by the end of October.2 Alibaba, MiniMax and DeepSeek each released new models during the week of October 17 to 21.3

In that field, Argon doesn't win on every measure. Its clearest advantages are long-output capacity and security-focused work, while Anthropic still leads several overall rankings.1016 Google's launch shows how flagship models are now released: in stages, priced to attract early users, and kept behind subscription tiers. Whether customers keep paying once the price doubles in December will be a better test than any benchmark.