Gemini 4 Argon is now broadly available. Google AI Pro and Ultra subscribers get it in the Gemini app and AI Mode in Search; developers can use the paid Gemini API. Free access is not part of the standard rollout. API pricing starts at $2 per million input tokens and $10 per million output tokens through November 30, then rises to $4 and $20 on December 1.
The catch is that access in one Google product does not buy access in every other one. Google's rollout announcement and API pricing make that distinction worth checking before upgrading.
Who gets Gemini 4 Argon now?
Consumer plans
| Surface / plan | Argon access | Notes |
|---|---|---|
| Gemini app, Free | Not included in the announced rollout | No general free Argon entitlement announced. |
| Gemini app, Google AI Plus | Not included in the announced rollout | A paid subscription alone does not guarantee Argon. |
| Gemini app, Google AI Pro | Yes | Subject to usage limits and rollout. |
| Gemini app, Google AI Ultra | Yes | Higher usage allowance than Pro. |
| AI Mode in Search | Pro and Ultra | Subject to the product's availability. |
Sources: Google's announcement, Gemini app announcement and usage limits. These rows cover personal accounts. Work and school accounts are licensed separately; Google lists Gemini for business among Argon's enterprise routes. Google broad rollout announcement, October 13
Developer access and coding tools
| Surface / billing route | Argon access | Notes |
|---|---|---|
| Gemini API Free tier | No | Argon requires paid access. |
| Paid Gemini API | Yes | Public model ID: gemini-4-argon. |
| Google AI Studio, paid API route | Yes | Use the paid Gemini API through AI Studio. |
| Google AI Studio, subscription route | AI Ultra | Google's pricing page excludes the Pro plan. |
| Google Antigravity | AI Ultra or eligible enterprise subscription | Antigravity 2.0, CLI and IDE extensions; excludes Antigravity IDE. |
| GitHub Copilot | Pro+, Max, Business and Enterprise | Gradual rollout; organization policy can affect access. |
| Cursor | Listed in official model documentation | Usage draws from its Other Models pool. |
Sources: developer announcement, API pricing, Antigravity models, GitHub Changelog and Cursor documentation.
What is the Gemini 4 release date?
Gemini 4 Argon became generally available through the Gemini API on October 13, 2026, according to Google's API changelog. That is the broad-release milestone this article had been waiting for.
Google originally announced Argon on September 30, beginning with trusted cybersecurity defenders. Announcement and general availability were separate events. This page was originally published on September 29 and has been updated as the story developed.
Google describes the new rollout as global, but that should not be read as every feature appearing immediately in every account. Gemini web and mobile availability can differ, and subscription eligibility still applies.
Who gets Argon in the Gemini app?
Google says Pro and Ultra subscribers can now use Argon in the Gemini app. Gemini app and macOS announcement The controls below come from Google's general model-selection guidance.
Subscribers can switch models manually and choose a Low, Medium or High thinking level. Low is the default if you do not pick one. With auto model selection on, Gemini routes each prompt to whichever model and thinking level it judges most efficient, so having Argon access does not mean every answer uses it. The model behind a response appears in its details on mobile and at the end of the thinking steps on the web. Gemini app model availability and selection
The limits are also less tidy than a daily message counter. Google describes compute-based allowances with five-hour resets and a weekly ceiling. Longer conversations, harder prompts and more demanding models consume more usage. Pro has four times the standard allowance; Ultra offers five or 20 times Pro's allowance, depending on the subscription. Those are plan-level ratios, not a guaranteed number of Argon responses. If you reach the limit on a paid plan, the chat continues on Flash-Lite. Gemini app usage limits and eligibility
If you already have Pro, check the model selector before buying Ultra purely to try Argon in chat.
Is Gemini 4 Argon free?
There is no standard free Argon tier in the rollout documented here. Google's consumer announcement names Pro and Ultra; it does not include Free or Plus. The Gemini API pricing page explicitly excludes the API Free tier. Google broad rollout announcement, October 13(https://ai.google.dev/gemini-api/docs/pricing)Gemini app and macOS announcement
That does not rule out specific offers. Google is offering up to 10,000 scientists six months of Antigravity's highest usage tier free of charge, onboarded on a rolling basis and subject to eligibility, capacity and selection. Applications are open. That is a research program, not free public access for everyone. Google broad rollout announcement, October 13
How Gemini 4 API access works
Developers now have a stable public model name: gemini-4-argon. Google's model directory replaces the previous uncertainty about whether a public endpoint existed.
Google names the paid Gemini API in AI Studio and Android Studio as developer routes. Google broad rollout announcement, October 13 Enterprise access also extends through the API in Gemini platform. Google broad rollout announcement, October 13
Treat migration as more than changing a string. Argon supports implicit caching only, not explicit caching. Gemini API pricing A low output cap can stop the model before it finishes thinking, so check the limit set in any agent framework or evaluation tool. Long-decode documentation Start with a small evaluation set from your own workload before sending production traffic across.
Gemini 4 Argon API pricing
All figures below are US dollars per million tokens. Output charges include thinking tokens. These are Gemini API rates, separate from monthly consumer subscriptions. Gemini API pricing
| API service | Input through Nov. 30 | Output through Nov. 30 | Input from Dec. 1 | Output from Dec. 1 |
|---|---|---|---|---|
| Standard | $2 | $10 | $4 | $20 |
| Batch | $1 | $5 | $2 | $10 |
| Flex | $1 | $5 | $2 | $10 |
| Priority | $3.60 | $18 | $7.20 | $36 |
Standard cached input costs $0.10 now and $0.20 from December 1: a 95% discount from the corresponding uncached rate. Argon supports implicit caching, not explicit cache storage. Tool charges can add to the token bill. Gemini API pricing
Across a workload totaling one million uncached input tokens and one million billable output tokens at standard rates, the arithmetic is $12 during the promotion and $24 afterward, before tool charges. That is an illustration, not an average job cost.
A trial can establish whether the model works for you. The December rates establish whether the economics still work afterward. Google has published those later rates; it has not promised they will remain unchanged forever.
Google AI Pro, API billing and AI Studio are different things
The apparent contradiction in Google's documentation is about the product surface.
Pro includes Argon in the Gemini app. Pro's AI Studio plan does not. Google's pricing page says Argon is unavailable on the Free tier and on the Google AI Pro plan, while Ultra subscribers can try it in AI Studio. Paid API billing is a separate route. Gemini API pricing(https://blog.google/innovation-and-ai/technology/developers-tools/build-with-gemini-4-argon/)
Google's AI Studio plan documentation says subscription benefits apply within the web interface. Direct API calls and external applications have separate billing and usage tiers. An Ultra subscription is not an unlimited API key, and a Pro subscriber can separately enable paid API access.
The same documentation separates Google One AI credits from Cloud credits. Eligible developer credits may offset some usage, but do not collapse those systems into one shared allowance.
Before paying, identify the job: chat in Gemini, prototype in AI Studio, or run an application through the API. The word "Pro" is doing rather too much work otherwise.
Argon in Antigravity and the Gemini Mac app
For Argon in Antigravity, Google's model table marks Free, Plus and Pro as unavailable; Ultra and enterprise access are listed. Its footnote limits support to the Antigravity app, CLI and IDE extensions, excluding Antigravity IDE. Antigravity model availability
Separately, Google has added Antigravity's coding interface to the Gemini macOS app and a chat-to-tasks toggle for work that can continue in the background. Gemini app and macOS announcement That integration should not be used to infer that every Gemini Pro subscriber receives the standalone Antigravity Argon entitlement. Check the surface you intend to use.
Where else can you use Argon?
GitHub Copilot: GitHub lists Argon as GA for Pro+, Max, Business and Enterprise, with gradual rollout across supported clients. It is billed at provider list pricing under usage-based billing. Business and Enterprise administrators can control model access. GitHub Copilot announcement
Cursor: official documentation now lists Argon and its pricing, so availability no longer rests on an X post. Cursor lists a 200K context window and 1M maximum context; its product configuration should not be assumed identical to Google's raw API. Cursor model documentation
Neither integration makes its vendor's benchmark a universal ranking. In particular, "on par with Opus" needs the benchmark, effort setting and cost attached before it means much.
Context window and the 1M-token output claim
Google's model directory and long-decode documentation now resolve the earlier uncertainty. Long-decode documentation(https://ai.google.dev/gemini-api/docs/models)
| Specification | Published limit |
|---|---|
| Unified context window | 1,048,576 tokens |
| Default synchronous maximum output | 65,536 tokens |
| Long decode | Up to 1,048,576 tokens, within the shared context budget |
Input, thinking and output share the same window. A million-token input does not leave another million tokens for an answer. Long-decode documentation
Google's long-decode guide explains that the Interactions API requires background execution and an output limit above 65,536 to continue beyond the normal ceiling. The server generates successive 64K hops; background runs can last up to 24 hours. Client-side continuation is also available through the GenerateContent API. More room permits longer reasoning; it does not guarantee a requested essay length.
There is a billing consequence too. Each hop sends the prompt plus everything generated so far back to the model as input, so reported input tokens can far exceed the real context size. Implicit caching discounts that repeated input by 95%, and every thinking token is billed. Long-decode documentation Inspect actual usage rather than estimating cost from the final text. A large ceiling is useful engineering headroom, not a target to fill.
Google's benchmarks are strong, but Argon does not win everything
The rows below are a selection from Google's current performance table; they are not Choosely test results. The selection includes every row where Argon does not lead outright. Across Google's full 19-row table, Argon has the outright top score on 13 rows, ties on one and trails on five, among the scores reported. This is a row count, not a weighted overall ranking.
| Benchmark | Gemini 4 Argon | GPT-6 Astra | Claude Fable 5.1 | Claude Opus 5.5 |
|---|---|---|---|---|
| Vals Index | 68.9% | 63.1% | 65.8% | 67.0% |
| AutomationBench | 51.3% | 41.4% | 31.4% | 42.5% |
| DeepSWE v1.1 | 77.9% | 74.1% | 67.4% | 74.2% |
| FrontierSWE v2 | 55.0% | 65.5% | 56.3% | 62.3% |
| Terminal-bench 4.0 | 57.4% | 58.2% | 57.9% | 66.4% |
| Terminal-Bench Science 0.1 | 57.6% | 68.1% | 52.6% | 63.3% |
| LABBench 2 | 88.8% | 85.4% | 68.6% | 73.1% |
| PostTrainBench | 45.3% | 44.3% | 40.2% | 49.3% |
| Agent's Last Exam | 39.5% | 34.2% | n/a | 38.2% |
| OSWorld-2.0, offline subset / partial score | 69.2% | 72.6% | n/a | n/a |
| LVBench | 91.7% | 87.5% | 79.7% | 83.7% |
| CWE-bench v1 | 68.0% | 68.0% | 58.0% | 67.0% |
In this selection, Argon leads six rows and ties CWE-bench. Astra and Opus lead the other five. The evidence supports a strong, competitive model, not a clean sweep.
Google's methodology says competitor scores default to providers' self-reported numbers. Many rows come from public leaderboards instead, and Google ran LABBench 2, PostTrainBench, GraphWalks and LVBench itself for every model. On LVBench, each model received a different number of video frames because of API limits. Harnesses and reasoning settings vary, so this is not a single controlled Choosely comparison of four identical products.
General availability now permits broader testing. It does not turn the launch evidence into independent proof of superiority. The public model card also acknowledges hallucinations, occasional slowness and timeouts, and gives a March 2026 knowledge cutoff.
What happened to the Fairwind-only rollout?
The September announcement prioritized trusted defenders while Google worked on safeguards. Original September announcement Broad commercial access has now opened, but Fairwind remains a separate controlled route for cybersecurity organizations. Fairwind Program
That distinction matters because access without cyber guardrails is not the same product promise as ordinary public access. The launch history remains relevant to understanding why there were two release milestones.
Google's original safety work covered misuse defenses, indirect prompt injection, monitoring and isolated training and evaluation environments. Original September announcement Those controls remain relevant context for agent deployment. An internal evaluation sandbox is not a consumer feature, and a model's safety claims do not remove the need to review its actions.
What about Gemini 3.5 Pro?
Argon is the publicly documented Gemini 4 model. Google's current API model directory does not list Gemini 3.5 Pro. The present access decision no longer depends on guessing whether an earlier planned flagship will arrive.
For a different Gemini capability, see Choosely's Gemini 3.8 Live Avatar analysis.
What remains unknown?
The sources reviewed do not establish a future free or Plus Argon rollout date, or a date for Pro access to Argon in standalone Antigravity. They also cannot tell you how well it will perform on your particular codebase or research workflow.
An account-specific rollout delay should not be mistaken for a promise of imminent access on an ineligible plan. Check eligibility first, then the product's current rollout information.
Choosely verdict
The useful change is that Argon can now be evaluated through ordinary commercial routes. Existing Pro users can start in Gemini; developers can compare a bounded set of API tasks without buying Ultra solely for access.
Keep the comparison practical: correct results, time spent checking them, latency and total cost. Budget beyond the introductory period before committing a workflow. A cheaper launch rate is welcome. It is not a reason to retire a system that already works.
For the wider choice, see Which AI Model Should You Use in 2026? and Choosely's Google Gemini profile.
Frequently asked questions
Is Gemini 4 released?
Yes. The Gemini API changelog records Argon GA on October 13, 2026. Consumer and integration access depends on the plan and surface described above. API changelog
Is Gemini 4 Pro a separate model?
The documented model is Gemini 4 Argon. Google AI Pro is a subscription plan. Do not confuse that plan name with the model ID. API model directory
Do I need Ultra if I already pay for the Gemini API?
No. Google separately names paid API access and Ultra subscription access. Choose the route appropriate to the product you are using. Google developer rollout
Can I get a million tokens of input and a million of output?
No. The limit is shared across input, thinking and output. Long decode does not double it. Long-decode documentation
Is Argon better than GPT-6 Astra or Claude Opus 5.5?
There is no universal winner established here. Google's published comparison varies by task, and Choosely has not performed a hands-on comparison. Use representative work to decide.
Keep your Gemini stack current
Argon is now available, but model access, billing and usage limits still differ by product.
Choosely Stack Intelligence helps keep the AI tools you rely on visible as models, pricing, access and capabilities change.
The Change Brief
Get the week’s AI changes in one clear read
Pricing moves, tool launches, free-tier changes and practical stack updates, filtered for people who actually use these tools.
Stay ahead of AI without following it all day. We’ll send you what matters each week.
Continue reading
Related reads
Product Update
Gemini 3.8 Live Avatar Gives AI Agents a Face. The Limits Matter More Than the Demo.
Live Avatar is enterprise-focused. Custom avatars are gated. Google says continuous avatar interactions currently last a few minutes rather than hours.
Product Update
Gemini 3.6 Flash Cuts Agent Costs: The AI Tool Changes That Matter This Week
Google's new Gemini Flash tiers sharpen the economics of production agents, while Kimi K3, Grok's Microsoft 365 add-ins and OpenAI Presence move capable models deeper into real work.
AI Strategy
GPT-6 Astra Explained: What the 99.9% ARC-AGI-3 Score Really Means
GPT-6 Astra makes a huge leap in adaptive reasoning and agentic work, but its viral 99.9% ARC-AGI-3 result depends on a provider-specific context-management harness.
AI Tool Recommendations
Which AI Model Should You Use? A Simple Guide to Fast, Thinking and Pro Modes
Start with the normal model and its default reasoning level, then escalate when the task gives you a reason to. A practical guide to Instant, Thinking, effort levels and Pro modes across ChatGPT, Claude and Gemini.
Sources
- Google broad rollout announcement, October 13
- Gemini API pricing
- Gemini app model availability and selection
- Gemini app usage limits and eligibility
- Google developer rollout
- Google AI subscription benefits in AI Studio
- Antigravity model availability
- GitHub Copilot announcement
- Cursor model documentation
- API changelog
- Original September announcement
- Gemini web availability
- Gemini app and macOS announcement
- Long-decode documentation
- Google performance table
- Google evaluation methodology
- Public model card
- Fairwind Program
- API model directory
