What Is Gemini 4 Argon? Pricing, Coding Tests, and the 1M-Token Catch

What Is Gemini 4 Argon?

Gemini 4 Argon is a native multimodal, sparse mixture-of-experts model with a March 2026 knowledge cutoff. According to Google’s Gemini 4 Argon model card, it is not a fine-tune or modification of an earlier Gemini model. It accepts text, images, video, audio, and PDF input, then returns text.

The model ID is gemini-4-argon. Google lists it as a stable model rather than a preview release.

You can view the current access option on the Gemini 4 Argon API page.

Is Argon the same as Gemini 4 Pro?

No. Google’s public model name and API identifier are Gemini 4 Argon and gemini-4-argon. Treat “Gemini 4 Pro” as an informal description unless Google releases a separate model under that name.

The naming also gives a useful clue about product positioning. Argon is not “Flash with a larger context window.” It is designed to spend more compute on multi-step reasoning, tool calls, and long outputs. That makes the model more capable on some hard jobs, but slower and more expensive on easy ones.

When Was Gemini 4 Argon Released?

Gemini 4 Argon has two dates that can both reasonably be called its release date.

If someone asks for the announcement date, the answer is September 30. If they mean general developer API availability, the answer is October 13. Many early explainers were published between those dates, which is why some still say that Argon has no public API or model ID.

Gemini 4 Argon Is Now Available
Call Gemini 4 Argon with one GPT Proto key and compare it with other frontier models without rebuilding your integration.
Try Gemini 4 Argon

What Is New in Gemini 4 Argon?

The important upgrade is not one isolated benchmark score. It is the combination of a large unified token budget, extended generation, multimodal input, and built-in tool support.

The 1M-token figure is a unified budget

Google gives Argon a 1,048,576-token unified budget. Input, internal reasoning, tool calls, and output all draw from that same allowance. A 700,000-token repository therefore does not leave another one million tokens for thinking and generation.

The normal synchronous maximum output is 65,536 tokens. For longer jobs, Google’s long decode documentation explains that the API continues generation across multiple hops. Each ordinary SDK or streaming hop can produce up to 64K tokens, while direct non-streaming REST calls use a 32K hop limit. The service joins those hops into a longer result.

That continuation system has two practical costs. First, hard tasks can run for several minutes, so Google recommends allowing a client timeout of up to 20 minutes. Second, each continuation hop rebills the accumulated input. Repeated context can receive the 95% implicit-caching discount, but a very long answer is still not cheap.

The takeaway is simple: use long decode when the model must finish a large code change, proof, research synthesis, or agent plan. Do not turn it on for ordinary chat, extraction, routing, or classification.

Argon is built to act, not just answer

The official model documentation lists function calling, code execution, file search, structured output, Google Search and Maps grounding, and computer use in preview. Thinking can be set to low, medium, or high; medium is the default.

That makes Argon a plausible model for agents that inspect a repository, call tools, test a change, and revise their own plan. It does not mean the model will perform equally well in every agent product. The surrounding prompts, tools, permissions, context management, and retry logic can change the result substantially.

An upgrade may require application changes

Moving from an earlier Gemini model is not always a model-name swap. Google advises developers to raise output limits for long jobs, use the new thinking levels, allow longer timeouts, and use background execution beyond 64K output. Explicit caching is not supported; Argon uses implicit caching. Google also says sampling controls such as temperature, top_p, and top_k are ignored.

Those details are easy to miss in a benchmark chart, but they determine whether an upgrade works in production.

Gemini 4 Argon Pricing

Google is using an introductory rate through November 30, 2026, followed by a higher standard rate on December 1. The official pricing page states that output pricing includes thinking tokens.

Google prices cached input at $0.10 per million tokens during the introductory period and $0.20 afterward, a 95% discount from the corresponding list rate. The billing detail to watch is output: internal thinking consumes output tokens even though those tokens may not all appear in the visible answer.

There is no Gemini API Free tier for Argon. Google’s API documentation also says Google AI Pro does not include Argon API access. Consumer access in a Gemini app subscription is a separate product from metered API usage.

The listed GPT Proto route is 40% below Google’s official token prices. Actual job cost still depends on reasoning length, repeated context, tool use, and continuation hops; a lower per-token price does not guarantee a lower bill per completed task.

How to Use Gemini 4 Argon Through GPT Proto

GPT Proto exposes an OpenAI-compatible chat-completions endpoint, so the basic request uses the standard messages structure and the real Argon model ID:

curl --request POST "https://gptproto.com/v1/chat/completions" \
  --header "Authorization: Bearer $GPTPROTO_API_KEY" \
  --header "Content-Type: application/json" \
  --data '{
    "model": "gemini-4-argon",
    "messages": [
      {
        "role": "user",
        "content": "Review this migration plan and identify the three highest-risk assumptions."
      }
    ],
    "stream": false
  }'

Keep the API key in an environment variable, not in source code. Start with a bounded task like the example above, record latency and token use, and only then test repository-scale or long-decode work. Provider-specific features may require different parameters from the basic compatible endpoint, so check the current Gemini 4 Argon API page before relying on long-running jobs in production.

Is Gemini 4 Argon Good for Coding and AI Agents?

The early evidence says yes, with important qualifications.

Google’s model card reports strong results on DeepSWE and long-context GraphWalks, while Argon trails both Astra and Opus on FrontierSWE v2 and trails Opus on Terminal-Bench 4.0.

This is not a clean sweep. My reading is that Argon’s clearest advantage appears when the job needs very long context and extended execution. Opus remains stronger on some terminal and post-training work, while Astra leads Argon on FrontierSWE and terminal science.

A Google engineering case study shows why instructions and tool setup matter as much as the model score. Argon was asked to optimize a Python codebase containing roughly 205,000 lines, 800 source files, and 16,892 tests. On its first attempt, without repository-specific guidance, it introduced a schema-breaking hook, missed edge cases, and ran only 216 nearby tests. After the team added reusable skills and a clear /goal, an unattended 11-hour run executed the full test suite and produced six proposals. Two were accepted and four were rejected. The shipped 22-file change cut session preparation time by about 40% and cold import time by 21%. Read this as a vendor-run production case study, not independent proof.

Independent coding-agent measurements tell a similar story. In Artificial Analysis testing of Antigravity and Claude Code, Argon with Antigravity scored 64 at $5.84 per task and averaged 34.5 minutes; Opus 5.5 at maximum effort with Claude Code scored 66 at $13.04 and averaged 1.1 hours. In its Antigravity versus Codex comparison, Argon scored 64 against 62 for GPT-6 Astra at maximum effort, while Astra finished faster.

Those are comparisons of a model plus its coding-agent environment, not isolated model intelligence. They answer “Which stack completed these tasks better?” rather than “Which model is universally smarter?”

Community reports add texture but should carry less weight. In one same-prompt Three.js test shared on Reddit, Argon finished in 5.5 minutes for $1.58, compared with 14 minutes and $2.54 for Opus. The tester also found Argon less creative and reported that it failed its own page check once. That is one user, one prompt, and one environment—not a benchmark—but it captures the tradeoff: faster and cheaper can coexist with weaker taste or self-verification.

Gemini 4 Argon vs GPT-6 Astra, Claude Opus 5.5, and Gemini 3.8 Flash

The best model depends on what you are optimizing. The following independent comparison uses Artificial Analysis’s Intelligence Index and its estimated cost per evaluated task; the token prices are list prices during Argon’s introductory period.

Gemini 4 Argon vs GPT-6 Astra

Argon leads Astra 53 to 51 on the Artificial Analysis model comparison, costs far less per token, and produces much longer answers in the evaluation. Yet its estimated cost per task is slightly higher—$1.99 versus $1.73—because Argon used about 62,000 output tokens per task compared with roughly 12,000 for Astra.

So, is Gemini 4 Argon better than Astra? For million-token context and exhaustive work, often yes. For science-heavy tasks, shorter completions, or jobs where latency matters, GPT-6 Astra can be the better choice. “Cheaper tokens” and “cheaper task” are not synonyms.

Gemini 4 Argon vs Claude Opus 5.5

Claude Opus 5.5 leads Argon 58 to 53 on the independent Intelligence Index and leads on Terminal-Bench 4.0 in Google’s own model card. Argon has the lower list price and a much lower estimated evaluation cost per task.

Choose Opus when terminal reliability, nuanced judgment, or its established Claude Code workflow matters more than price. Choose Argon when long context, multimodal input, and extended generation are central to the job.

Gemini 4 Argon vs Gemini 3.8 Flash

Argon scores 53 versus 41 for Flash on the Artificial Analysis comparison, but the introductory Argon price is about 2.7 times higher for both input and output. Estimated evaluation cost is $1.99 per task for Argon and $1.24 for Flash.

Use Gemini 3.8 Flash for classification, extraction, routine support, lightweight generation, and high-volume workloads. Route only the difficult cases to Argon. That design is usually more economical than replacing Flash across an entire application.

One Key, Four Frontier Models
Compare Argon, Astra, Opus, and Gemini Flash through the same GPT Proto account before committing your production workload.
Compare Models on GPT Proto

Should You Upgrade to Gemini 4 Argon?

Upgrade when your current model loses track of large repositories or document sets, stops before completing a complex deliverable, or needs repeated handoffs to finish a multi-tool workflow. Argon is also worth testing when image, audio, video, PDF, and code context must be considered together.

Keep a smaller model when requests are short, predictable, and frequent. Flash remains the more sensible default for high-volume extraction, tagging, routing, short summaries, and simple chat. Keep Astra or Opus in the evaluation pool when their stronger benchmark areas match your workload.

I would put Argon in a model router before making it the default for every request. Send hard, long-context jobs to Argon; keep routine traffic on Flash; and compare Astra or Opus on the task categories where they lead. Production traces—success rate, human corrections, latency, and total cost per accepted result—should decide the final route.

Build With Gemini 4 Argon for 40% Less
Use one GPT Proto key to call Gemini 4 Argon and switch to Astra, Opus, or Flash when your workload changes.
Get Gemini 4 Argon API Access