Gemini 4 Argon Is Now Available to Developers and Subscribers: Benchmarks, Pricing, and Where to Use It

Google made Gemini 4 Argon generally available on October 13, 2026. Developers and enterprises get it through the paid Gemini API in AI Studio and the Gemini platform. Google AI Ultra subscribers get it in Google Antigravity, and Ultra and Pro subscribers get it in the Gemini app and AI Mode in Search (Google). Google first announced Argon on September 30, when it began rolling out to trusted cyber defenders in its Fairwind Program (Google's first Argon post).

Argon is built for long, hard tasks. It ranks first on the Vals Index at 68.90%, ahead of Claude Sonnet 5.5 at 67.04% and Claude Opus 5.5 at 66.97% (Vals AI). Google raised its output token limit to 1M tokens, up from the previous 64K (Google's first Argon post).

The verdict: try Argon now for large refactors, long-horizon coding tasks, and financial research, while the introductory price of $2 input and $10 output per million tokens lasts through November 30 (Gemini API pricing). Don't switch your terminal agent or your language-porting work yet. On Vals AI's own runs, Claude Opus 5.5 still leads Terminal-Bench 4.0 and Claude Sonnet 5.5 leads Code Migration (Vals AI).

Key Takeaways

• Available: October 13, 2026, globally, in the paid Gemini API, Antigravity for Ultra, and the Gemini app and AI Mode for Pro and Ultra (Google).
• Model ID: gemini-4-argon, a stable model with a 1,048,576-token window shared by input and output (Gemini API docs).
• Price: $2 input and $10 output per million tokens through November 30, 2026, then $4 and $20 (Gemini API pricing).
• Not on free or Pro API access: Argon is not available on the Gemini API free tier or on the Google AI Pro plan (Gemini API pricing).
• Long decode: output up to 1,048,576 tokens, less your input, instead of the default 65,536 (Gemini API docs; long decode docs).
• Benchmarks: first on the Vals Index at 68.90%, and 77.9% on DeepSWE v1.1 by Google's count (Vals AI; Google's first Argon post).
• Where it trails: of the eight Vals Index components, Argon leads only Finance Agent v2; Claude Opus 5.5 leads Terminal-Bench 4.0, and Claude Sonnet 5.5 leads Vibe Code Bench and Code Migration (Vals AI).
• Knowledge cutoff: March 2026 (model card).

Who Gets Gemini 4 Argon and Where

Gemini 4 Argon is open to paid API customers and to Google AI Ultra and Pro subscribers, but each group gets a different set of surfaces. Antigravity needs Ultra or an eligible business plan, and the Gemini API needs the paid tier. Pro gets the consumer apps only.

Google AI Ultra subscribers can also reach Argon from the latest Android Studio Canary by signing in to Antigravity in the agent tool window (Android Developers Blog). In Antigravity, Google says to update to the latest version to get the model (Antigravity).

Pricing: Introductory Rates Through November 30

Argon's API price doubles on December 1, 2026. Until then it costs $2 per million input tokens and $10 per million output tokens, with thinking tokens billed as output. Plan any large job you want to run on it before the change.

Source: Gemini API pricing. Caching is implicit only; Argon does not support explicit caching. Grounding with Google Search includes 5,000 free requests a month, shared across Gemini 3 and newer models, then costs $14 per 1,000 (Gemini API pricing).

On the Vals Index, Argon cost $15.68 per test at the $4 and $20 rates, against $21.34 for Claude Sonnet 5.5 and $32.14 for Claude Opus 5.5 (Vals AI). It scored higher than both and cost less to run.

Benchmarks: Where Argon Leads and Where It Trails

Argon leads most of the benchmarks Google highlights, and it tops Vals AI's independent index overall. Inside that index, though, it leads only one of eight components: Claude models lead the terminal, app-building, and code migration tests. The table below uses only scores published in text by Vals AI or Google.

Google computed the DeepSWE v1.1 score itself with a mini-swe agent harness, and ran its own scores at the highest thinking setting (Google evaluation methodology). Vals AI notes that some Claude component tasks fell back to another model after a provider refusal, while Argon had no fallbacks (Vals AI).

Specs and Long Decode

Argon takes text, images, video, audio, and PDF, and returns text. Its 1,048,576-token window is shared by input and output (Gemini API docs).

• Max output: 65,536 tokens by default, up to 1,048,576 with long decode (Gemini API docs).
• Thinking levels: low, medium, and high (Gemini API docs).
• Supported: function calling, structured outputs, code execution, search grounding, URL context, file search, and computer use in preview (Gemini API docs).
• Not supported: explicit caching, the Live API, and image or audio generation (Gemini API docs).
• Consumption options: Batch, Flex, and Priority inference (Gemini API docs).
• Architecture: a sparse mixture-of-experts model, not a fine-tune of an earlier Gemini (model card).

Long decode is how you get past 64K tokens of output. On the Interactions API, Google's server-side option, you create the request with background: true and a max_output_tokens above 65,536. The server then generates the response in 64K hops, each continuing where the last stopped, and you poll or stream the interaction to read the joined result (long decode docs).

Watch out: the model card lists "occasional slowness or timeout issues" among Argon's known limitations (model card). Use background execution for long jobs rather than holding a synchronous connection open.

What Google Says Argon Did Inside Google

Google's case for Argon rests on internal engineering results it reports itself. Treat them as vendor claims, not independent measurements.

• Antigravity at Google: over 100K Alphabet users run Argon in Google's internal Antigravity each month; over four weeks, Google reports a 35% increase in agentic code and a 30% decrease in post-submit rollbacks (Antigravity).
• Agent Development Kit: working overnight from a one-paragraph goal, Argon cut in-memory session prep time by 40% in long sessions and sped up cold imports by 21% in google/adk-python (Google).
• Gemma inference: a 2.85x end-to-end inference speedup for Gemma 4 E2B on an A100 via vLLM, with 100% token parity (Google).
• Bug backlog: in under 3 hours, Argon closed 97 of 134 open bug reports as stale or duplicate, with documented commit proof, and sorted the remaining 37 by owner (Antigravity).
• Rust optimization: Argon agents took an existing Rust port of the libgav1 video decoder and replaced 32K lines of SIMD code, producing a memory-safe decoder that runs 2.7x faster than the port (Google's first Argon post).

Safety and the Cyber Program

Google's model card says it cannot rule out that Argon reached its critical capability levels for biological information risk and for cybersecurity. Google developed safety cases for both and says its safeguards lower the risk below those levels (model card).

• Prompt injection: Argon leads the Gray Swan Indirect Prompt Injection benchmark, with attack success at 0.7% across 15 adaptive attempts (Google).
• Shortcut-taking: on coding benchmarks, shortcut attempts dropped to 2%, Google's best result across its Gemini models (Google).
• Cyber defenders: members of the Fairwind Program can get Argon without cyber guardrails, and Google is expanding the program to hundreds of additional partners (Google).
• Scientists: Google is offering 10,000 scientists Antigravity's highest usage tier free for 6 months, subject to eligibility, capacity, and selection (Google).

When to Choose Gemini 4 Argon

Choose Argon when a task is long, deep, and benefits from a very large output: big refactors, long-horizon coding tasks, financial research, and long-video work. Keep a Claude model for terminal-driven agent loops, quick app builds, and porting code between languages, where Vals AI's runs still favor Opus 5.5 and Sonnet 5.5.

For the small, cheap end of the market, see our Claude Haiku 5.5 coverage; for earlier Gemini Flash coverage, see Gemini 3.6 Flash and Flash Cyber.

Sources

• Google, "Gemini 4 Argon: Built to solve your hardest problems," October 13, 2026: https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-argon/
• Google, "Gemini 4 Argon: our next era of frontier intelligence," September 30, 2026: https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/
• Google AI for Developers, "Gemini 4 Argon": https://ai.google.dev/gemini-api/docs/models/gemini-4-argon
• Google AI for Developers, "Gemini Developer API pricing": https://ai.google.dev/gemini-api/docs/pricing
• Google AI for Developers, "Long decode": https://ai.google.dev/gemini-api/docs/long-decode
• Google DeepMind, "Gemini 4 Argon Model Card": https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-4-Argon-Model-Card.pdf
• Google DeepMind, "Gemini 4 Argon model evaluation": https://deepmind.google/models/evals-methodology/gemini-4-argon
• Google Antigravity, "Gemini 4 Argon in Google Antigravity," October 13, 2026: https://antigravity.google/blog/gemini-4-argon-in-google-antigravity
• Android Developers Blog, "Bringing our latest frontier model to Android Developers": https://android-developers.googleblog.com/2026/10/bringing-gemini-4-argon-to-android-agentic-engineering-antigravity-cli.html
• Vals AI, "Vals Index": https://www.vals.ai/benchmarks/vals_index