Gemini 4 Argon Goes GA, and Its 1M Output Arrives 65,536 Tokens at a Time
Google opened Gemini 4 Argon to everyone on October 13. The $2 per million input price runs only until November 30, and the headline 1M output ceiling is a sum of 65,536-token segments stitched together.
- Gemini 4 Argon is open to any paid API key, and
$2input holds until November 30. - Its 1M output ceiling is a sum of 65,536-token segments, not a single response.
- Every segment resends the previous output as input, so billed prompt tokens keep growing.
Google opened Gemini 4 Argon to everyone on October 13, 2026. A paid API key calls gemini-4-argon starting today, and a Google AI Pro or Ultra subscription exposes it in the Gemini app and in AI Mode in Search. Two weeks ago the terms were different: at the September 30 announcement only cyber defense organizations inside the Fairwind Program and Google's own internal teams could run it, and no API model ID had been published at all.
Google added several hundred more Fairwind partners the same day. Approved defense organizations continue to receive a build with the cyber guardrails removed. The thinking_level section of the developer guide names a separate model ID for it, gemini-4-argon-cyber-permissive, whose access terms are not documented anywhere public.
The number missing from the September 30 announcement: a 1M context limit
The largest blank filled in is the input context window. The September 30 post recorded only that the output ceiling had risen from 64,000 tokens to 1 million, and said nothing about how much the model accepts. The October 13 model page states it for the first time: 1,048,576 tokens, shared between input and output.
Shared means exactly that. Load an 800,000-token codebase and 240,000 tokens of output headroom remain; getting a full million tokens back requires an almost empty prompt. By comparison, Claude Opus 5.5 counts its input window and its output ceiling separately, with output capped at 128,000 tokens.
Reasoning strength is set through thinking_level, with LOW, MEDIUM, and HIGH as the choices. MEDIUM is the default, and only the gemini-4-argon-cyber-permissive build defaults to HIGH. It resembles the knob that flipped per-task cost on Claude Sonnet 5.5 depending on the effort value, with two differences: Argon has no MINIMAL step, and it does not accept thinking_budget for specifying a token budget directly. Both come back as HTTP 400.
The rest of the documented support surface: text goes in and out, while images, audio, and video are input only. Google Search and Maps grounding, code execution, and function calling all work, and computer use, which drives a screen directly, is in preview. On the other side of the line, the Live API for realtime voice, explicit context caching, tuning, and RAG Engine are all unsupported. Caching exists only as implicit caching that triggers on repeated prefixes, and only from 6,144 tokens up.
A million output tokens does not arrive in one response
The sharpest gap between the announcement and the documentation is the output ceiling. Argon does not write a million tokens straight through from one request. It emits roughly 65,536 tokens, stops, then resumes from where it stopped, repeating until the total reaches a million. Google calls this Long Decode.
| Call style | Output ceiling per response |
|---|---|
| Streaming, or any call without background execution | 65,536 tokens |
| Non-streaming direct HTTP call | 32,768 tokens |
Background execution with a raised max_output_tokens | up to 1,048,576 minus input tokens |
The mechanism: when a response hits the segment ceiling, the API returns finish_reason=CONTINUATION along with an encrypted continuation_token holding a hash of the original prompt and the generation state up to that point. Writing the loop yourself means resending the original contents, system_instruction, and tools unchanged while attaching only the latest token, and not appending the partial output you received to contents. The run is finished when no further continuation_token arrives and finish_reason=STOP comes back.
On google-genai 2.30.1 or newer the SDK runs that loop for you. Turning it off means passing GenerateContentConfig(automatic_continuation=False).
How the segments chain for a 20,000-token prompt that requests a long response, counted for a streaming call.
| Segment | Input billed on this request | Output returned | Finish reason |
|---|---|---|---|
| 1 | prompt 20,000 | 65,536 | CONTINUATION |
| 2 | 20,000 + prior output 65,536 = 85,536 | 65,536 | CONTINUATION |
| 3 | 20,000 + prior output 131,072 = 151,072 | 65,536 | CONTINUATION |
| … | grows by 65,536 per segment | 65,536 | CONTINUATION |
| last | all accumulated prior output | remainder | STOP |
Resent prior output gets the 95% implicit cache discount. It applies only when segment calls continue within the 15 to 20 minute cache window, and the shared prefix must be at least 6,144 tokens. Sources: the Gemini API Long Decode documentation and the Gemini 4 Argon developer guide.
The bill is where this lands. Each time a segment chains, the output produced by the previous segment goes back in as input on the next request. Google's documentation states this outright. As segments accumulate, prompt_token_count keeps climbing and total_input_tokens becomes a figure that counts the same content several times over. That is why the docs tell you to read initial_prompt_token_count instead of that number when deciding whether to compress context.
Not all of the resent volume is billed at list price. When segment calls run back to back they land inside the 15 to 20 minute implicit cache window, which puts a 95% discount on the resent portion: $0.10 per million tokens instead of the $2 promotional rate. Cancelling or failing partway through still bills whatever was produced up to that point. Background execution runs for up to 24 hours and requires store: true.
The documentation also lists three conditions under which existing code will not work as written.
- Long Decode runs on the
v1beta1API only, not onv1. - The Interactions API does not support it.
- Older SDKs and the OpenAI compatibility layer cannot resume at all. Long responses get cut off mid-stream.
The last one bites hardest in practice. Teams running a gateway built on the OpenAI-compatible endpoint so models can be swapped out cannot receive Argon's headline feature, long output, through that path.
Independent measurements split between first place and mid-pack
Google's own 19 row comparison table was covered in the October 1 post. Two external measurements have landed since, and they disagree.
Vals AI puts Argon first on its composite index at 68.90%, ahead of Claude Sonnet 5.5 at 67.04%, Claude Opus 5.5 at 66.97%, and Claude Fable 5.1 at 65.83%. It also takes first on Finance Agent v2 at 65.40% and on International Olympiad in Informatics problems at 100%. Cost per test run is $15.68, below Sonnet 5.5 at $21.34 and Opus 5.5 at $32.14.
Inside that same Vals run, tasks that execute a shell and check the result still go the other way. Terminal-Bench 4.0 comes in at 57.58%, sixth of 48 models, and CUA-bench, which drives a screen directly, scores 4.83%, seventh of eight. The weakness the October 1 post flagged using Google's own numbers reappears in the same shape under outside measurement.
Artificial Analysis reaches the opposite ranking. Its composite index scores Argon at 53, tied with GPT-6 Astra and Claude Fable 5.1 and below Claude Opus 5.5 at 58 and Claude Sonnet 5.5 at 56. VentureBeat reported those figures.
Google published internal numbers separately. Teams inside the company using Argon through Antigravity raised agent code submissions by more than 35% and cut post-submission rollbacks by roughly a third. VentureBeat reported that more than 100,000 Google employees use Argon monthly through the internal Antigravity build. In the same piece, Sam Witteveen of Red Dragon AI said it is "a really good model" but that he would not call it the best model in the world, and added that Antigravity itself is not yet anywhere near Claude Code or Codex.
What you can call today, and what changes on December 1
Access depends on which path you come in through.
| Item | Detail |
|---|---|
| Who can use it | Paid Gemini API developers (AI Studio, Android Studio), Google AI Ultra subscribers (Antigravity, AI Studio), Google AI Pro and Ultra subscribers (Gemini app, AI Mode in Search), enterprise customers |
| Free tier | Not available. Also unavailable through the API on the Google AI Pro plan |
| Region availability | The Gemini API and Google AI Studio are available globally, Singapore and the rest of APAC included. Serving region cannot be pinned (see below) |
| Requirements | A paid API key with billing attached. Long Decode additionally needs the v1beta1 API and google-genai 2.30.1 or newer, plus store: true for background runs |
| Extra | Six months of top-tier Antigravity limits free for 10,000 scientists, approved in batches on review and capacity rather than first come |
All four pricing paths double on December 1. Dollars per million tokens.
| Path | Input (through Nov 30) | Output (through Nov 30) | From Dec 1 |
|---|---|---|---|
| Standard | 2.00 | 10.00 | 4.00 / 20.00 |
| Batch | 1.00 | 5.00 | 2.00 / 10.00 |
| Flex | 1.00 | 5.00 | 2.00 / 10.00 |
| Priority | 3.60 | 18.00 | 7.20 / 36.00 |
Cached input goes from $0.10 to $0.20. Google Search grounding is free up to 5,000 requests per month, a quota shared across Gemini 3 and newer models, and $14 per 1,000 requests beyond it.
There is one more condition for anyone with a data residency requirement. The API is reachable from APAC, but the model does not run there. Google's available-regions list includes Singapore and South Korea, so obtaining a key is not the constraint; where inference happens is. Per the Gemini Enterprise Agent Platform model spec, Argon's model availability regions are global, us, and eu, and its machine learning processing regions are us and eu only. Neither asia-southeast1 (Singapore) nor asia-northeast3 (Seoul) appears on either list. If your deployment has to keep processing inside an APAC region, whether because of commitments you have made to customers under Singapore's PDPA or because of a sector rule in finance, healthcare, or the public sector, Argon is not currently a candidate. AWS opened in-region Claude inference on Bedrock in Singapore and Seoul for exactly that constraint, and that is the comparison Argon loses today.
The same Gemini API changelog carried deprecations. As of October 8, gemini-3.7-flash and gemini-3.5-flash are retired and requests route automatically to gemini-3.8-flash and gemini-3.6-flash. The Deep Research agent deep-research-pro-preview-12-2025 shuts down on October 23.
If Argon is on your shortlist, the measurement is worth running before November 30. Upgrade google-genai to 2.30.1 or newer and run one long-output task with max_output_tokens set above 65,536. The numbers to record are not response quality but how many segments chained and how far prompt_token_count climbed. Apply the December 1 rates of $4 input and $20 output to that total and you have the real monthly bill after the promotion ends. If your stack swaps models through an OpenAI compatibility layer, start by confirming that long output gets truncated on that path.