Google DeepMind's frontier reasoning model above the Gemini 3.8 Flash line, built for long coding, knowledge-work and cyber-defense jobs.
As of 15 October 2026, Gemini 4 Argon is generally available: it reached the paid Gemini API this week and is rolling out to Google AI Pro and Ultra subscribers in the Gemini app. It suits long codebase migrations and legal or finance research, where Vals AI scored it 68.9% on its cross-industry index.
Checked 30 Oct against Google DeepMind's own pages.
| Job | Input | Output | Total |
|---|---|---|---|
| Summarise a 20-page PDF | $0.060 | $0.010 | $0.070 |
| Support reply | $0.0040 | $0.0030 | $0.0070 |
| One coding agent run | $0.400 | $0.200 | $0.600 |
Budgets: 20-page PDF = 30k in / 1k out · Support reply = 2k in / 300 out · Coding agent run = 200k in / 20k out. Computed from the vendor's per-token prices at render time; cached-input discounts are not applied.
Gemini API paid tier only; Google lists no free tier. Standard rates are $2.00 input and $10.00 output per 1M tokens (output includes thinking tokens) through 30 November 2026, then $4.00 and $20.00 from 1 December 2026. Context caching is $0.10 then $0.20 per 1M, implicit only; explicit caching is not supported. Batch and Flex inference run at $1.00 input and $5.00 output per 1M through 30 November, then $2.00 and $10.00. Priority inference is $3.60 and $18.00, then $7.20 and $36.00. Grounding with Google Search and Google Maps gives 5,000 free requests a month shared across Gemini 3 and newer models, then $14 per 1,000. Google states that paid-tier data is not used to improve its products. Cost per job runs higher than the per-token rate suggests because of long reasoning traces; compare it with GPT-6 Astra and Gemini 3.8 Flash on a real task.
Gemini 4 Argon is Google DeepMind's frontier reasoning model, scoring 53 with high reasoning on the Artificial Analysis composite of ten agentic, coding and knowledge tests. It reads text, images, video and audio, writes text, and stands out for calibration: its AA-Omniscience hallucination rate is 15%, so it usually declines rather than guesses.
Reads
Writes
One-million-token output limit
A single response can run to 1M output tokens, up from 64K on earlier Gemini models, so a long migration or report can finish in one trajectory.
Long Decode Continuation
A new Gemini API mechanism pauses a very long response and resumes it through follow-up requests, so extended reasoning does not hit a request timeout.
Cyber defense mode
Trained to find, validate and patch software vulnerabilities, it ties for first on CWE-bench v1 at 68% in Google's results, and Fairwind partners can run it inside the CodeMender security agent.
Long video understanding
Google reports 91.7% on LVBench, a test of locating details across long videos, alongside chart and multi-document analysis.
Multimodal input, text output
Takes text, images and video as input and answers in text. Artificial Analysis's launch analysis also lists speech input, though its model page lists only text and image, and Google's announcement names video but not audio.
output tokens in a single response, up from 64K on earlier Gemini models
Source: Google announcement, 30 Sep 2026Pricier than 73% of the 104 GA models with a published price, and one of 41 that document a zero-data-retention option.
| Measure | This model | Rank | Median |
|---|---|---|---|
| Blended price (3:1)$ per 1M tokens | $4.00/M | #71 / 104 | peer median $1.73/M |
Ranks are against GA models on Hokai that publish the same figure; ties share a rank.
A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.
| Model | Lab | Context | In / out per 1M | Reads and writes |
|---|---|---|---|---|
| 1M | $2 / $10 | text, image, video, audio → text | ||
| OpenAI | 1.1M | $10 / $50 | text, image → text | |
| Anthropic | 1M | $4 / $20 | text, image, pdf → text | |
| Anthropic | 1M | $2 / $10 | text, image, pdf → text |
| Name | Endpoint | Auth |
|---|---|---|
| -- | -- | REST |
Spends many tokens per task
Artificial Analysis counted about 62,000 output tokens per Intelligence Index task, versus about 27,000 for GPT-6 Astra. Per-token price comparisons understate its real cost.
Workaround: Budget by measured cost per task on your own workload, and cap maximum output tokens on routine requests.
Long answers arrive in pieces
Responses that run toward the 1M-token limit are paused and resumed through Long Decode Continuation rather than returned in one call.
Workaround: Write your client to detect a paused response, issue the follow-up request, and persist partial output between calls.
Abstains more than it recalls
Its low hallucination rate comes with 50% accuracy on AA-Omniscience, so it often declines questions a less cautious model would answer correctly.
Workaround: Ground factual lookups with retrieval or search, and treat an abstention as a cue to fetch a source.
Price doubles on 1 December 2026
Google lists $2 and $10 per 1M tokens through 30 November 2026, then $4 and $20. Batch, Flex and Priority double on the same date.
Workaround: Model your costs at the standard rate so the switch does not break a budget.
Gemini 4 Argon (general availability) · 13 Oct 2026
Generally available on the paid Gemini API as gemini-4-argon (Gemini API release notes) and rolling out to Google AI Pro and Ultra subscribers in the Gemini app (Gemini app release notes).
Gemini 4 Argon (Fairwind preview) · 30 Sep 2026
Announced; restricted to Fairwind Program partners and Google internal teams. Output limit raised to 1M tokens.
Gemini 3.1 Pro (predecessor, non-Flash) · 19 Feb 2026
Previous Google frontier model above the Flash class; Artificial Analysis Index v4.3.2 score of 30 as Gemini 3.1 Pro Preview.
There is no free API tier. Through the end of November 2026 the paid tier charges $2 per million input tokens and $10 per million output tokens, with implicit caching at $0.10; from 1 December it moves to $4 and $20, with caching at $0.20. Batch and Flex run at half the standard rate. On Artificial Analysis testing a task costs $1.99 now and $3.98 at the standard rate, against $3.26 for GPT-6 Astra.
Not through the API, where Google lists the model for the paid tier only. In the Gemini app, Google's release notes name Google AI Pro and Ultra subscribers as the rollout audience, so a plain free account is not listed. Google AI Studio offers a place to try the gemini-4-argon model ID, billed at the paid-tier rates.
On the overall Artificial Analysis score they tie, so behaviour decides it. Argon is far more cautious about inventing facts and does better on business-workflow automation, while Astra answers more factual questions correctly and uses fewer than half the output tokens per task. Teams should run both on their own workload, since cost per task, not per token, is what differs.
Google announced it in late September 2026 and gave it first to vetted security teams in the Fairwind Program, which works with over 650 partners. General availability on the Gemini API followed in mid-October 2026, with a global Gemini app rollout for AI Pro and Ultra subscribers the same day. Fairwind partners still get a version without the cyber guardrails.
Use it for long document-heavy work: large code migrations, legal drafting or financial research, where a single pass can now run far longer than before. Look elsewhere if you need terminal-driven coding agents (Claude Sonnet 5.5 scores higher on Terminal Bench 4), audio or image output, or a fixed budget that cannot absorb the 1 December price change.
Gemini 4 Argon is the frontier reasoning model from Google DeepMind, announced in late September 2026 by Koray Kavukcuoglu, Google's Chief AI Architect. It is the first Gemini 4 model Google has announced, and Artificial Analysis describes it as Google's first proprietary model above the Flash class in more than seven months, after Gemini 3.1 Pro. Google has not disclosed the architecture, parameter count or training cutoff. Its API documentation lists a unified context window of 1,048,576 tokens and an output limit of 65,536 tokens by default, rising to 1,048,576 with long decode. The rest of the lineup sits on the Google DeepMind models page.
Access was staged and has now opened. Argon first went to vetted security teams in Google's Fairwind Program, who get a version with the cyber guardrails removed for vulnerability research, and that restricted route still exists. On 13 October 2026 Google's Gemini API release notes marked the model generally available under the ID gemini-4-argon, on the paid tier only, and the same day the Gemini app release notes said it was rolling out globally to Google AI Pro and Ultra subscribers. The consumer route is the Gemini app, and developers can try the model ID in the AI Studio playground. The API takes text, image, video, audio and PDF input and returns text.
On Google's own evaluations, Argon scores 77.9% on DeepSWE v1.1, a long-horizon software engineering test, and 51.3% on Zapier's AutomationBench. These are vendor-run numbers. The independent picture from Artificial Analysis is close to Google's story but less flattering at the top: its composite index puts Argon level with GPT-6 Astra and Anthropic's Fable 5.1, and a point ahead of GPT-6.1 Sol, while Claude Opus 5.5 sits at 58. That is a 23-point jump over Gemini 3.1 Pro Preview on the same index. On the AA variant of AutomationBench it scores 78%, and on Terminal Bench 4 it scores 57%, behind Claude Sonnet 5.5 at 64%, Claude Opus 5.5 at 60% and GPT-6 Astra at 59%.
What Argon has already done inside Google is the strongest evidence so far, though it is Google's account. Argon agents replaced 32,000 lines of SIMD code in the Rust port of the libgav1 video decoder, producing safe Rust that the vendor reports runs 2.7 times faster than the earlier port with identical output. Another group of agents worked through fleet-wide profiling data to find memory optimizations across Google's data centers; Google says they will free more than 300 TiB once rolled out, with an estimated 500 TiB to 1 PiB in total savings. On the security side, Google reports that Wiz, running it through its Scan for Good program, found a critical flaw in hospital software that earlier frontier models had missed.
The non-obvious cost is token volume. Artificial Analysis measured an average of about 62,000 output tokens per Intelligence Index task, more than twice what GPT-6 Astra used, so its low per-token rate does not translate one for one into a cheap job. Teams comparing it with Gemini 3.8 Flash or the earlier Gemini 3.8 Flash Cyber should price a real task, not a token. Google pairs the release with safety work it describes in some detail: activation monitoring to spot misuse, chain-of-thought monitoring that can stop an agent mid-run, refusal training for chemical, biological, radiological, nuclear and offensive cyber requests under its Frontier Safety Framework, and adversarial training against indirect prompt injection.
Who should care now: engineering, legal and finance groups that want long single-pass work, and security teams, whether through Fairwind or the standard API. Who should not rush: teams that need audio or image output, which the model does not produce, or a model with a long production track record; the release is days old. For those cases a model that has been stable for longer, such as the ones behind the Gemini app, or a rival frontier model is the practical choice. Our walkthrough on how to choose the right LLM covers that decision by job, the guide to the previous round of frontier launches shows where the previous Gemini stood, and our analysis of Google's Gemini line covers the consumer side. To compare like with like, browse models that accept video or models with contexts above 400K tokens.
Specifications on this page (context window, token prices, benchmark scores, licence) are read from Google DeepMind's own model card and documentation. The verdict is Hokai's reading of them, not the provider's claim.