Gemini 4 Argon API

Gemini 4 Argon is Google DeepMind's new frontier model, and the first model of the Gemini 4 generation. Google built it for work that takes a long time to get right: real software engineering, legal and financial analysis, and defensive cybersecurity. It reads up to about a million tokens of text, images, audio and video, it can think through hundreds of thousands of tokens in a single run, and on Google's own comparison table it beats GPT-6 Astra and Claude Opus 5.5 in more categories than it loses.

That last point matters because Google has spent most of 2026 chasing. Its previous frontier family arrived in November 2025, and while OpenAI and Anthropic shipped several flagship updates, Google's next big Pro model was reported as delayed. Argon is the answer. It was announced on September 30 for a small group of cyber defenders, and opened to everyone on October 13 through the Gemini API, AI Studio, the Gemini app and Google's business products.

The headline numbers are a 1,048,576-token context window, an output limit that grows from 64K tokens on the previous generation to as much as a million, and first place in Google's table on legal, finance, automation, long-context and video benchmarks. The catch is equally clear: Argon is a heavy thinker, and it uses more tokens and more time than its rivals to get there.

Google DeepMind's first Gemini 4 model, built from scratch

Gemini 4 Argon comes from Google DeepMind, the company's merged research lab, and was introduced by Koray Kavukcuoglu. Google describes Gemini 4 as "the next generation in the Gemini series", a suite of natively multimodal reasoning models, and Argon is the first member to ship.

The model card makes a point that is easy to miss: Argon "is not a modification or a fine-tune of a prior model." This is a fresh pre-training run, not Gemini 3 with more post-training on top. The direct comparison point Google uses throughout its safety evaluations is Gemini 3.1 Pro, the February 2026 update to the Gemini 3 line, and the catalog's older Gemini 3 Pro is the model many teams still run today.

Argon is a closed model. Google has not released weights and has not published a parameter count. Access runs through Google's own channels: the Gemini app (for Google AI Pro and Ultra subscribers), AI Mode in Search, Google Antigravity for coding (Ultra), Gemini for business, AI Studio and the Gemini API. Use is governed by the Gemini API Additional Terms of Service and Google's Generative AI Prohibited Use Policy.

The staged launch was deliberate. For two weeks Argon went only to partners in Google's Fairwind Program, a set of trusted security teams, and through the U.S. government's voluntary pre-release access process. Those defenders receive a version without the usual cyber guardrails, so they can use its full ability to find and fix vulnerabilities.

What changed from Gemini 3.1 Pro

A million tokens out, not just in. Gemini models have read a million tokens for a while. What is new is how much Argon can write back. Google calls its output limit "industry-leading 1M tokens, up from the previous 64K tokens." On the Gemini API the default output cap is 65,536 tokens, and a long-decode mode raises it to the full 1,048,576-token window. Google's argument is that depth of reasoning grows when the model "has the headroom to think deeply and generate hundreds of thousands of tokens in a single trajectory."

Trained for defensive security. Argon was trained specifically for cyber defence, and Google says it can autonomously find, validate and patch critical software vulnerabilities. It ties GPT-6 Astra for first place on CWE-bench v1 at 68.0%, ahead of Claude Opus 5.5 at 67.0%. In an early test, the security firm Wiz used it to find a critical flaw in healthcare software used by hospitals, one Google says earlier frontier models had missed.

Harder to hijack. Google calls Argon its most resilient model against indirect prompt injection. On the Gray Swan benchmark, attackers succeeded 0.7% of the time, against 1.0% for Claude Opus 5.5 and 8.5% for GPT-6 Astra. For agents that read web pages, e-mails and documents written by strangers, that gap is meaningful.

Fewer pointless refusals. Against Gemini 3.1 Pro, Google's automated tests show unjustified refusals down 9.6 percentage points, a more neutral refusal tone (up 5.5 points), and a 9.29-point improvement in safety on image-plus-text prompts. In the Gemini app Google says Argon "answers questions more directly with less hedging, doesn't repeat itself more than necessary, and skips clichés."

Proof from inside Google. Google says its own teams already work with Argon. It replaced 32,000 lines of SIMD code in the libgav1 video decoder, producing a decoder 2.7 times faster than the earlier Rust port, and Chrome engineers using it raised agentic code submissions by more than 35% while cutting rollbacks by nearly a third.

Inside Argon: a sparse mixture of experts

The model card describes Argon as a large-scale sparse mixture-of-experts transformer with native support for text, vision and audio inputs. In a mixture-of-experts model only a fraction of the network's weights run for each token, which keeps the compute per token manageable even for a very large model. Google has not said how many experts there are or how large the model is.

Pre-training used a broad mix of public web documents, code, images, audio and video, plus licensed data, data from Google products under their terms, and synthetic data. Post-training combined instruction tuning, human preference data and reinforcement learning on multi-step reasoning, problem-solving and theorem-proving data. Training ran on Google's TPUs with JAX and ML Pathways. The knowledge cutoff is March 2026.

Gemini 4 Argon specs

Argon does not generate images and has no realtime voice mode. On Google's API it supports implicit caching only.

Where Argon leads: knowledge work, long context and video

Google tested Argon against GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5 on 18 benchmarks. By VentureBeat's count it leads outright on 12 and ties for first on one more. The wins cluster in a few areas.

Professional knowledge work. Argon scores 68.9% on the Vals Index, which covers finance, coding, legal and tax tasks, ahead of Claude Opus 5.5 at 67.0%. On Vals Finance Agent v2 it reaches 65.4% against 58.6%. The biggest gap is Harvey's Legal Agent Benchmark: 19.6% for Argon, 5.4% for GPT-6 Astra and 3.8% for Opus 5.5. The absolute score is low, which says the test is hard, but nobody else is close.

Business automation. On Zapier's AutomationBench, which asks a model to carry out complete business processes across apps, Argon scores 51.3% against 42.5% for Opus 5.5 and 41.4% for GPT-6 Astra.

Very long context. GraphWalks asks a model to follow links through a large graph described in text. Up to 128K tokens Argon is near perfect at 99.7%. Between 256K and 1M tokens it scores 84.2%, while GPT-6 Astra drops to 71.8% and Opus 5.5 to 66.8%. If your workload is a whole codebase, a data room or a year of contracts in one prompt, this is the number to look at.

Video and charts. Argon reaches 91.7% on LVBench, a long-video understanding test, and 71.6% on Chartography. Google pitches it for professional chart analysis and finding details in long videos.

Science and maths. It leads LABBench 2 at 88.8% and RiemannBench at 76.0%.

Writing. Dan Shipper of Every called Argon a strong structural editor: asked about a weak metaphor, it explained what the metaphor was trying to do, offered four possible emphases and flagged what would need to be cut. Testers on Hacker News described its answers as to-the-point and logically structured, "like a really well made mini essay."

Where it falls short

Agentic coding in a terminal. Argon wins DeepSWE v1.1 with 77.9%, but it is not the best coding agent across the board. On FrontierSWE v2 it scores 55.0% against GPT-6 Astra's 65.5%. On Terminal-bench 4.0 Claude Opus 5.5 leads with 66.4% to Argon's 57.4%, and on Terminal-Bench Science GPT-6 Astra leads 68.1% to 57.6%. Opus 5.5 also wins PostTrainBench, and GPT-6 Astra edges it on the OSWorld-2.0 computer-use subset.

Token appetite and latency. Argon thinks a lot. Artificial Analysis, which places it eighth of 214 models on its Intelligence Index with a score of 53, counted 110 million output tokens to run that index and measured a time to first token of about 48 seconds at 81.3 output tokens per second. Early users complain that it deliberates at length over simple questions, and some reported runs where the reasoning looped. Google's model card itself warns of "occasional slowness or timeout issues." Because thinking tokens count toward max_tokens, long deliberation also eats into the room left for the visible answer.

Safeguards around risky domains. Under Google's Frontier Safety Framework, the company says it cannot rule out that Argon reaches the critical capability levels for biological information and for cyber offence, and it has added safeguards in response. Expect careful handling, and sometimes refusals, for requests that drift toward attack tooling or biology protocols.

The usual limits. It can hallucinate, its knowledge stops in March 2026, and it produces text only.

How to prompt Gemini 4 Argon

Argon rewards complete briefs. Give it the whole document set, state the deliverable and the format, and tell it what "done" looks like. Because it reasons at length by default, use reasoning_effort: "low" for quick, well-defined tasks and keep high for problems that need real search.

For a legal or finance review, put the documents first and the question last:

Here are the share purchase agreement and the disclosure letter. List every warranty that the disclosure letter qualifies, quote the qualifying sentence, and rate the buyer's residual risk as low, medium or high with one line of reasoning each. Return a Markdown table.

For defensive code review across a large repository:

You are reviewing this service for security flaws before release. Trace every path where user input reaches a database query or a shell call. For each finding give the file, the line, a proof-of-concept input, and a minimal patch. Ignore style issues.

For video or chart analysis, ask for timestamps or values you can check:

Watch this 40-minute product walkthrough. List each feature shown, the timestamp where it first appears, and any moment where the presenter's claim does not match what is on screen.

Set max_tokens with room for thinking. A reply that needs 3,000 visible tokens may spend many more on reasoning, and the cap includes both.

Gemini 4 Argon versus Claude, GPT and older Gemini models

Claude Opus 5.5 is the closest rival on Google's own table. Argon wins on knowledge work, long context and video, Opus wins on Terminal-bench 4.0 and PostTrainBench. Argon also edges Opus on DeepSWE v1.1 (77.9% against 74.2%) and is slightly harder to hijack through prompt injection.

GPT-6 Astra is the pick for hard terminal-style and frontier software engineering, where it leads FrontierSWE v2 and Terminal-Bench Science, and it edges Argon on the OSWorld-2.0 computer-use subset (72.6% against 69.2%). It is also far less talkative: one review put its average at about 27,000 output tokens per task, against about 62,000 for Argon.

Claude Sonnet 5.5 is worth testing side by side for everyday coding and writing. For high-volume work where a frontier model is overkill, a Flash-tier model such as Gemini 3.8 Flash answers far faster and deliberates far less. Teams already on Gemini 3 Pro get the most direct upgrade path: same multimodal inputs, much longer outputs, and a new model underneath.

Getting started

Open Argon in the Playground to try a prompt without writing code. When you are ready to build, create a key on the API keys page and call the OpenAI-compatible /v1/chat/completions endpoint with model: "gemini-4-argon". Images and video go in as image_url message parts, tools use the standard tools array, and response_format: { "type": "json_object" } returns JSON. The gemini-4-argon-openai id is a second route to the same model with the same parameters, which makes it a natural fallback. The docs cover streaming, tools and error handling.

Verdict

Gemini 4 Argon puts Google back at the front of the frontier race, at least on the work Google chose to measure: legal and financial analysis, business automation, million-token context and video. It is a new model rather than a retune, it resists prompt injection better than its rivals, and it beats Claude Opus 5.5 on most of Google's knowledge-work table. It is not the best terminal coding agent, and its habit of thinking long and starting slowly makes it a poor fit for quick chat. For deep, document-heavy and security-minded work where accuracy matters more than speed, it is the model to try first.