Gemini 4 Argon — online chat, and the five rows its own model card does not win

Five of the nineteen rows, and the leading figures are printed by the model card itself. On FrontierSWE v2 this model reads 55.0% while GPT-6 Astra reads 65.5%, Claude Opus 5.5 62.3% and Claude Fable 5.1 56.3%, which places it fourth — 10.5 percentage points behind the leader. On Terminal-bench 4.0 it reads 57.4% against Claude Opus 5.5’s 66.4%, GPT-6 Astra’s 58.2% and Claude Fable 5.1’s 57.9%, which is fourth and 9.0 points off the lead. On Terminal-Bench Science 0.1 it reads 57.6% against GPT-6 Astra’s 68.1% and Claude Opus 5.5’s 63.3%. On PostTrainBench it reads 45.3% against Claude Opus 5.5’s 49.3%, and on OSWorld-2.0 it reads 69.2% against GPT-6 Astra’s 72.6%. Across all nineteen rows that is thirteen outright firsts, one tie for first, two second places, one third and two fourths. Those five rows are exactly why the comparison above is written against GPT-6 Astra row by row rather than as a claim about the whole table.

Google describes it as a new frontier model, and says so in the model card’s dependencies section: Argon “is not a modification or a fine-tune of a prior model.” That is a deliberately different sentence from the one used for the Flash releases before it, which were documented as building on the generation below them. The architecture section describes a large-scale sparse mixture-of-experts transformer with native multimodal support for text, vision and audio, trained with reinforcement learning techniques that can use multi-step reasoning, problem-solving and theorem-proving data, and the hardware and software sections name Google’s TPUs with JAX and ML Pathways. The models index files it under a new Gemini 4 family marked stable, and the model page carries a single stable version line with no preview row.

Google publishes one shared window rather than two separate limits: input, thinking and output all draw from 1,048,576 tokens. The output ceiling is 65,536 tokens by default for synchronous requests and up to 1,048,576 with long decode, which the official guide defines as up to one million tokens minus your input, delivered in hops of up to 64K tokens through background execution. Two conditions are easy to miss. If you enable background execution without setting max_output_tokens, the server stops at 64K; and max_output_tokens caps thinking plus output together, so a value set too low can truncate the answer. On this site the picture is different again: each reply is requested with at most 4,096 output tokens, which is a workspace setting, and because the catalog entry behind this page publishes no context length the interface shows no context figure at all. Ask for long deliverables in parts.

Google’s pricing page quotes four tiers, and every tier carries two figures with two dates. Standard input is $2.00 per million tokens through November 30, 2026 and $4.00 starting December 1, 2026; Standard output is $10.00 and then $20.00, and a cache read $0.10 and then $0.20. Batch and Flex both sit at $1.00 and $5.00 with cache reads of $0.05 and $0.10, and Priority, the dearest tier, at $3.60 and $18.00 with a cache read of $0.18, each doubling on the same date. The release blog states the same $2.00 and $10.00 as an introductory price but attaches no date. Caching is implicit only on all four tiers, and the official guide says implicit caching is on by default for prompts of at least 4,096 tokens with a 95% discount on cached input tokens. Google also states that this model is not available on the free tier or on the Google AI Pro plan, so every tier here is a paid one. This site does not use those dollar figures: it bills in points, and its catalog reference price is its own basis.

On this site it is always on and it cannot be changed, and that is a difference between the catalog entry behind this page and the official model. Google documents thinking as supported at low, medium and high with a default of medium, and describes the three levels as different trades between token use and latency. The catalog entry declares reasoning as required rather than optional, so the composer shows the thinking control as engaged, offers no level selector, and provides no way to switch reasoning off for a cheap task. If working at a lower effort level matters to your cost or latency, that choice is not available in this interface — it would have to be made through the API. Read the official levels as the model’s behaviour and the fixed switch here as this site’s setting.

Images and PDFs, yes; audio and video, no. Google documents all five — text, images, video, audio and PDF — as inputs. The catalog entry behind this page publishes image, document, audio and video as supported input modalities, so the attach button offers JPEG, PNG and WebP images alongside PDF files, each up to 24 MB. A PDF is converted to text before it is sent, which means a scan the extractor cannot read returns an error rather than an empty answer. Audio and video cannot be uploaded through this interface at all, so a task that depends on listening to a recording or watching a clip would need a different route. Google also documents a max output of text only, so nothing here produces a generated image or audio file.

Because the catalog entry behind this page declares web search as unsupported, so the composer disables that control. The official model differs: Google lists search grounding and grounding with Google Maps among its capabilities, and the pricing page even publishes a monthly allowance of 5,000 free search requests shared across Gemini 3 and newer models. This is a gap between the official model and this interface rather than a property of the model, and it is stated here instead of left to be discovered. The practical consequence is that anything requiring current information has to arrive in the conversation as material you supply — pasted text, an image, or a PDF.

It reports two things at once and they should be read together. Against Gemini 3.1 Pro, the card’s automated safety evaluations show improvement on multilingual safety at 3.16 percentage points better, image-to-text safety at 9.29 points better, refusal tone 5.50% better and unjustified refusals 9.60 points better, while text-to-text safety is 1.57 points worse — that row is marked lower-is-better, so it is a regression, and the card’s own summary calls it a similarly high safe response rate within confidence intervals. The card also notes that these results use improved evaluations and are not directly comparable with earlier Gemini model cards. Separately, under Google DeepMind’s Frontier Safety Framework of April 2026, the card states that the chemical, radiological and nuclear threshold was not reached, that the biological and cybersecurity critical capability levels cannot be ruled out, and that the stealth and situational awareness threshold cannot be ruled out after the model passed 14 of 15 challenges there. In each of those three cases the same section says a safety case or residual-risk assessment was produced and that Google assesses the risk as lowered below the threshold.

Google dates this release twice, and the two dates are not the same day, so both are given. The release blog announcing the model is dated 30 September 2026, and there it is described as rolling out to a set of trusted cyber defenders through the Fairwind Program, with access widening to developers, enterprises and consumers as soon as possible. The model card itself is published 13 October 2026, and the Gemini API model page carries a last-updated stamp of the same day. The models index files the model under a new Gemini 4 family marked stable, with no preview row on its version line. This page does not merge those dates into a single release day, because doing so would mean choosing between two official statements.

On the official side the answer is specific and negative on the free tier. Google’s pricing page states that this model is not available on the free tier or on the Google AI Pro plan, and every one of the four rate tables carries not-available in its free column while the paid figure doubles on 1 December 2026. The release blog describes the rollout as starting with paid API customers and Google AI Ultra subscribers. On this site the answer is separate: chatting through this page needs an account with available points, and this page makes no claim about eligibility for any Google plan or about free access to the API. Read the specification card above as Google’s availability policy and the platform note as this site’s.

The same model card scores them against each other in every row except one, and the result is not a clean sweep in either direction. This model leads Claude Opus 5.5 on fourteen of the nineteen rows, including Vals Index at 68.9% against 67.0%, AutomationBench at 51.3% against 42.5%, Vals Finance Agent v2 at 65.4% against 58.6%, Harvey’s Legal Agent Benchmark at 19.6% against 3.8%, DeepSWE v1.1 at 77.9% against 74.2% and LVBench at 91.7% against 83.7%. Claude Opus 5.5 leads on four: Terminal-bench 4.0 at 66.4% against 57.4%, FrontierSWE v2 at 62.3% against 55.0%, Terminal-Bench Science 0.1 at 63.3% against 57.6% and PostTrainBench at 49.3% against 45.3%. One row cannot be compared at all, because the card prints no score for Claude Opus 5.5 on OSWorld-2.0, where this model reads 69.2%. That is four rows for the rival and fourteen for this model, all taken from the vendor’s own table rather than from a third party that re-ran both.

No, and the page keeps them labelled. Every one of the nineteen rows comes from Google’s own model card, produced under Google’s own harness, so a reader comparing this model with a rival is reading one vendor’s measurement of both rather than two parties measuring each other. Two details follow from that. The card’s safety evaluations use a different baseline again: they are stated as change against Gemini 3.1 Pro, which is not one of the four columns in the performance table, so the two tables cannot be read as one scoreboard. And the card publishes its stopping points alongside its wins, printing the rows where Claude Opus 5.5 or GPT-6 Astra leads — which is why this page reproduces all nineteen rows instead of only the leading ones.

No. This page covers Google’s Gemini 4 Argon model, a large sparse mixture-of-experts system that takes text, images, video, audio and PDF as input. Argon is also the name of a chemical element, a colourless noble gas used as a shielding gas in welding and in some kinds of lighting, and Google’s own suggestions for the model name mix the two together. Nothing on this page describes that element, its properties, its uses or its safety profile. If you came here for the model, everything above applies; if you came looking for the gas, this is a different subject.