Gemini 4 Argon Is Out. Test It on Your Own Work First

Gemini 4 Argon is generally available, and Google is pitching it at enterprise knowledge work such as legal and finance. A launch is information, not evidence. Google's own model card shows Gemini 4 Argon behind at least one rival on 5 of the 19 benchmark rows it publishes. It leads or ties on the other 14, including all four knowledge-work rows, yet none of those rows is your work. The test that decides whether to switch is one you run yourself, on a task you would actually move.

What changed on 13 October?

Google's Gemini API changelog carries an entry dated October 13, 2026: "Gemini 4 Argon (gemini-4-argon) generally available (GA)". Google's launch post is timestamped 16:00 UTC that day. That followed a 30 September announcement in which Google said the model was "rolling out to a set of trusted cyber defenders" first.

Access now depends on the surface. Google's 13 October announcement lists the paid tier of the Gemini API for developers, Gemini for business, the API and Google Antigravity for enterprises, and the Gemini app and AI Mode in Search for Google AI Ultra and Pro subscribers, with subscriber access in AI Studio and Google Antigravity reserved for Ultra. The developer guide and the pricing page add that the model is not available on the API's unpaid tier or on the Google AI Pro plan. Check the surface your organisation actually uses.

The model card names the target work directly, describing complex workflows across software engineering, "enterprise knowledge work like legal and finance", and cybersecurity defence.

The model page lists a context window of 1,048,576 tokens shared by input and output, and a maximum output of 65,536 tokens by default, or up to 1,048,576 with long decode. The developer guide is plain: "A single response can now go past 64K tokens, up to 1M tokens minus your input." The card's knowledge cutoff is March 2026.

The developer guide lists introductory prices of US$2 per million input tokens and US$10 per million output tokens, thinking tokens included, through 30 November 2026, rising to US$4 and US$20 from 1 December 2026. Those are Google's standard paid-tier prices for the Gemini API; batch, flex and priority rates differ.

The catch in Google's own model card

The results table fills page 6 of the model card PDF, straight after the line "Gemini 4 Argon results as of October, 2026 are below:" at the foot of page 5. It compares Gemini 4 Argon with GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5. GraphWalks takes two of its rows, so the table holds 19 rows rather than 19 benchmarks. Gemini 4 Argon trails at least one rival on these five:

• FrontierSWE v2: 55.0%, against 65.5% for GPT-6 Astra
• Terminal-bench 4.0: 57.4%, against 66.4% for Claude Opus 5.5
• PostTrainBench: 45.3%, against 49.3% for Claude Opus 5.5
• Terminal-Bench Science 0.1: 57.6%, against 68.1% for GPT-6 Astra
• OSWorld-2.0: 69.2%, against 72.6% for GPT-6 Astra

It ties GPT-6 Astra on CWE-bench v1 at 68.0% and is top alone on the other 13 rows, including all four knowledge-work rows: Vals Index 68.9%, AutomationBench 51.3%, Vals Finance Agent v2 65.4% and Harvey's Legal Agent Benchmark 19.6%. A lead is not a pass mark. The top score on the legal agent row is 19.6%, and it is Gemini 4 Argon's.

Two lines in Google's evaluation methodology note matter here. It states that "All of the results are run with the Gemini API with the highest thinking settings unless indicated otherwise below", while the developer guide lists medium as the default thinking level. It also states that "All the results for non-Gemini models are sourced from providers' self reported numbers unless otherwise mentioned below", and its notes then take many rows from public leaderboards. The table is a vendor's comparison, and its published results default to a thinking setting your deployment may not use.

The card also lists limits: "Gemini 4 Argon may exhibit some of the general limitations of foundation models, such as hallucinations. There may also be occasional slowness or timeout issues." Model cards "may be updated from time to time", so read the table again on decision day.

How do you test it on your own work?

The method is the one this site set out when Google missed its own release date: evaluation on your own work, not a launch date, moves the decision. It is the argument for testing an AI tool before you buy it, applied to a model switch.

1. Pick one real task. Build a small set of examples from the one or two tasks you would actually hand to a new model, replace every name and identifier with placeholders such as [CLAIMANT_NAME] and [CLAIM_NUMBER], and freeze the set so every model faces the same inputs. Run it in an approved work account.
2. Run both models. Give your current model and Gemini 4 Argon the identical task, with the same prompt and the same inputs. If you use the API, Google's guide sets out the method for its thinking levels: "Run your evaluations at all three levels, then pick the lowest level that meets your quality bar."
3. Score, then decide. Score both outputs against criteria agreed before the run, then decide to adopt, pilot or hold, and record why, with the trigger that would change the call.

A scoring sheet keeps it honest:

Your own work decides, not the launch.

If a developer is switching an existing integration

The developer guide lists changes that, if missed, make a like-for-like test unfair:

• For existing apps with a limit lower than 32K: "Raise max_output_tokens to at least 32,768. Gemini 4 Argon can spend more of the limit on thinking than earlier models."
• "Remove temperature, top_p, and top_k from your requests. If provided, they are ignored."
• "Explicit caching (cachedContents / caches.create) is not supported on Gemini 4 Argon."
• Hard prompts can take several minutes, so the guide says to use background execution "or set HTTP client timeouts to at least 15 minutes".

Do this Monday

Choose the first task you would move and name the person who signs off its output. With that owner, agree the scoring criteria in writing, freeze a placeholder-only example set and record a baseline on your current model. Once Gemini 4 Argon is approved for your account, run the same set unchanged. The decision memo is the evidence.

TheAICommand works to the Verified Draft Method: de-identify the inputs, ground the model in your own source material, keep a person at the decision point, verify against the primary source, and log what happened.

The bottom line

Gemini 4 Argon is a real release, with a shared 1,048,576 token window, long decode and introductory pricing. Google's own table shows it top alone on 13 of 19 rows and behind on five, and its published results use the highest thinking settings unless indicated otherwise. Treat the launch as information and your own three-step test as the evidence. This is general information, not procurement or legal advice.