What changed
On 30 September 2026, Google announced Gemini 4 Argon. Google calls it the first model in the Gemini 4 family. Its main pitch is cyber defence: finding, checking and fixing weak spots in software before attackers use them.
Three things stand out. The model is not open to the general public yet. Google says it can produce up to 1 million output tokens, up from 64K on earlier models. And Google says it is being trained and tested for security work, with its own cyber benchmark results to back that up.
- Name and family: Gemini 4 Argon, the first Gemini 4 model (Google's announcement; Engadget).
- Output limit: 1 million tokens, up from 64K (Google's announcement; 9to5Google).
- Access today: a trusted-defender programme called the Fairwind Program (Google's announcement; 9to5Google; Engadget).
- Price: $2 per million input tokens and $10 per million output tokens as an introductory rate (9to5Google; Artificial Analysis).
Confirmed facts, vendor claims and open questions
Launch coverage mixes three kinds of statement. Some are facts that several outlets and a third-party tester agree on. Some are Google's own claims, which we can report but not yet check. And some are details that appear in only one place. The table below sorts the main points.
| Point | Status | Where it comes from |
|---|---|---|
| Announced 30 September 2026 | Confirmed | 9to5Google, Artificial Analysis |
| Output limit of 1M tokens (from 64K) | Confirmed as Google's figure | Google's announcement, 9to5Google |
| Ties for first on CWE-bench v1 at 68% | Google's figure | Google's announcement, 9to5Google; Engadget confirms the tie |
| Can find, check and patch serious flaws on its own | Google's claim | Google's announcement only |
| Intelligence Index of 53, one point ahead of GPT-6.1 Sol | Single source | According to Artificial Analysis (launch snapshot); Engadget on the ranking |
| 15% hallucination rate | Single source | According to Artificial Analysis; Engadget |
| Successor to Gemini 3.5 | Single source | According to Engadget |
| Cached input tokens at a 95% discount | Single source | According to 9to5Google |
Benchmark numbers in launch posts are chosen by the company that made the model. Treat Google's cyber benchmark results, such as CWE-bench, as Google's own figures until independent labs publish their runs. The Artificial Analysis figures are third-party, but they are still a snapshot taken at launch.
Why a cyber-defence model, and why first to defenders
Cyber defence is the first use case Google has chosen to name. Google describes Argon as able to autonomously find, validate and patch critical software vulnerabilities. According to Google's announcement, an early test found a serious flaw in healthcare software used by hospitals, one that exposed sensitive patient information and that earlier models had missed. Google's announcement credits the discovery to Wiz. Only Google's own account is available for this example, so treat it as a claim.
The release pattern is also worth noting. Argon is going first to vetted defenders, not to open API users. This resembles the vetted-access approach covered in our pieces on OpenAI's GPT-5.6 cyber programme and Anthropic's Mythos preview. Our reading is that the reason is simple: a model that is very good at finding flaws is also very good at finding flaws to exploit. Vetted access lets the vendor see who is using it and for what. For buyers, it means the strongest security tooling may arrive through approved programmes first, with paperwork attached.
The 1M output limit: what it does and does not mean
The output limit is the amount of text the model can write in a single response. Going from 64K to 1 million tokens is a large jump, and it matters for jobs such as rewriting a large codebase, producing a full test suite or writing a long audit report in one go.
It does not mean every answer will be long. It is a ceiling. The price is set per token, so the cost of a long answer rises in step with its length. Output tokens are also the more expensive side of the bill at the standard rate. The table below shows the arithmetic for a sample job. These are our own calculations from the published rates, not benchmark results.
| Sample job | Input | Output | Cost at introductory rate | Cost at standard rate |
|---|---|---|---|---|
| Scan of one service, short report | 100k tokens | 10k tokens | $0.30 | $0.60 |
| Repo-wide patch, long output | 800k tokens | 200k tokens | $3.60 | $7.20 |
| Filling the full 1M output limit once | Not counted | 1M tokens | $10.00 | $20.00 |
Before you plan a pipeline around the 1M limit, set a hard cap on output tokens in your own code. Ask for structured output such as a findings list, a patch file and a short summary, rather than a full narrative. That keeps the bill predictable and makes results easier to review.
Pricing: introductory rate versus standard rate
Google's introductory price is $2 per million input tokens and $10 per million output tokens. Reported standard pricing after the introductory period is $4 per million input tokens and $20 per million output tokens. Artificial Analysis reports that Google has not confirmed when the introductory rate ends, and that a launch discount is in effect.
For comparison, according to Artificial Analysis, Argon's cost per task is about 60% of GPT-6 Astra's on its Intelligence Index, at current discounted pricing. At standard pricing, Artificial Analysis puts Argon at $3.98 per task, above GPT-6 Astra's $3.26. These are single-source figures from one tester's launch snapshot. Budget for the standard rate, not the launch rate, when you plan a yearly contract.
Do not sign a multi-month commitment priced at the introductory rate. The discount is temporary. According to Artificial Analysis, Google has not said when it ends. If procurement needs a fixed number, use the standard rate as the base case.
What it means for UK and Indian security teams
The UK and India have different buying habits, but the questions are the same. Here is how Argon fits into the three areas that matter most to hiring managers, tool buyers and procurement leads.
- Hiring. Teams that can use a model like this well will need people who understand vulnerability triage, can write clear evaluation harnesses and can check model output before it touches production. Those are skills that are hard to find on a CV alone. A verified profile with real work shown is a better signal than a list of certificates.
- Tooling. Do not wire a cyber model straight into a CI pipeline on day one. Start with read-only scans, log every finding, and require a human sign-off before any patch is merged. Keep the model's access to repositories and secrets as narrow as possible.
- Procurement. Ask vendors whether a model is available through a programme like Fairwind, what checks apply to your organisation, and where your data is processed. In the UK, that means checking your obligations under UK GDPR and any sector rules that apply to you. In India, it means checking the Digital Personal Data Protection Act, 2023 and its rules, which are being phased in, with the main duties due from May 2027.
What it means for AI builders outside security
Most builders will not get Argon this quarter. The useful lesson is about cost and shape, not access. A 1M output ceiling changes how you think about long-running agents, batch rewrites and report generation. It also raises the value of good caching, good output limits and a clear cost per task.
If you run agents today, compare your current model against the cost-per-task numbers and not just the price per million tokens. Our earlier coverage of cost per task in Gemini's Flash line explains why that difference matters. If your work touches security, read how vetted access works for a comparable model in OpenAI's GPT-5.6 cyber programme.
A practical test before you adopt any frontier model is to run a small bake-off on your own data. Pick twenty real tasks from your backlog, including two or three that are messy and long. Measure accuracy, time taken and total cost for each model, then compare the results. Launch benchmarks rarely reflect your codebase, your documents or your customers' languages, so your own numbers are the ones that should drive the decision.
Every article here is written by a Verified Builder community. Want your name on the next one?
AI Tech Connect lists AI engineers, founders and researchers across India and the UK. The people hiring browse it to find them. Adding your profile is free.
Become a Verified Builder →What to watch next
Four things will tell you more than the launch post did. First, whether independent labs publish their own cyber benchmark runs for Argon. Second, the date on which the introductory price ends. Third, when Argon reaches paid API customers and AI Ultra subscribers. Fourth, whether the 1M output limit holds up under real workloads, which is where most launch claims get tested. Each of these answers will change the economics for UK and Indian teams, because a cheaper, widely available model reaches a small agency in Pune or a start-up in Leeds just as quickly as it reaches a large enterprise.
For background on how defensive cyber models are being released across the industry, see our coverage of Anthropic's Claude Mythos preview for cybersecurity and our policy section for the regulatory side. For practical checklists on running AI agents safely, see this approval checklist for browser agents.
Primary sources: Google's announcement at blog.google, and independent measurements from Artificial Analysis.