Google has unveiled Gemini 4 Argon, a model for complex multi-step work: software development, professional work with documents and software security. The most unusual technical detail is a limit of up to a million output tokens. The main catch is that, for now, access has gone to selected partners rather than to every Gemini user.
Below is what the primary sources actually say, where the numbers need caveats and what a developer can take from all this. Terms and availability are as of 1 October 2026; the figures come from Google’s announcement and its methodology notes. We have not run any tests of our own.
What Gemini 4 Argon is
In its 30 September announcement Google describes Argon as a new frontier model that can work on a hard task for a long time. That means not just answering a question about code but going through a sequence of steps: studying the project, testing a hypothesis, making changes and judging the result.
The main areas are software engineering, financial research, legal work and cyber defence. Google also reports strong results in analysing charts and long videos. It is a general-purpose model with an emphasis on professional workflows, not just a specialised security assistant.
That emphasis matches where the whole market is heading: vendors are no longer selling only the quality of an answer but the ability to carry out long chains of actions on their own. You can see a similar shift in the OpenAI DevDay 2026 announcements, although the companies’ products, access models and limits differ.
A million output tokens: not the same as context size
Google has raised the stated output token limit to 1 million, up from the previous 64 thousand. The announcement ties this to the model’s ability to reason and generate at length within a single run of work.
It is easy to mix up two specs here. The input context determines how much material you can hand the model. The output limit is how many tokens it can produce. The million-output-token news should not be paraphrased as “now you can load a million tokens of documents”: that is a different parameter.
For a complex migration or research project a large limit is potentially useful, because the model hits the length ceiling less often. But on its own it does not guarantee a correct result, and it does not make the budget go away. Being able to generate for a very long time does not mean it has to on every request.
More room to reason also does not replace organising the inputs. If you give an agent outdated documents and contradictory instructions, the extra limit will not fix the underlying task. More on that in our piece on context engineering.
What the benchmarks show, and where they diverge
In the official announcement Google highlights several results:
- DeepSWE v1.1, 77.9%: long-running software engineering tasks.
- AutomationBench, 51.3%: end-to-end business processes; the company claims first place.
- LVBench, 91.7%: understanding long videos.
- CWE-bench v1, 68%: fixing vulnerabilities; according to Google, the model shares first place.
Google’s summary table also includes Vals Index, 68.9%. This index combines results on professional tasks in finance, programming, law and tax accounting, weighting the sectors by their share of US GDP. It is not a universal measure of “intelligence”, nor the share of any job the model could take from a human.
Vals AI itself confirms first place here. In the public Vals Index table, version 2.1 from 30 September, Argon leads with 68.90%, followed by Claude Sonnet 5.5 (67.04%) and Claude Opus 5.5 (66.97%). The gap is about two percentage points: noticeable, but not a chasm.
The terminal tasks deserve a closer look. Google’s table lists Terminal-Bench 4.0, 57.4% and Terminal-Bench Science 0.1, 57.6%. These are different tests with similar numbers, and they are easy to confuse. The 57.6% figure should not simply be labelled Terminal-Bench and compared with an earlier result without checking the version and conditions.
The DeepMind methodology explains that the results were obtained with maximum reasoning settings and that conditions vary between individual tests. Google computed Argon’s Terminal-Bench 4.0 score itself, while the other models’ numbers were taken from the public leaderboard. For Terminal-Bench Science the grading system’s timeout was increased sixfold. The figures are useful as a guide, but don’t assume they carry over directly to your own repository.
How Google is already using Argon internally
More interesting than general promises are the concrete engineering examples in the announcement. They come from Google itself; we have no independent verification of these cases.
Memory optimisation. A team of agents analysed infrastructure profiling data and found changes that cut memory use. According to the company, the optimisations it rolled out freed up more than 300 TiB. That is the result of changes to the infrastructure, not the memory size of the model itself.
Migrations from C/C++ to Rust. Google describes agents working on projects of different scales, from libraries to the Zircon kernel of the Fuchsia operating system. At the same time it explicitly mentions manual and automated auditing, testing in emulation and review before rollout. The story is not “the model rewrote a critical system and it went straight into production”.
Optimising the libgav1 video decoder. In an existing Rust port, the agents replaced 32 thousand lines of SIMD code, running experiments with profiling and studying the compiler output. Google claims a 2.7x speed-up over that Rust port with identical decoding output. The comparison applies to that specific starting version, not to any decoder or any C++ code.
What these examples have in common is that the model’s work is measured and checked. There is a baseline version, a metric, experiments and verification. That approach is more useful than the slogan “developers are no longer needed”.
API pricing: the introductory rate won’t last
Google has announced the following introductory prices per million tokens:
- input: $2;
- output: $10;
- cached input: a 95% discount on the standard input price.
A footnote carries an important caveat: after the introductory period the price will be $4 per million input and $20 per million output tokens. The announcement does not say when that period ends.
These are the announced terms for future access, not proof that any developer can already call the model at that rate. Before integrating, check availability and current prices in your own account.
For a project, what matters more than the price per token is the cost of an accepted result, counting retries, tools and the human time spent checking it.
Who can use the model right now
Initial access is going to trusted cyber defence specialists through the Fairwind Program. Google is gathering feedback and strengthening safeguards ahead of a wider launch.
The company plans to expand access to developers, organisations and everyday users, starting with paying API customers and Google AI Ultra subscribers. The announcement gives no exact date for general availability. So “Gemini 4 is out” in practice means the model has been presented and is in limited rollout, not that every app user has Argon.
The Fairwind Program is a managed-access programme for vetted organisations: government agencies, critical infrastructure operators, large technology platforms and academic labs. The programme page already lists Argon among its tools, alongside the code security agent CodeMender.
For Google itself, cyber defence is one of the first use cases: the model is meant to find, confirm and fix vulnerabilities. But as such capabilities grow, they call for isolated environments, access restrictions and protection against malicious instructions hidden in external data. A claimed resistance to prompt injection does not mean the problem has gone away.
What a developer should do about it
Until there is broad access, it is more useful to prepare a trial than to change your working stack based on a screenshot of a leaderboard. Pick limited tasks you already do with another model: fixing a reproducible bug, migrating a small module, research with verifiable references.
For each task, set the criteria for “done” in advance. For code, that means tests and an acceptable size of change; for research, the quality of sources and no invented conclusions. Compare not the length of the answer but the share of accepted results, the full cost and the amount of manual fixing. We covered the practical basis for this kind of process in our article on working with an AI agent like a grown-up.
Gemini 4 Argon is interesting for its combination of long-running work, a large output limit and concrete engineering scenarios. But the announcement does not yet let you honestly conclude that it beats Claude across the board, that it guarantees savings, or that it is ready to replace human review. The next useful step is to wait for access and test the model on your own tasks, keeping your tests, your budget and control over its actions.
