Frontier model
Gemini 4 Argon:
The AI Race Is No Longer About Who Has the Smartest Chatbot
Google has a new frontier AI model.
But unusually, you probably cannot use it yet.
On September 30, Google announced Gemini 4 Argon, the first model in its new Gemini 4 generation. It is designed around software engineering, cybersecurity, legal and financial work, computer use, and the increasingly important category that connects all of them:
long-horizon agency.
And after spending some time looking through what Google actually published, I think focusing only on whether Gemini 4 beats GPT-6 Astra or Claude Opus 5.5 misses the bigger story.
Because something much more interesting is happening.
For most of the generative-AI era, we interacted with models roughly like this:
Ask → Think → Answer
Now the frontier is moving toward something closer to:
Goal → Investigate → Plan → Act → Test → Observe → Revise → Continue
That sounds like a small difference.
It isn’t.
It changes the fundamental unit of interaction with AI from a prompt to a job.
And Gemini 4 Argon may be one of the clearest signs yet that this transition is accelerating.
First, what exactly is Gemini 4 Argon?
Google describes Argon as a frontier model designed to maintain deep reasoning across complicated, long-running workflows.
Its strongest advertised areas are:
- real-world software engineering
- enterprise knowledge work
- finance
- legal workflows
- computer use
- multimodal reasoning
- long-context reasoning
- cybersecurity defense
Google says Argon can autonomously find, validate and patch critical software vulnerabilities, and it is initially giving access to selected cybersecurity specialists through its Fairwind program rather than releasing the model broadly.
That restricted launch is important.
Usually a model announcement follows something like:
Model → API → developers → experimentation
Argon begins more like:
Model → controlled deployment → safety evaluation → broader access later
As of October 1, Google has not announced a general public-release date.
So Gemini 4 currently exists in a strange position:
It is simultaneously one of Google’s most important AI announcements and a model most developers cannot actually call.
That is not incidental to the story.
It is part of the story.
What happened to Gemini 3.5 Pro?
To understand why Gemini 4 matters, we have to go backwards.
At Google I/O in May, Google announced Gemini 3.5 Flash and said Gemini 3.5 Pro was already being used internally and would arrive the following month.
Google’s wording was quite explicit:
“We’re also hard at work on 3.5 Pro. It’s already being used internally, and we look forward to rolling it out next month.”
That was May 19.
June came.
Then July.
The model still didn’t arrive broadly.
Google acknowledged that 3.5 Pro remained in testing while continuing to release Flash models instead. By July, Sundar Pichai was still describing 3.5 Pro as being tested while saying Google had already begun building the next generation.
Now we know how that story ended.
Gemini 3.5 Pro was ultimately scrapped as a public release, and Google jumped forward to Gemini 4 Argon.
That makes Gemini 4 more interesting than an ordinary annual upgrade.
Google effectively skipped a promised flagship generation.
And during the delay, OpenAI and Anthropic kept moving.
OpenAI introduced GPT-6 Astra.
Anthropic pushed forward with Claude Opus 5.5 and Sonnet 5.5.
Google, meanwhile, concentrated heavily on its increasingly capable Flash family.
Which brings us to one of the most important distinctions in Google’s current model strategy.
Gemini 3.8 Flash and Gemini 4 Argon are solving different problems
Gemini 3.8 Flash is already extraordinarily capable.
Google positions it as a model for running agentic workloads at scale.
It handles multi-file software work, tool orchestration, enterprise workflows and autonomous agents, and it is the default model behind Google’s managed Antigravity agent.
Its pricing also tells you exactly what Google wants it to be.
During the introductory period, Gemini 3.8 Flash costs:
$0.75 / million input tokens $3.75 / million output tokens
Standard pricing becomes:
$1.50 input $7.50 output.
This is Google’s workhorse.
Argon is something else.
Reported introductory pricing for Gemini 4 Argon is:
$2 / million input tokens $10 / million output tokens
with standard pricing expected to rise to approximately:
$4 input $20 output.
So the distinction appears to be:
Gemini Flash: capable intelligence everywhere, cheaply.
Gemini Argon: spend more intelligence on the hardest jobs.
That is probably a more useful way to think about modern AI families than “small model versus big model.”
They are becoming different computational products.
So how good is Gemini 4 actually?
This is where things get interesting.
Google published comparisons against GPT-6 Astra, Claude Opus 5.5 and Claude Fable 5.1 across knowledge work, coding, science, long-context reasoning, computer use, multimodality and cybersecurity.
A necessary warning:
These are Google-published evaluations.
They are useful evidence, but they should not be treated as a neutral universal ranking of the models.
Different laboratories use different prompting setups, reasoning budgets, tools and evaluation configurations.
Still, the results show something important.
Gemini 4 is clearly competitive at the frontier.
Here are some of the more revealing numbers from Google’s published table:
This is actually more interesting than if Argon simply won everything.
It doesn’t.
Gemini 4 Argon vs GPT-6 Astra vs Claude Opus 5.5
The benchmark results show that frontier models are developing personalities
Look closely at those results.
Argon performs extremely well on:
knowledge work
financial workflows
legal agent tasks
long-context reasoning
some forms of software engineering
and multimodal understanding.
But GPT-6 Astra beats it on FrontierSWE and OSWorld.
Claude Opus 5.5 leads substantially on Terminal-Bench 4.0 in Google’s own published comparison.
Astra is also ahead on Google’s Terminal-Bench Science evaluation.
And this is an important development.
The frontier may be moving away from a world where one model is obviously “the smartest.”
Models are becoming good at different shapes of work.
That is exactly what happened with programming languages.
There isn’t one universally “best” language.
The question becomes:
Best for what?
We may increasingly ask the same thing about models.
Gemini 4 vs GPT-6 Astra
OpenAI’s GPT-6 Astra is perhaps the most obvious comparison.
Astra is designed heavily around computer use, professional work, software engineering, browsing, science and autonomous action.
OpenAI reports:
- 59.3% on Agents’ Last Exam
- 72.6% on OSWorld 2.0
- 41.4% on AutomationBench
- 91.5% on BrowseComp
and very high scores on several reasoning evaluations.
Its API pricing is:
$10 / million input tokens
and
$50 / million output tokens.
That creates an interesting competitive dynamic.
On Google’s published comparisons, Argon is ahead of Astra on several knowledge-work and long-context evaluations.
Astra remains stronger on some computer-use, software-engineering and science workloads.
But Argon’s announced introductory token pricing is dramatically lower.
That’s potentially significant.
Because the AI race is beginning to have another dimension:
Not simply:
How intelligent is the model?
But:
How much intelligence can I afford to use?
For agentic systems, this matters enormously.
A chatbot might generate 1,000 tokens.
An autonomous coding agent might consume millions.
Price becomes part of capability.
Gemini 4 vs Claude Opus 5.5
Then there is Anthropic.
Claude has become particularly associated with coding and long-running agentic work.
Anthropic launched Claude Opus 5.5 on September 22, followed by Sonnet 5.5 on September 28.
Anthropic positions Opus 5.5 for difficult work requiring careful judgment, while Sonnet handles the cheaper, faster everyday layer.
Notice something?
That increasingly resembles Google’s own structure:
Gemini 3.8 Flash → high-volume workhorse
Gemini 4 Argon → frontier reasoning
versus
Claude Sonnet 5.5 → efficient default
Claude Opus 5.5 → difficult frontier work
The industry is converging on a tiered architecture.
And Claude remains extremely strong in software engineering.
In Google’s own measurements, Opus 5.5 scored:
66.4% on Terminal-Bench 4.0
compared with:
Gemini Argon — 57.4% GPT-6 Astra — 58.2%.
Anthropic also says an early Opus 5.5 tester completed a migration involving roughly 680,000 lines of code in less than a day.
That is the kind of example I think matters more than a benchmark leaderboard.
Because it tells us what the new unit of AI work is becoming.
Not:
“Write this function.”
But:
“Take this engineering project.”
And that brings us to the real change
Think about how AI development has progressed.
Stage 1 — Chat
You ask:
“Why is this function broken?”
The model responds.
Human executes everything.
Stage 2 — Copilot
You write code.
The model predicts and suggests the next part.
Human remains firmly inside the loop.
Stage 3 — Agent
You ask:
“Fix this bug.”
The model:
reads files searches the repository changes code runs tests returns the result.
Stage 4 — Long-horizon agent
You say:
“Migrate this service from C++ to Rust without breaking production.”
Now the model may need to:
understand the architecture map dependencies create a migration strategy modify hundreds of files run tests observe failures debug those failures update its plan continue working validate performance generate documentation and possibly repeat that loop hundreds of times.
That is fundamentally different.
The model isn’t merely generating text anymore.
It is maintaining state, intent and responsibility across time.
And that may be the real frontier.
The evolution of AI: Chat → Copilot → Agent → Long-horizon Agent
Google’s own use of Argon makes this much more than theory
Google says its engineers are already using Gemini 4 internally.
One particularly striking example involved analyzing fleet-wide telemetry and identifying infrastructure optimizations that saved more than 300 TiB of memory.
Google has also discussed using Argon-assisted systems for major code migrations, including C/C++ to Rust work connected to the Zircon kernel underlying Fuchsia.
This is precisely the type of environment where today’s agents normally struggle.
Large codebases are messy.
Dependencies are poorly documented.
Tests fail for unexpected reasons.
Internal conventions matter.
A change that looks locally correct can create problems twenty modules away.
Real engineering requires keeping an evolving mental model of the system.
If AI becomes reliable there, something profound changes.
The model stops being a tool that helps engineers type.
It begins helping engineers engineer.
Then Google gave Argon a one-million-token output limit
This might initially sound ridiculous.
Who needs an AI response containing one million tokens?
Almost nobody.
Get Pascal’s stories in your inbox
Join Medium for free to get updates from this writer.
But that question assumes the output is an answer.
Imagine instead that the output represents work.
A model performing an enormous migration could potentially generate:
code patches tool commands intermediate artifacts test results new implementations documentation configuration files revisions additional patches and thousands of intermediate operations.
Suddenly a huge output allowance makes more sense.
Google has increased Argon’s output limit dramatically from earlier Gemini generations, with reporting placing it at as much as one million tokens.
We spent years expanding how much information AI could consume.
Now we may be entering another phase:
expanding how much work AI can produce before stopping.
But longer does not automatically mean better
This distinction matters.
A model being capable of acting for longer is only useful if its probability of making catastrophic errors does not increase with every action.
Imagine an agent that succeeds at each step 99% of the time.
That sounds incredible.
Until you give it 500 consequential steps.
Reliability compounds.
Long-horizon agents therefore require much more than raw intelligence.
They need:
error recovery
tool permissions
memory
verification
sandboxing
observability
rollback mechanisms
authorization boundaries
and probably human escalation when uncertainty becomes too high.
This is why the safety architecture around Gemini 4 is as interesting as its benchmark results.
Gemini 4’s most important feature might be that Google won’t fully release it yet
Google says Argon is its most resilient Gemini model yet against indirect prompt injection attacks.
The company also describes systems for monitoring model actions, stopping execution when necessary and isolating high-risk environments.
And the model’s cybersecurity abilities make that especially important.
Google claims Argon can autonomously:
find vulnerabilities
validate them
and
patch them.
Those capabilities are obviously useful to defenders.
But the intellectual capability required to discover exploitable software is inherently dual-use.
The AI doesn’t magically become less capable because the person operating it has malicious intentions.
Which creates an uncomfortable problem.
As AI capabilities increase, access itself becomes a safety mechanism.
That is a very different world from ordinary software.
The AI model is becoming only one layer of the product
This may eventually be the biggest conceptual shift.
When we talk about frontier AI today, we tend to obsess over the model.
Gemini.
GPT.
Claude.
But consider what a truly autonomous agent requires.
It needs:
Model
for reasoning.
↓
Context + memory
for maintaining understanding.
↓
Tools
for interacting with software and the world.
↓
Permissions
for deciding what those tools can touch.
↓
Sandbox
for controlling failures.
↓
Verification
for checking whether actions actually worked.
↓
Human escalation
when confidence breaks down.
The model is only the intelligence layer.
Everything surrounding it determines whether that intelligence can safely become useful.
This is why the future of AI development may increasingly resemble operating-system design.
It isn’t enough to make intelligence smarter.
You have to decide what intelligence is allowed to do.
The modern AI agent stack
And Google has an enormous advantage here
Google owns environments in which agents can operate.
Think about the stack:
Search.
Chrome.
Android.
Workspace.
Cloud.
YouTube.
GitHub-scale internal repositories.
Global infrastructure.
Cybersecurity systems.
TPUs.
Enterprise software.
Gemini does not need to remain a separate chatbot floating above that ecosystem.
It can become an intelligence layer running through it.
And that creates an extraordinary feedback loop.
Google uses Gemini internally.
↓
Gemini encounters real problems.
↓
Those failures expose model weaknesses.
↓
The model improves.
↓
A better model can handle more Google workloads.
↓
Those workloads generate more experience.
↓
Repeat.
Few companies have environments of that scale.
That may turn out to matter more than who temporarily occupies the top position on an AI leaderboard.
There is also a fascinating price war developing
Consider the rough API economics at current announced rates:
Now imagine you’re building an agent that processes billions of tokens.
Suddenly the difference isn’t cosmetic.
Model efficiency becomes infrastructure economics.
This is where Google’s vertically integrated position matters.
Google designs TPUs.
Google operates enormous data centers.
Google trains Gemini.
Google owns Cloud.
Google owns products capable of distributing Gemini to billions of people.
The ultimate AI competition therefore isn’t just:
Who builds the smartest model?
It might become:
Who can manufacture useful intelligence most efficiently?
That’s a much more complicated race.
Benchmarks are becoming less useful precisely as AI becomes more useful
There is an irony here.
We have more AI benchmarks than ever.
But measuring frontier models is becoming harder.
A static question is relatively easy to score.
Long-running work is not.
Imagine evaluating an agent responsible for migrating an enormous production application.
Do you measure:
whether the code compiles?
whether tests pass?
how many bugs were introduced?
how long the migration took?
how much money it cost?
whether the architecture improved?
how often humans intervened?
whether the model recognized when it was uncertain?
whether it broke something six hours later?
Suddenly one benchmark number looks inadequate.
So I wouldn’t interpret Gemini 4’s release primarily as:
Google is ahead.
Or:
OpenAI is ahead.
Or:
Anthropic is ahead.
Those claims can become obsolete after a single model release.
Something more durable is happening.
The laboratories are all pushing toward AI systems capable of accepting increasingly large pieces of human work.
That is the important trajectory.
The frontier has moved from intelligence to responsibility
For years we asked:
Can AI write?
Then:
Can AI reason?
Then:
Can AI code?
Now we’re beginning to ask:
Can AI be trusted with a job?
That is a much harder question.
Because intelligence is only one requirement for responsibility.
A useful autonomous system also needs judgment.
Persistence.
Reliability.
Memory.
Security.
Restraint.
The ability to recognize failure.
And the ability to stop.
Gemini 4 Argon feels important because Google’s announcement touches nearly all of these problems simultaneously.
Its strongest story is not that it generates better answers.
It is that Google believes the model can increasingly participate in workflows where the answer itself is not the product.
The completed work is.
That might be the real meaning of Gemini 4
Gemini 4 arrives after a somewhat strange year for Google.
Gemini 3.5 Pro was promised.
Delayed.
Tested.
And ultimately never broadly released.
Meanwhile Google kept improving Flash.
OpenAI pushed Astra.
Anthropic pushed Claude further into serious engineering work.
For a while, it looked like Google might have lost momentum at the absolute frontier.
Argon is Google’s answer.
But what makes the release fascinating isn’t whether Google has retaken some imaginary AI crown.
It’s what all these companies are now racing toward.
OpenAI is building models that operate computers.
Anthropic is showing Claude working through massive engineering projects.
Google is using Gemini internally to optimize infrastructure and attack long-running software problems.
The direction is becoming clear.
We’re moving from:
AI that tells you how to do something
to
AI that helps you do it
and now toward:
AI that can potentially be trusted to keep doing it after you walk away.
That final transition is enormous.
And it raises an entirely different engineering problem.
The next great AI model doesn’t merely need to be intelligent enough to begin a task.
It needs to be reliable enough to finish one.
Gemini 4 Argon is one more step toward that world.
And perhaps the most revealing thing about this release is that Google has built a model capable enough that, for now, it isn’t comfortable simply handing all of it to everyone.
The race for intelligence is continuing.
But another race has begun alongside it:
the race to turn intelligence into reliable, affordable, controllable action.
That may ultimately matter much more.
