For the last few years, the AI race has often felt like a race between companies trying to build the smartest model.
OpenAI became one of the biggest names in that race with GPT.
Google responded with Gemini.
But Google has sometimes struggled to keep pace with OpenAI at the very top of the market.
Now, that may be changing.
On September 30, 2026, Google introduced Gemini 4 Argon, its new flagship frontier model. Google says Argon is designed for difficult, long-running work across software engineering, business knowledge, finance, legal work, and cybersecurity. It is initially being released to selected cyber defenders through Google's Fairwind program before a wider rollout.
OpenAI's GPT-6 Astra, meanwhile, has already established itself as one of the leading frontier models. OpenAI describes Astra as its most capable broadly deployed model, with strong performance in computer use, browsing, software engineering, cybersecurity, science, and professional work.
So the big question is:
Has Google finally caught OpenAI?
The early answer is: Google has clearly closed the gap, but it is too early to say that Argon has beaten Astra overall.
And that is what makes this comparison interesting.
Gemini 4 Argon vs GPT-6 Astra: What Has Changed?
The most important difference is not simply that Google released a new model.
It is how much stronger Argon appears to be compared with Google's previous generation.
Independent testing from Artificial Analysis puts Gemini 4 Argon at about 52.6 on its Intelligence Index at Google's highest tested reasoning setting, while GPT-6 Astra scores about 52.7. In other words, the two models are almost tied on that particular overall measure.
That is a major result for Google.
It suggests that Argon is no longer clearly behind OpenAI at the frontier.
But benchmark scores need context.
A model can win one test and lose another. Different models may also use different amounts of computation to reach their answers.
That is exactly what we see with Argon and Astra.
Google Claims Strong Results Across Many Tests
Google says Gemini 4 Argon beats competing frontier models on many of its internal benchmark comparisons.
The company reports strong performance in areas including coding, enterprise knowledge work, cybersecurity, automation, and long-context reasoning. Google also says Argon can work with up to one million tokens of context, allowing it to handle very large amounts of information in a single task.
A million-token context window is particularly interesting for businesses.
Imagine giving an AI:
- a large collection of contracts
- years of company documents
- a large software project
- financial reports
- customer research
- product information
- technical documentation
Instead of splitting everything into many smaller conversations, a model with a very large context window can potentially reason over much more information at once.
This does not automatically mean better answers.
But it can make complex workflows much easier.
GPT-6 Astra Is Still a Very Strong Competitor
OpenAI launched GPT-6 Astra earlier in September and positioned it as a major step forward in reasoning and agentic work.
Astra is designed not only to generate text but also to work with computers, browse the web, write software, perform research, and handle professional tasks. OpenAI says it reached a 98% score on FrontierMath Tier 4 and 99.9% on ARC-AGI-3, while also setting new results for computer and browser use.
OpenAI also says Astra is more aligned than previous models and is better at staying within its authorized scope.
That matters because frontier AI is moving toward agents.
The question is no longer only:
“Can the AI answer my question?”
It is increasingly:
“Can the AI complete the task correctly?”
A model that can reason well but cannot reliably use tools may be less useful than a slightly weaker model that can complete a real workflow.
This is one reason the Gemini vs GPT competition is becoming much harder to judge with a single benchmark.
Where Gemini 4 Argon Looks Especially Strong
One of the most interesting areas for Argon is agentic work.
Artificial Analysis reported that Argon reached 78% on AutomationBench, which tests automated workflows in business software. That result placed it ahead of the models included in that particular comparison.
This is important because businesses are increasingly interested in AI agents.
An employee may not want an AI that simply explains how to complete a task.
They want an AI that can actually do it.
For example:
“Look at our sales data, identify the biggest changes, investigate why they happened, and prepare a report.”
That requires several steps.
The AI needs to understand the goal, analyze information, perform research, connect different pieces of evidence, and create a useful result.
Argon's focus on long-horizon workflows is therefore strategically important.
Google is not only trying to make Gemini better at answering questions.
It is trying to make Gemini better at doing work.
Coding Is More Complicated
Coding is one area where the early results do not give Google a simple victory.
On Terminal-Bench 4.0, an independent Artificial Analysis test put Argon at around 57%, while GPT-6 Astra scored around 59%. Claude's latest models were also ahead on that test.
Google's own testing shows Argon performing strongly on coding-related tasks, but Reuters noted that Argon still trailed competitors on some of the coding benchmarks included in Google's comparison.
This is a good example of why the phrase “Google beats GPT-6” is too simple.
The reality is more interesting.
Argon is competitive.
It may lead on some coding tasks.
Astra may lead on others.
And another frontier model may win a different test.
For developers, the best model may depend on the type of software work being done.
One of Argon's Most Interesting Advantages: Hallucinations
Another early test gives Argon an interesting advantage.
Artificial Analysis reported a lower hallucination rate for Argon than GPT-6 Astra in its evaluation. It reported about 15% for Argon compared with 51% for Astra in that specific test. However, the same analysis found that Astra had higher raw accuracy on the task.
This distinction is important.
A model that says “I don't know” can be safer than one that confidently invents an answer.
But refusing to answer is not the same as being correct.
Businesses therefore need to look at both:
How often does the model make things up?
and
How often does it actually get the answer right?
This becomes especially important in areas such as finance, law, healthcare, research, and enterprise decision-making.
The Price Difference Is Huge
This is one area where Google is making an aggressive move.
Google says Gemini 4 Argon launches at an introductory API price of $2 per million input tokens and $10 per million output tokens.
After the introductory period, Google says the standard price will become $4 per million input tokens and $20 per million output tokens.
GPT-6 Astra is considerably more expensive through the OpenAI API.
OpenAI lists Astra at $10 per million input tokens and $50 per million output tokens.
At first glance, this makes Argon look like the obvious winner.
But there is a catch.
Token price is not the same as task cost.
A model that uses many more tokens to complete a task can become expensive even if each token is cheap.
Artificial Analysis found that Argon used substantially more output tokens per task than Astra in its comparison. It estimated an average task cost of about $1.99 for Argon during the launch discount versus about $3.26 for Astra. Once the discount ends, its estimate puts Argon at about $3.98 per task compared with $3.26 for Astra.
So businesses should not simply ask:
“Which model has the cheaper API?”
They should ask:
“How much does it cost to complete the work I actually need?”
That is a much better question.
The Million-Token Context Could Change Enterprise AI
One of the biggest technical stories around Argon is its very large context window.
Google says Argon supports up to one million tokens.
For normal chatbot conversations, this may not matter much.
For enterprise work, it could matter a lot.
Consider a law firm.
A lawyer might need to examine hundreds of documents.
A financial analyst might need to study years of reports.
A software team might need to understand a huge codebase.
A large retailer might have millions of product details and internal documents.
A model that can maintain a large amount of context can potentially reason across more information without constantly losing the earlier parts of the task.
This is one area where Google has a strong story.
But again, context size alone does not guarantee intelligence.
The model still needs to find the right information, understand it correctly, and reason over it.
Google Is Also Using Argon Internally
Argon is not only a public product announcement.
Google says the model is already being used internally by thousands of employees for coding, research, writing, and specialized engineering work.
Google gives some impressive examples.
In one quantum computing task, the company says Argon helped researchers optimize a subroutine and beat a published baseline by 40%.
Google also says a group of Argon agents analyzed data-center memory usage and identified optimizations that could free more than 300 TiB of memory once rolled out, with a potential total savings estimate of 500 TiB to 1 PiB.
These examples are interesting because they show where frontier models are heading.
The goal is not simply to produce better text.
It is to help organizations solve complicated technical problems.
Astra Has a Similar Enterprise Direction
OpenAI is taking the same general path.
GPT-6 Astra is designed for professional work, software engineering, computer use, browsing, science, and cybersecurity. It is also available through the OpenAI API and enterprise channels such as Microsoft Azure and Amazon Bedrock.
That means Astra is becoming part of a larger ecosystem.
Businesses can potentially build AI agents around it.
Developers can connect it to tools.
Companies can use it for internal workflows.
The competition is therefore not just:
Gemini vs GPT
It is:
Google's AI ecosystem vs OpenAI's AI ecosystem.
That is a much bigger battle.
Safety Is Becoming Part of the Model Race
As models become more capable, safety becomes increasingly important.
Google has chosen a phased rollout for Argon.
The model is initially being provided to trusted cybersecurity teams through its Fairwind program. Google says it wants to gather feedback and improve safeguards before making Argon more broadly available to developers, businesses, and consumers.
OpenAI has also emphasized safety around Astra.
OpenAI says Astra reaches its Critical level for cybersecurity capability under its Preparedness Framework. The company says it has strengthened protections, isolation, monitoring, jailbreak testing, and alignment evaluations as a result.
This is an important shift.
The more powerful the model becomes, the more important it becomes to control what the model can do.
Frontier AI is no longer just a competition for intelligence.
It is also a competition for reliability, security, and controlled autonomy.
So, Is Google Finally Catching OpenAI?
The answer depends on what “catching” means.
If it means:
Can Google build a frontier model that competes directly with GPT-6 Astra?
Yes.
The early evidence strongly suggests that Google has done that.
Artificial Analysis places Argon almost exactly level with Astra on its overall Intelligence Index.
If it means:
Does Argon beat Astra at everything?
No.
It clearly does not.
Some coding tests still favor Astra.
Other benchmarks favor Argon.
And independent testing shows that the overall picture is much closer than a simple winner-versus-loser story.
If it means:
Has Google returned to the very top tier of frontier AI?
That is probably the most reasonable conclusion.
Reuters described Argon as Google's new attempt to close the gap with OpenAI and Anthropic, while independent testing now places Google back among the leading frontier labs.
Which Model Should You Choose?
There is no universal winner.
For complex enterprise workflows, Argon's long context, agentic capabilities, and competitive pricing make it very interesting.
For coding and computer-use tasks, Astra remains extremely strong and should be tested directly against the specific development workflow.
For large document analysis, Argon's million-token context could be a major advantage.
For API cost, Argon currently looks attractive, especially during its introductory pricing period.
For safety-sensitive applications, businesses should evaluate the models using their own security requirements rather than trusting general benchmark results.
For everyday users, the best choice may simply be whichever ecosystem they already use.
Google users may naturally benefit from Gemini's connection to Google's products.
OpenAI users may prefer Astra because of the existing ChatGPT and developer ecosystem.
The Bigger Story Is Not Google vs OpenAI
The most important lesson from Gemini 4 Argon is that frontier AI is becoming harder to separate into clear winners.
One company may lead in coding.
Another may lead in reasoning.
Another may lead in agentic workflows.
Another may offer lower prices.
And a model that looks slightly weaker on a benchmark may be much better for a particular business.
This means the AI race is becoming less about one giant leaderboard.
It is becoming a race across many dimensions:
intelligence, cost, context, speed, coding, agents, multimodality, reliability, safety, and real-world usefulness.
That is good news for businesses and developers.
Competition forces the companies to improve.
Final Thoughts
Gemini 4 Argon is one of the clearest signs yet that Google is back in the serious frontier-model conversation.
It is not simply another Gemini upgrade.
Google has built Argon for long-running professional work, coding, enterprise knowledge tasks, cybersecurity, and AI agents. The company is also pricing it aggressively and giving it a very large context window.
GPT-6 Astra remains a powerful competitor, with strong results in reasoning, computer use, coding, cybersecurity, science, and professional work.
Early independent testing suggests the two models are remarkably close overall.
So, is Google finally catching OpenAI?
Yes — and arguably the more important point is that the gap is no longer the story.
The real story is that the frontier has become incredibly competitive.
Google no longer needs to prove that Gemini belongs in the conversation.
Now it needs to prove that Argon can consistently turn impressive benchmark results into better real-world work.
And OpenAI has a new problem too.
It can no longer assume that being at the frontier is enough.
Google is right there.
Anthropic is right there.
New models are arriving faster.
Prices are changing.
AI agents are becoming more capable.
The next stage of the AI race will therefore not be decided by who can release the biggest model.
It will be decided by which model can deliver the most useful intelligence, at the right cost, with the fewest mistakes, while safely completing real work.
That is a much harder race.
And Gemini 4 Argon has made it much more interesting.