Gemini 4 Argon Beats GPT-6 Astra on Key Benchmarks at Lower Cost

Google has unveiled Gemini 4 Argon, its latest frontier artificial intelligence model aimed at complex software engineering, professional knowledge work and cybersecurity.

Early benchmark results put Argon ahead of OpenAI’s GPT-6 Astra in several high-profile evaluations, including long-horizon software engineering, business automation and the independent Vals Index.

Its pricing could prove equally significant.

Google says Gemini 4 Argon will initially cost $2 per million input tokens and $10 per million output tokens when paid access expands. After the introductory period, those prices will increase to $4 and $20 respectively.

By comparison, OpenAI lists GPT-6 Astra at $10 per million input tokens and $50 per million output tokens under standard API pricing.

That means Argon’s introductory token rates are only 20% of Astra’s standard rates, while its eventual $4/$20 pricing will equal 40% of Astra’s standard API price.

What Is Gemini 4 Argon?

Gemini 4 Argon is Google’s new frontier model designed specifically for tasks requiring sustained reasoning over long, complicated workflows.

Google says the model is particularly strong in:

• Software engineering
• Financial research
• Legal work
• Business automation
• Multimodal analysis
• Cybersecurity
• Large codebase migrations
• Long-running agentic tasks

Google announced the model on September 30, 2026.

However, Argon is not immediately available to everyone.

Google is first providing the model to selected cybersecurity defenders through its Fairwind Program while conducting further safety testing.

Broader access is planned later for paid API customers, enterprises, developers and Google AI Ultra subscribers.

Gemini 4 Argon vs GPT-6 Astra Pricing

Price is one of the most striking differences between the two models.

Google says the $2/$10 price is introductory. Once that promotional period ends, Argon will move to $4 per million input tokens and $20 per million output tokens.

OpenAI lists GPT-6 Astra Standard API pricing at $10 for input and $50 for output.

This creates two useful comparisons.

At launch, Gemini 4 Argon costs approximately 20% as much as Astra per token.

After the introductory period, it will cost approximately 40% as much as Astra at the listed standard rates.

Therefore, headlines stating that Argon costs “40% of Astra” refer to Google’s eventual standard price rather than its cheaper introductory offer.

Does Gemini 4 Argon Actually Beat GPT-6 Astra?

On some important benchmarks, yes.

But saying that Argon simply “beats GPT-6 Astra” in every respect would be inaccurate.

Different benchmarks measure very different capabilities, and no current frontier model leads every category.

For example, Google reports that Gemini 4 Argon achieved 77.9% on DeepSWE v1.1, a benchmark focused on complex real-world software engineering.

OpenAI reports GPT-6 Astra at 74.1% on the same benchmark.

That gives Argon a 3.8-percentage-point advantage.

Argon Leads GPT-6 Astra on DeepSWE

DeepSWE v1.1 evaluates models on long-horizon software engineering work rather than simple coding questions.

The reported scores are:

Google describes Argon’s 77.9% result as a new state of the art on the benchmark.

This matters because AI coding is increasingly moving beyond writing individual functions toward tasks such as debugging entire projects, migrating codebases and managing multi-step software-development workflows.

Argon Also Leads on AutomationBench

Gemini 4 Argon performs strongly on AutomationBench, which evaluates whether an AI system can complete end-to-end workflows across business functions.

Google reports a score of:

51.3% for Gemini 4 Argon

OpenAI’s published benchmark table lists:

41.4% for GPT-6 Astra.

That represents a sizable advantage for Argon on this particular evaluation.

AutomationBench is especially relevant to businesses exploring AI agents that do more than answer questions and instead complete multi-step work involving applications, documents and tools.

Independent Vals AI Ranking Puts Argon First

Perhaps more significant than company-published benchmarks is the current Vals Index, because it comes from an independent benchmarking organisation.

Vals AI’s leaderboard, updated September 30, ranks Gemini 4 Argon first with an overall score of 68.90%.

GPT-6 Astra scores 63.13% and ranks below Argon on the index.

The Vals Index combines performance across real-world professional tasks involving:

• Finance
• Coding
• Legal work
• Tax

The sectors are weighted partly according to their contribution to the U.S. economy.

According to the current leaderboard:

This provides independent evidence that Argon can outperform Astra on a basket of professional knowledge-work tasks.

But Vals Test Costs Tell a More Complicated Story

Token prices are not the only factor that determines how expensive an AI model is to run.

Models can consume different numbers of tokens when solving the same task.

Vals reports an average cost per test of $15.68 for Gemini 4 Argon, compared with $18.46 for GPT-6 Astra in its current evaluation.

That difference is much smaller than simply comparing their advertised per-token prices.

Why?

Because total AI inference cost depends on:

• Number of input tokens
• Number of output tokens
• Amount of reasoning performed
• Caching
• Number of attempts or tool calls
• Length of each task

A cheaper price per token therefore does not automatically mean every completed job will cost 60% or 80% less.

For developers, cost per successful task may ultimately be more useful than token price alone.

GPT-6 Astra Still Leads in Other Areas

Gemini 4 Argon’s strong results should not be interpreted as proving it is universally more capable than GPT-6 Astra.

OpenAI reports particularly strong Astra results in several areas.

GPT-6 Astra scores:

• 97.6% on FrontierMath Tier 4
• 99.9% on ARC-AGI-3
• 100% on ExploitBench
• 72.6% on OSWorld 2.0
• 57.9% on Terminal-Bench 4.0
• 96.0% on GPQA Diamond

Google has not published directly comparable Argon results for every one of those evaluations.

It would therefore be incorrect to use a handful of winning benchmarks to conclude that Argon is superior to Astra at every task.

The more accurate conclusion is that Argon has established frontier-level performance and leads Astra on several notable professional and coding evaluations while offering substantially cheaper listed token pricing.

Argon Can Generate Up to 1 Million Output Tokens

Another unusual feature is Argon’s extremely large output capacity.

Google says Gemini 4 Argon supports an output-token limit of up to 1 million tokens, up dramatically from the 64,000-token output ceiling cited for previous Gemini models.

That is different from a context window.

A context window determines how much information a model can process within a request.

The output-token limit determines how much content the model can potentially generate in one trajectory.

OpenAI lists GPT-6 Astra with:

• 1.05 million-token context window
• 128,000 maximum output tokens

Argon’s extremely high output allowance could be useful for long-running agents, large coding projects and complicated research workflows.

In normal chatbot usage, however, most people would never need anywhere near one million generated tokens.

Why Google Built Such a Large Output Limit

Google says Argon is designed for problems where a model may need to reason and work for an extended period rather than produce a short response.

Examples could include:

• Migrating a large software project
• Investigating a complicated financial question
• Analysing hundreds of documents
• Running a lengthy debugging workflow
• Completing a multi-stage research task
• Finding and patching vulnerabilities

Google argues that giving the model considerably more room to work allows it to maintain deeper reasoning trajectories rather than stopping early because of output restrictions.

This reflects a broader shift in AI development away from simple chatbot interactions toward autonomous and semi-autonomous agents.

Google Is Already Using Argon Internally

Google says thousands of its employees are already using Gemini 4 Argon in internal workflows.

One particularly ambitious application involves migrating large C and C++ codebases to Rust.

According to Google, Argon agents have worked on codebases ranging from tens of thousands of lines to more than 800,000 lines in the Fuchsia operating system’s Zircon kernel.

Google also says Argon helped rework approximately 32,000 lines of SIMD code in the Rust version of its libgav1 video decoder.

The resulting implementation reportedly ran 2.7 times faster than the previous Rust port while maintaining identical video output.

These are Google-reported internal examples rather than independent benchmark results, so they should be understood in that context.

Argon Is Also Being Used in Google Data Centers

Google says a team of Argon agents analysed profiling data across its computing fleet to find memory-efficiency improvements.

The company claims the resulting optimisations have already freed more than 300 TiB of memory, with estimated potential savings between 500 TiB and 1 PiB once fully deployed.

Again, these figures come directly from Google.

They nevertheless illustrate what companies increasingly want from frontier AI models: not simply text generation, but autonomous analysis of highly complex technical systems.

Cybersecurity Is a Major Focus

Cybersecurity is one of the most significant parts of the Gemini 4 Argon launch.

Google says Argon can autonomously:

• Find software vulnerabilities
• Validate whether vulnerabilities are genuine
• Develop fixes
• Patch affected code

On CWE-bench v1, which tests vulnerability remediation, Google reports that Argon scored 68% and tied for first place.

Cloud-security company Wiz is already testing the model through its Scan for Good initiative.

Google says Argon identified a serious vulnerability affecting healthcare software that previous frontier systems had missed.

Why Google Is Restricting Argon Initially

The same cybersecurity capability that makes Argon useful to defenders could potentially make it useful to attackers.

Google is therefore taking a phased approach to deployment.

Initial access is being provided through the Fairwind Program to trusted cyber defenders.

Google says it is also participating in a U.S. government voluntary pre-release process while strengthening safeguards before expanding availability.

The company specifically highlights safeguards against:

• Cyber misuse
• Chemical and biological misuse
• Prompt-injection attacks
• Agent misalignment
• Unauthorized actions

Google says broader availability will follow once those protections are strengthened.

Prompt-Injection Resistance Gets Special Attention

Prompt injection has become an increasingly important problem as AI agents gain access to browsers, emails, documents and external applications.

A malicious instruction hidden inside a webpage or document could potentially cause an agent to behave in ways its user did not intend.

Google says Argon is its strongest model yet against indirect prompt-injection attacks and reports leading performance on Gray Swan’s Indirect Prompt Injection benchmark.

This could become particularly important for enterprise AI systems that operate across sensitive internal data.

Google Is Monitoring Argon for Misalignment

Google also describes a monitoring system designed to identify situations where an AI agent attempts to accomplish a task by going beyond its authorised scope.

The company says it monitors Argon’s reasoning and actions and can stop execution when necessary.

This reflects one of the emerging challenges in advanced AI development.

As models become more capable of independently taking actions, developers increasingly need safeguards not only against harmful user requests but also against unintended model behaviour during legitimate tasks.

How Does GPT-6 Astra Compare?

OpenAI positions GPT-6 Astra as its highest-capability model for demanding professional work.

It supports:

• Complex reasoning
• Coding
• Computer use
• Web research
• Document creation
• Scientific work
• Cybersecurity
• Professional workflows

OpenAI lists a 1.05 million-token context window and up to 128,000 output tokens.

Astra is also widely available through ChatGPT, the OpenAI API, Microsoft Azure and Amazon Bedrock, unlike Gemini 4 Argon’s initial restricted rollout.

That availability difference matters.

A model may look attractive on benchmarks, but businesses cannot deploy it widely until the provider actually makes commercial access available.

Gemini 4 Argon vs GPT-6 Astra at a Glance

Prices shown are the providers’ listed standard or introductory rates and may not reflect every processing tier, caching discount or enterprise agreement.

Is Gemini 4 Argon Really 60% Cheaper Than Astra?

At its eventual standard pricing, yes — on a per-token list-price basis.

Argon’s planned $4 input / $20 output pricing is 40% of Astra’s $10 / $50 pricing.

In other words, it is 60% cheaper per token.

At Argon’s introductory $2 / $10 pricing, the difference is even larger: Argon costs only 20% as much per token, or 80% less than Astra’s standard rates.

But this does not guarantee that every workload will cost exactly 60% or 80% less.

Models consume different quantities of tokens and can require different numbers of reasoning steps.

Vals, for example, currently reports costs of $15.68 per test for Argon and $18.46 for Astra, despite their much larger difference in list pricing.

Businesses should therefore compare completed-task cost rather than relying only on token prices.

Why Price Matters So Much for AI Agents

Traditional chatbot usage may involve only a few thousand tokens at a time.

AI agents can consume vastly more.

A coding agent might repeatedly:

1. Read a large repository.
2. Search files.
3. Generate code.
4. Run tests.
5. Inspect errors.
6. Modify code.
7. Run tests again.
8. Document its changes.

A complex task can therefore consume millions of tokens.

For companies operating thousands of agent tasks every day, differences of several dollars per million tokens can translate into substantial monthly expenses.

This is why Google’s aggressive Argon pricing could put pressure on competitors even if benchmark differences remain relatively small.

Argon Is Not Yet a Drop-In Replacement for Astra

There is another major qualification.

GPT-6 Astra is already available through OpenAI’s API and multiple cloud platforms.

Gemini 4 Argon is initially restricted.

Google says broader availability will begin with paid API customers and Google AI Ultra subscribers, but it has not announced a specific general-release date in its initial announcement.

Developers therefore cannot yet assume they can immediately migrate production workloads from Astra to Argon.

Availability, rate limits, latency, tool support, reliability and enterprise contracts will all influence real-world adoption.

Benchmark Scores Are Not the Same as Real-World Performance

AI benchmark comparisons need to be interpreted carefully.

A model can perform exceptionally well on one evaluation while performing less well on another.

Results can also be affected by:

• Prompting strategy
• Reasoning effort
• Tool access
• Test environment
• Sampling settings
• Time limits
• Token budgets

Even the Vals Index cautions against treating one score as a complete measure of every possible use case. Its index aggregates different professional sectors and uses its own weighting methodology.

A software company, law firm and scientific laboratory may therefore reach very different conclusions about which model performs best for their specific workloads.

What Gemini 4 Argon Means for the AI Race

The significance of Argon is not simply that Google has produced another high-scoring model.

Three developments stand out.

First, Google is demonstrating that it can compete directly with the strongest frontier systems on complicated professional workloads.

Second, it is significantly increasing how long a model can continue generating and reasoning in a single trajectory.

Third, Google is applying aggressive pricing pressure.

If Argon’s benchmark performance translates reliably into production workloads once access expands, competing AI providers could face pressure to improve either capability, efficiency or pricing.

That competition could ultimately benefit developers and businesses by reducing the cost of advanced AI inference.

Bottom Line

Google’s Gemini 4 Argon has emerged as a serious competitor to OpenAI’s GPT-6 Astra.

Argon scores 77.9% versus Astra’s 74.1% on DeepSWE v1.1, 51.3% versus 41.4% on AutomationBench, and 68.90% versus 63.13% on the independent Vals Index.

Pricing is equally notable.

Argon will initially cost $2 per million input tokens and $10 per million output tokens, before rising to $4 and $20. GPT-6 Astra’s standard API pricing is $10 and $50 respectively.

Therefore, the claim that Argon operates at “40% of Astra’s cost” is accurate when comparing Google’s post-introductory token rates with Astra’s standard rates. During Argon’s introductory period, its listed token rates are actually only 20% of Astra’s.

However, Argon does not beat Astra on every benchmark, and per-token price does not equal the cost of completing a real-world task.

The clearest conclusion is that Gemini 4 Argon has reached the top tier of frontier AI, while its combination of strong professional-work benchmarks, unusually large output capacity and aggressive pricing could intensify competition among Google, OpenAI and other AI developers once the model becomes broadly available.