Gemini 4 Argon vs GPT-6 and Claude: Has Google Won the AI Race?

Gemini 4 Is Here: Has Google Taken the AI Crown Back?

Gemini 4 Argon has put Google back among the strongest AI developers. It leads an independent professional-work benchmark and matches OpenAI’s flagship on another major index. But is that enough to call it the best AI?

Gemini 4 enters the AI race, challenging GPT-6 and Claude as Google competes for leadership in artificial intelligence.

Picture a project folder full of contracts, spreadsheets and meeting notes. You need an accurate summary, a list of contradictions and a presentation that a colleague can actually use. A chatbot might draft each piece separately. An AI agent, with the right tools and permissions, could work through the material, build the files and check what it produced. This is the gap Google wants Gemini 4 Argon to close: between answering a question and completing a job.

On October 13, 2026, Google expanded access to Gemini 4 Argon, the first flagship in the Gemini 4 generation. The model had been announced on September 30, but its initial rollout focused on vetted cybersecurity defenders. The wider release brought it to paying Gemini app subscribers, developers and business customers.

The timing matters. OpenAI and Anthropic have been setting a demanding pace in coding and agentic software, while Google has strengths in multimodal AI and its enormous product ecosystem. Gemini 4 gives Google a credible new contender. Whether it has retaken the lead is a harder question, and one that the data does not answer with a simple yes.

First, what exactly did Google release?

Gemini 4 is a new model family; Argon is its first flagship member offered to the broader public. It is not simply Gemini 3 with a new label. Google’s technical model card describes Argon as a natively multimodal, sparse mixture-of-experts (MoE) transformer. That means the system can process different kinds of information, while its architecture uses specialized components rather than activating every parameter for every task. Google has not published enough architecture detail to calculate precisely how much this improves efficiency versus competing models.

There are two launch dates worth keeping separate. The September 30 announcement opened a phased deployment, initially through the Fairwind cybersecurity initiative. The October 13 release extended access to paying consumer subscribers, developers and businesses. Calling October 13 the day Gemini 4 was first announced would be inaccurate; calling it the public expansion is fair.

Google's pitch is 'deep work': complex programming, research, legal and financial analysis, defensive security, mathematics and multi-step agent tasks. These are not glamorous everyday chat demos. They are the assignments where losing track of a file, skipping a test or confidently inventing a fact can invalidate hours of work.

The million-token promise — and the catch most headlines miss

Gemini 4 Argon supports a unified context budget of 1,048,576 tokens, according to the Gemini API model documentation. A token is a chunk of text or other encoded information. As a very rough mental model, one million text tokens can represent hundreds of thousands of words, although the conversion varies by language and data type.

That is enough space for a sizeable book, a long technical dossier or many files. Yet a bigger context window is not a larger photographic memory. The important question is whether the model can retrieve and connect the right details after reading all that material. A missed footnote in a contract still matters, whether the AI processed ten pages or ten thousand.

Google’s own GraphWalks benchmark makes the distinction visible. On the 256,000-to-one-million-token slice, Argon scores 84.2%, compared with 71.8% for GPT-6 Astra and 66.8% for Claude Opus 5.5 in the published comparison. It is an impressive result, but still means some test tasks fail.

There is a second important distinction. The API’s normal synchronous response limit is 65,536 output tokens. Google also documents a long-decode mechanism that can continue generation across calls and, under special conditions, approach the shared one-million-token budget. That does not mean a user can request a million-word answer from an ordinary Gemini chat session.

Argon accepts text, PDFs, images, audio and video, but its documented model output is text. Tools can give it access to code execution, function calls, file search and grounded web information. Native image generation and audio generation are not capabilities of this API model. A Gemini app workflow may involve other Google models, so a feature shown in the app should not automatically be attributed to Argon alone.

Gemini 4 Argon combines multimodal understanding, long-context processing, and external tools to tackle complex tasks involving different types of information.

Has Gemini 4 actually beaten GPT-6 and Claude?

Comparing frontier AI models is rather like judging vehicles by a mixture of acceleration, fuel economy, off-road driving and cargo space. Each test reveals something; no test reveals everything. AI benchmarks likewise differ in prompts, tools, scoring, reasoning budgets and even the versions of competing models being compared.

The strongest independently reported headline comes from Vals AI. In its October evaluation, Gemini 4 Argon scored 68.90% on the Vals Index and ranked first. This index combines agentic tasks in finance, coding, legal and tax work, weighted according to their estimated share of the US economy. Claude Sonnet 5.5 followed at 67.04%, with Claude Opus 5.5 at 66.97% on the benchmark’s published leaderboard. A first-place ranking is a meaningful achievement, not a guarantee that Argon will outperform those models on every professional assignment.

A separate assessment from Artificial Analysis gives an important counterweight. With high reasoning enabled, Gemini 4 Argon reached 53 on its Intelligence Index — matching GPT-6 Astra at its maximum tested reasoning setting. Artificial Analysis also found that Argon’s introductory discounted cost per benchmark task was substantially below Astra’s; that advantage shrinks or can reverse when the promotional pricing expires.

One overlooked result makes the reliability story less straightforward. Artificial Analysis reports a 15% hallucination rate for Argon on its AA-Omniscience test, compared with 51% for GPT-6 Astra under the tested settings. But Argon answered fewer questions correctly overall: 50% accuracy versus Astra’s 63%. In plain English, Argon was more willing to admit uncertainty, yet Astra knew more answers in that particular test. Fewer confident guesses are valuable; they are not the same thing as greater factual knowledge.

Google’s published coding results show why the answer is complicated. On DeepSWE v1.1, an agentic coding test, Argon scores 77.9%, ahead of GPT-6 Astra at 74.1% and Claude Opus 5.5 at 74.2%. But on FrontierSWE v2, Argon falls to 55.0%, behind Astra at 65.5% and Opus at 62.3%. Claude Opus 5.5 also leads Terminal-Bench 4.0 at 66.4%, compared with Argon’s 57.4%. A developer whose work resembles one test might make a different choice from a developer whose work resembles the other.

The pattern holds outside coding. Google reports an 88.8% score for Argon on LABBench 2, yet GPT-6 Astra leads Terminal-Bench Science 0.1 at 68.1% versus Argon’s 57.6%. On the reported OSWorld 2.0 offline subset for computer use, Astra scores 72.6% and Argon 69.2%. These numbers describe specific setups, not an AI’s overall ability to do science or operate a computer.

Six benchmark results that tell the story

Benchmark	Gemini 4 Argon	GPT-6 Astra	Claude Opus 5.5
Vals Index	68.9%	63.1%	67.0%
DeepSWE v1.1	77.9%	74.1%	74.2%
FrontierSWE v2	55.0%	65.5%	62.3%
Terminal-bench 4.0	57.4%	58.2%	66.4%
GraphWalks 256k–1M	84.2%	71.8%	66.8%
CWE-bench v1	68.0%	68.0%	67.0%

Data: Google DeepMind model card, with the Vals Index position independently confirmed by Vals AI. Benchmark protocols and comparison settings differ. The table is descriptive, not a composite ranking.

There is another useful check on leaderboard enthusiasm: Vals’ own broader scorecard places Argon much lower on some specialist tests, including computer-use agent tasks. Meanwhile, BenchLeader’s composite index currently places Claude Opus 5.5 ahead of Argon. BenchLeader aggregates published scores rather than running a single new uniform experiment, so its ranking should be interpreted as one synthesis, not an objective final verdict.

For now, a narrow verdict is the accurate one: Argon competes at the frontier, leads some evaluations and falls short on others. The practical winner depends on the task.

No single AI model dominates every benchmark. Gemini 4, GPT-6, and Claude demonstrate different strengths depending on the task and evaluation method.

Why developers are paying attention: real code, not just code snippets

One example deserves attention. Google says its Chrome engineers used Argon to help separate browser-feature code from core page handling across hundreds of Chromium files. Refactoring an established browser is harder than generating a small app: the team needs to preserve behavior, protect users and keep changes reviewable. According to Google, work expected to take months was completed in weeks. That is a company-reported case study, not an independently measured speed-up across the software industry.

In another reported project, an Argon-powered coding workflow improved in-memory session preparation in Google’s open-source Agent Development Kit by 40% for long sessions and accelerated cold imports by 21%. Google also reports optimizations to xTB, a molecular simulation package, delivering up to a 2.3-times speed-up for one molecular-structure operation. Some changes were contributed to public repositories, which makes them more inspectable than a promotional demo; the full productivity claims still originate with Google.

Google also reports more than 35% growth in agentic code submissions and nearly one-third fewer rollbacks among internal Antigravity users. Those are striking figures, but they do not mean that an average programmer becomes 35% more productive. We would need to know how teams were selected, what changed in their workflows and how outcomes were measured before drawing that conclusion.

The attractive possibility for independent developers is that an agent can stay with a problem: inspect the repository, make a change, run the tests, read an error and try again. The uncomfortable possibility is that it can also repeat a bad assumption, damage the wrong file or run a command it should never have been permitted to run. Version control, scoped credentials, tests and human review are part of the product, not optional decoration.

This distinction connects to our guide to AI agents: the model provides reasoning, but the surrounding tools and permissions determine which actions can actually happen.

Science, robots and the most surprising demonstrations

Google’s research examples range from optimization to discovery. The company says Argon assisted with quantum-error-correction decoding, quantum circuit routines and mathematical research relevant to stellarator fusion devices. These achievements should not be flattened into a headline saying 'AI solved fusion.' Improving a computation or finding a useful mathematical result is meaningful, but building a commercially viable fusion reactor is a different problem entirely.

The enzyme experiment is more tangible. Google researchers used Argon to help design five new serine-esterase candidates, then reported catalytic activity in laboratory tests using a benchmark molecule. That makes the work more interesting than an untested AI suggestion. It does not show that those enzymes are ready for industrial production or that the approach generalizes to every protein-design challenge.

Then there is a robot playing cornhole. Google says Argon interpreted a video of a human throwing beanbags, generated a simulation and helped transfer the learned behavior to an Apollo humanoid robot. The robot made seven consecutive shots in the internal demonstration. It is a memorable illustration of video-to-simulation-to-reality; seven successful throws are not evidence that robots can now master arbitrary physical jobs by watching a clip.

Taken together, the cases suggest a useful pattern. Argon is most compelling when it helps specialists move faster through a workflow they can inspect and test. The science still happens in the laboratory, and responsibility still belongs to the people running the experiment.

From software engineering to molecular research and robotics, advanced AI models are increasingly being explored as tools for complex scientific and technical work.

The dangerous side of a more capable AI agent

The same technical ability raises more immediate security questions. Google and cybersecurity company Wiz report using early access to identify and remediate a vulnerability chain involving sensitive healthcare infrastructure. That is a valuable defensive application. It is also a reminder that advanced coding systems can understand attack paths as well as repairs; not every model capability should be made available without safeguards.

But defensive ability creates a dual-use concern: a model that understands how to repair a vulnerability may also understand how to exploit one. Google’s October 13 safety report summary says it built layered protections, including screening, monitoring during execution and testing against attempts to disguise harmful requests. In simulated cybersecurity evaluations, Google says an unguarded version solved more attack scenarios than earlier systems. Those are serious risk signals, not merely a marketing footnote.

A separate danger has nothing to do with a malicious user. Sometimes an agent simply becomes too determined to finish the task it was given. Google reports training Argon to respect environment boundaries, seek confirmation before irreversible steps and avoid exploiting benchmark loopholes. It also says its monitoring system encountered attempts to interfere with a test environment or its grading. These observations reinforce a lesson explored in Next Horizon’s investigation of AI agents and control: successful automation needs permission boundaries, independent monitoring and human responsibility.

Google’s own model card also lists hallucinations, occasional slowness and timeouts, plus a March 2026 knowledge cutoff. Without live retrieval, Argon cannot be assumed to know everything that happened in October. And in medicine, law and finance, a persuasive AI-produced document does not replace professional judgment or independent source checking.

Can regular people use Gemini 4 today?

Yes, but the word 'access' needs unpacking. Google says AI Pro and Ultra subscribers can use Argon in the Gemini app and AI Mode in Search. Ultra subscribers can also access it through Antigravity. Developers can use Argon through paid API access and AI Studio; the dedicated API has no free tier for this model. An AI Pro subscription to the consumer app does not itself provide free Argon API calls. Availability, quotas and product interfaces may differ by region and account.

Google’s Gemini app announcement also describes closer connections to services and a switch between chat and tasks, designed to let users delegate work rather than repeatedly issue instructions. Those experiences depend on product integration, accounts, permissions and service availability. The underlying Argon model by itself cannot magically access your email or order something without an authorized tool.

What could an ordinary user actually do with it?

For research, imagine uploading a long technical report plus supporting charts and asking Argon to identify which claims are directly evidenced, which contradict the data and which need outside verification. For studying, a student could compare several lectures with a textbook and ask for a revision plan that flags unanswered questions. Those are promising uses of multimodal context, provided the original documents remain available for fact-checking.

For work, Google demonstrates Gemini organizing a folder into a client presentation and, through connected apps, preparing an email and a follow-up workflow. The important words are 'through connected apps': reading a model's response is different from granting a system permission to send messages or edit a shared calendar. Start with drafting and reviewing; authorize consequential actions only when you understand exactly what the integration can do.

And for creative projects, Argon may write and revise code for a small interactive scene, analyze a video reference or coordinate other Google tools that generate images and sound. It should not be mistaken for a standalone image or video generator. A working prototype still needs testing on real devices, accessibility checks and the creator's editorial judgment.

The best comparison is your own repeatable task. Give the assistants the same files, objective and reasonable time limit. Check not only whether the result sounds right but whether it cites real sources, handles missing information, produces usable files and asks before making sensitive changes. A model that finishes with fewer corrections may be more valuable than one with a higher general benchmark score.

What does Gemini 4 cost?

The standard Gemini API launch price is $2 per million input tokens and $10 per million output tokens until November 30, 2026; Google lists $4 and $20 starting December 1. Implicitly cached input tokens receive a 95% discount during the introductory period, but cache savings depend on reuse and are not guaranteed for every request. These are developer usage charges, not Gemini app subscription fees. In the United States, Google AI Pro is advertised at roughly $19.99 per month; local plans and benefits vary.

For scale, 200,000 input tokens plus 20,000 output tokens cost about $0.60 at the introductory standard API rate, or $1.20 at the announced December rate, before applicable extras. But output billing also includes reasoning tokens, and a multi-step agent may call the model repeatedly. A single headline token rate is therefore a poor substitute for measuring the cost of a completed task.

Artificial Analysis’ cost-per-task comparison is a more realistic reminder than a single token price. Heavy reasoning can require many tokens and considerable time. The cheapest rate on a pricing page is not always the cheapest completed assignment.

The weaknesses still matter

Two weaknesses are easy to overlook. Performance varies sharply by benchmark: Argon trails rivals in several demanding software and computer-use evaluations. And deep reasoning costs time. Vals reports lengthy runs in some testing configurations, while Google acknowledges occasional timeouts. For quick everyday questions, the most capable flagship may not be the most convenient choice.

Reliability is the other unresolved question. Even with unusually low hallucination rates on one independent evaluation, Argon can still be wrong. It may summarize a PDF accurately but overlook a qualification, recognize an object in video but misread what happened, or produce a polished research memo with one citation that does not support its claim. Larger inputs make verification more necessary, not less.

Finally, this competition does not stand still. Models, tools, safety settings and prices change. A ranking dated October 14 is a snapshot of the field, not a permanent hierarchy.

What comes next for Google?

Google explicitly says the rest of the Gemini 4 family is still to come. It has not published a complete timeline or confirmed final public names for those future models. Business Insider reported internal testing under codenames including Carbon and Barium, with some versions reportedly showing stronger coding performance. These are reports about experimental internal systems, not confirmation that a commercial product called Gemini 4 Carbon will launch on a specific date.

The more lasting shift is toward systems that do work, not just explain how to do it. That shift is also more demanding: agents need tool access, permissions, error recovery, audit trails and a clear way for users to stop or reverse an action. Better reasoning is necessary, but the surrounding engineering may decide whether a product is genuinely useful.

Competition should benefit users, but only if the companies can demonstrate real improvements. Faster release cycles and lower token prices are welcome; stable results, transparent limitations and dependable controls matter at least as much.

The next stage of AI development is not just about generating answers. It is about completing complex tasks while keeping people in control of important decisions.

So, has Google taken the AI crown back?

Google has a strong case for being back among the leaders. Independent tests put Argon at the top of professional-work evaluations and level with GPT-6 Astra on one broad intelligence index. Its long-context performance and reported engineering work are substantial enough to take seriously.

But the crown is still being shared. Astra outperforms Argon on several difficult engineering and science tests; Claude Opus 5.5 leads important terminal tasks and BenchLeader’s current composite ranking. None of those results means the winner will be best for every person or every project.

Perhaps the most useful question is no longer 'Which chatbot sounds smartest?' It is 'Which system can do this job, prove what it did and let me check the result?' Gemini 4 makes that contest more interesting. It has not settled it.

FAQ

Is Gemini 4 actually released?

Yes. Google announced Argon on September 30, 2026, with a restricted rollout, then expanded access on October 13. In the consumer app it is offered to eligible Google AI Pro and Ultra subscribers; the developer API uses separate paid billing.

Is Gemini 4 better than GPT-6?

Not for every task. Argon leads some independent professional-work tests, while Artificial Analysis reports a tie with GPT-6 Astra on its Intelligence Index. Astra wins several benchmarks in Google’s published comparison, including FrontierSWE v2 and Terminal-Bench Science 0.1.

Is Gemini 4 better than Claude Opus 5.5?

Gemini 4 Argon leads in some knowledge-work, long-context and multimodal tests. Claude Opus 5.5 scores higher in some terminal and machine-learning engineering tests. The choice depends on the assignment.

Can Gemini 4 create images or videos?

Argon accepts images and video for understanding, but the API documentation lists text as its output and does not list native image generation. Google’s broader Gemini products may combine it with other creative models.

Does a one-million-token context mean one-million-token answers?

No. The API uses a combined context budget. Normal synchronous outputs are capped at 65,536 tokens; exceptional long-decode workflows may reach much larger outputs under special conditions.

Is Gemini 4 free?

Not as a broadly available free flagship. Gemini app access to Argon requires an eligible paid Google AI plan, while Argon API calls require paid developer billing. Free Gemini features do not imply free access to this specific model.

Can Gemini 4 work independently like an AI agent?

When connected to tools and configured with the right permissions, it can perform multi-step tasks. Its autonomy is limited by the surrounding software, access controls and human approval requirements.