What Is Gemini 4 Argon? Gemini 4 vs Gemini 3 Pro Benchmarks and Features
Gemini 4 Argon is Google DeepMind’s newest frontier AI model, designed for long-running software engineering, enterprise knowledge work and defensive cybersecurity. Compared with Gemini 3 Pro, Google is placing much greater emphasis on agents that can stay inside a difficult task for a long time, operate tools, analyze large amounts of information and complete complex workflows instead of simply producing a strong single response.
Its most unusual specification is the maximum output limit. Google has increased it from the previous 64,000 tokens to one million tokens, giving Argon enough generation headroom for extremely long coding, research and agent trajectories. Google has not yet published a complete public developer specification for the model, and broader API availability is still pending.
What Is Gemini 4 Argon?
Gemini 4 Argon was announced on September 30, 2026. Google describes it as a frontier model built to sustain deep reasoning through complex, long-horizon professional tasks, particularly software engineering, legal and financial work, scientific analysis and cybersecurity defense.
Unlike a conventional chat model that answers a question and stops, Argon is being positioned as the intelligence layer for autonomous workflows. Google engineers are already using Argon internally for debugging, codebase migration, infrastructure optimization and algorithm development. One internal deployment used multiple Argon agents to analyze data-center telemetry and identify memory optimizations that freed more than 300 TiB of memory, with Google estimating substantially larger eventual savings.
A One-Million-Token Output Limit
The most visible change from Gemini 3 is not just reasoning quality but how long Argon can continue working.
Gemini 4 Argon supports up to one million output tokens in a single trajectory, compared with the previous 64K limit. That is approximately a 15.6-times increase in maximum output capacity. Google argues that this gives the model enough space to investigate a repository, run tools, revise code, produce documentation and continue correcting mistakes without forcing the workflow to terminate prematurely.
This does not mean developers should routinely generate one million tokens. Extremely long outputs can increase latency and cost. The importance of the feature is that the model can now support workflows whose natural execution path is much longer than a normal chat response.
What Google’s Own Benchmarks Actually Tell Us
Google’s official Gemini 4 Argon comparison is more useful than a simple leaderboard because it places Argon next to GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5 across coding, knowledge work, science, long context, computer use and multimodal understanding. The results show a genuinely strong model, but they also make one thing clear: Argon does not dominate every category.
| Benchmark | Gemini 4 Argon | GPT-6 Astra | Claude Fable 5.1 | Claude Opus 5.5 |
|---|---|---|---|---|
| Vals Index | 68.9% | 63.1% | 65.8% | 67.0% |
| AutomationBench | 51.3% | 41.4% | 31.4% | 42.5% |
| Vals Finance Agent v2 | 65.4% | 53.5% | 58.9% | 58.6% |
| Harvey’s Legal Agent Benchmark | 19.6% | 5.4% | 6.7% | 3.8% |
| DeepSWE v1.1 | 77.9% | 74.1% | 67.4% | 74.2% |
| FrontierSWE v2 | 55.0% | 65.5% | 56.3% | 62.3% |
| Terminal-Bench 4.0 | 57.4% | 58.2% | 57.9% | 66.4% |
| Terminal-Bench Science 0.1 | 57.6% | 68.1% | 52.6% | 63.3% |
| GraphWalks 256K–1M | 84.2% | 71.8% | 65.0% | 66.8% |
| Agent’s Last Exam | 39.5% | 34.2% | — | 38.2% |
| OSWorld 2.0 | 69.2% | 72.6% | — | — |
| Chartography | 71.6% | 71.0% | 46.2% | 66.3% |
| LVBench | 91.7% | 87.5% | 79.7% | 83.7% |
| CWE-bench v1 | 68.0% | 68.0% | 58.0% | 67.0% |
Source: Google DeepMind’s official Gemini 4 Argon benchmark table.
The strongest pattern is not “Gemini 4 wins everything.” Its clearest advantages are enterprise knowledge work, long-context reasoning and several multimodal tasks. Argon leads Vals Index, AutomationBench, Finance Agent v2 and Harvey’s Legal Agent Benchmark, while its 84.2% result on GraphWalks between 256K and 1M tokens is substantially ahead of the other models Google tested.
Coding is more complicated. Argon leads DeepSWE v1.1 at 77.9% and performs extremely well on Vibe Code Bench, but it loses to GPT-6 Astra and Claude Opus 5.5 on FrontierSWE v2. Claude Opus 5.5 also has a sizable advantage on Terminal-Bench 4.0, while GPT-6 Astra leads Terminal-Bench Science 0.1. This suggests that Gemini 4 may be exceptionally strong at long-horizon repository work without necessarily being the best model for every terminal or scientific coding workflow.
Gemini 4 vs Gemini 3: The Comparison Is Less Straightforward Than It Looks
There is one important omission in Google’s Gemini 4 launch materials: the official benchmark table does not include Gemini 3 Pro.
That makes a precise generation-over-generation comparison surprisingly difficult. Gemini 3 Pro was evaluated on a somewhat different benchmark suite. Google previously reported 76.2% on SWE-Bench Verified, 56.9% on Terminal-Bench 2.0, 91.9% on GPQA Diamond and 37.5% on Humanity’s Last Exam. Gemini 4 Argon is now being evaluated much more heavily on DeepSWE, FrontierSWE, AutomationBench, Vals and long-context agent benchmarks.
This change in benchmark selection is itself revealing. Gemini 3 was marketed largely around reasoning, multimodal understanding and coding quality. Gemini 4 is being judged much more heavily on whether it can complete long, economically useful workflows.
However, it also raises a reasonable question: if Gemini 4 represents such a large generational improvement, why did Google not publish Gemini 3 results using the same harnesses and benchmark versions as Argon?
Without that matched comparison, statements such as “Gemini 4 improves coding by X% over Gemini 3” would be difficult to justify. The safer conclusion is that Google has changed both the model and the type of work it considers important.
The 1M Output Limit Sounds Impressive — But Will Developers Actually Use It?
The headline Gemini 4 specification is its one-million-token maximum output, up from 64K. Google argues that this allows Argon to continue reasoning, calling tools and revising work over much longer trajectories. Technically, this is impressive. Economically, the question is more complicated. At the introductory $10 per million output tokens, a trajectory that actually consumes the full allowance would cost $10 in output alone. After introductory pricing ends, Google says output pricing will rise to $20 per million tokens. That excludes input, tools and any repeated attempts.
The more practical metric is therefore not “Can Gemini 4 generate one million tokens?” but “How many tokens does Gemini 4 need to successfully complete a task?” If Argon completes a six-hour coding workflow with half the tool calls and fewer failed attempts than Gemini 3, the larger output allowance is valuable. If it simply produces much longer reasoning trajectories, the feature could increase cost and latency without delivering proportional value.
Google has not yet published enough production data to answer that question.
Another Missing Number: What Is the Actual Input Context Window?
This is probably the most confusing part of the launch. Google repeatedly emphasizes a one-million-token output limit, but its announcement does not state the official maximum input context window. Third-party listings currently describe Argon as having a one-million-token context window, but Google has not confirmed that specification in the developer documentation available today.
Those are two very different concepts. An
capable of generating one million tokens is useful, but enterprise applications also need to know how much repository code, documentation, conversation history and tool output can be fed into the model at once.
Until Google publishes the full API specification, developers should avoid casually describing Argon as having a “1M context window.” The confirmed specification is currently a 1M maximum output.
Are Google’s Internal Success Stories Representative?
Google provides some extraordinary internal examples. Argon agents reportedly identified data-center changes that freed more than 300 TiB of memory, with eventual savings potentially reaching 500 TiB to 1 PiB. Google also says Argon has been used for migrations of C/C++ projects to Rust, including codebases approaching 800,000 lines, and that an Argon workflow produced a Rust implementation of libgav1 that was 2.7 times faster than the previous Rust port.
These examples are genuinely interesting, but they should not be treated like independent benchmarks. Google controls the model, infrastructure, tools, datasets and engineering environment used in these projects. The company also explicitly notes that critical migrations still go through automated testing, emulation, manual audits and human review before deployment. blog.google
For ordinary developers, the unanswered questions are how much human intervention was required, how many agent attempts failed, how many tokens were consumed, how long the jobs ran and what the total inference cost was.
Those numbers would tell us much more about Argon’s practical productivity than the headline “800K-line migration.”
Cybersecurity Performance May Not Be the Same for Every User
Gemini 4 Argon is also being launched in an unusual way because of its cybersecurity capabilities. Google says trusted defenders in the Fairwind Program will receive Argon without cyber guardrails so they can use its full vulnerability-discovery and remediation capabilities. Public and broader enterprise releases will include strengthened safeguards designed to prevent misuse.
That raises another practical issue: will the version ordinary API customers eventually receive reproduce the same cybersecurity benchmark results? Argon scores 68% on CWE-bench v1, tying GPT-6 Astra in Google’s comparison. Google also reports 85.8% on its real-world vulnerability-discovery benchmark versus 71.0% for Gemini 3.8 Flash Cyber, and 70.9% versus 58.2% on Wiz’s penetration-testing benchmark.
Those results are impressive, but developers need to know which safeguard configuration was evaluated. A restricted Fairwind model and a general commercial API model may behave differently on exactly the kinds of tasks these benchmarks test.
Price per Token Is Attractive, but Cost per Task Will Matter More
Gemini 4 Argon’s introductory pricing of $2 input and $10 output per million tokens looks extremely competitive for a frontier model. Cached inputs receive a 95% discount, while the eventual standard rate will rise to $4 input and $20 output.
For long-running agents, however, price per million tokens is only the starting point. A real production comparison should measure total tokens, cache-hit rate, number of tool calls, wall-clock time, failed attempts and human interventions required to finish the same task. Argon could be cheaper than a more expensive competitor even if it uses more tokens, or considerably more expensive than its headline rate suggests if a long-horizon agent generates hundreds of thousands of unnecessary tokens.
This is why the most interesting Gemini 4 test will not be another static benchmark. It will be running the same software-engineering or research task on Gemini 4 Argon, GPT-6 Astra and Claude Opus/Fable and measuring the total cost of reaching a successful result.
A More Practical Verdict on Gemini 4 Argon
Google’s own numbers make a convincing case that Gemini 4 Argon is a frontier model. It is particularly impressive in long context, enterprise knowledge work, multimodal reasoning and several forms of agentic coding. At the same time, the official table also shows clear areas where competing models remain stronger, including Terminal-Bench 4.0, FrontierSWE v2, Terminal-Bench Science and OSWorld.
That makes the launch more interesting, not less. Gemini 4 Argon does not appear to be simply “Gemini 3 but smarter.” Google is optimizing for a different definition of intelligence: whether a model can remain useful across a long, messy workflow involving code, tools, documents and real business software.
The major unanswered questions are now operational rather than academic: latency, throughput, actual input context, public API limits, cost per completed task and whether the broader commercial version matches the capability demonstrated by Google’s trusted-access deployments. Until developers can test those factors independently, Gemini 4 should be viewed as an extremely promising frontier agent model rather than an automatic replacement for every existing model.
Final Thoughts
Gemini 4 Argon looks like a more fundamental change than a routine model upgrade. Though Gemini 3 proved that Google could compete at the top of reasoning, multimodal understanding and coding, 4th takes the next step by focusing on whether an AI system can remain productive inside a complex task for a much longer period of time.
Its one-million-token output limit, strong DeepSWE and AutomationBench results, leading Arena debut and focus on enterprise agents all point in the same direction: Google wants Gemini to move from answering difficult questions to completing difficult work.