Assure your AI deployment - across enterprise and real world

We assure that your AI works as intended. Safely and securely.

AIAgent Testing

Gemini 4 Argon Explained: Benchmarks, Pricing, and Access

Gemini 4 Argon is now generally available. See its benchmarks against GPT-6 Astra and Claude Opus 5.5, API pricing, long-decode limits, and where to use it.

Last Updated on:

Gemini 4 Argon raises Google's output token limit to 1 million tokens, up from the previous 64K, according to Google's announcement. Google calls it its next frontier model, built to sustain deep reasoning across complex, long-horizon workflows.

Argon became globally available on 13 October 2026. Developers and enterprises can use it through the Gemini API in AI Studio and the Gemini platform, and Google AI Pro and Ultra subscribers can use it in the Gemini app.

For teams building on Gemini, Google's benchmark table shows where Argon trails rival models, and its API changes how long one agent run can take, what it costs, and how repeatable your evals are. TestMu AI checks for coding agents and conversational agents fit in before you switch a production agent over.

TL;DR

Gemini 4 Argon is Google's frontier AI model, built for long-horizon software engineering, legal and finance work, and cybersecurity defense. It became generally available on 13 October 2026 on the paid Gemini API tier and in the Gemini app, and it can extend one response to 1M tokens through long decode.

  • API availability: Can developers use Gemini 4 Argon today? Yes. It is on the paid tier of the Gemini API in Google AI Studio and on the Gemini platform, but not on the Gemini API free tier or the Google AI Pro plan.
  • App availability: Is Gemini 4 Argon in the Gemini app? Yes. Google AI Pro and Ultra subscribers can use it in the Gemini app and AI Mode in Search, and Google Antigravity access requires Google AI Ultra.
  • Pricing: Gemini 4 Argon costs an introductory $2 per million input tokens and $10 per million output tokens through 30 November 2026, then $4 and $20, with cached input 95% off the input price.
  • Output limit: Is Gemini 4 Argon's output limit larger than earlier Gemini models? Yes. By default a response stops at 64K tokens, but long decode extends it up to 1M tokens, and input and output share one 1,048,576-token limit.
  • Benchmarks: Does Gemini 4 Argon beat Claude Opus 5.5 on every benchmark? No. In Google's table it leads on most rows but trails Claude Opus 5.5 on Terminal-bench 4.0 and PostTrainBench.
  • Testing agents: TestMu AI Kane CLI checks what a Gemini 4 Argon coding agent ships in a real browser. TestMu AI Agent Testing evaluates chat, voice, and phone agents before you move them to a new model.

What Is Gemini 4 Argon?

Gemini 4 Argon is Google's new frontier model, announced on 30 September 2026 by Koray Kavukcuoglu, SVP of Google DeepMind and Chief AI Architect at Google. Google aims it at real-world software engineering, enterprise knowledge work such as legal and finance, and cybersecurity defense.

Google says thousands of Googlers already use Argon, and its launch posts list these internal results:

  • Agentic code submissions - in Google Antigravity, Argon helped Google teams raise agentic code submissions by more than 35% while cutting code rollbacks by nearly a third.
  • Quantum algorithms - optimizing the spacetime resources of subroutines that bottleneck quantum applications, where in one example it beat the published baseline by 40% in minutes.
  • Memory efficiency - a team of Argon agents analyzed fleet-wide profiling telemetry and freed over 300 TiB of memory across Google's data centers, with an estimated 500 TiB to 1 PiB in total savings.
  • C/C++ to Rust migrations - Argon agents are migrating codebases from tens of thousands of lines in libraries such as re2 and libgav1 up to 800K+ lines for the Fuchsia Zircon kernel, with automated and manual auditing before production.
  • Faster safe code - for libgav1, Argon replaced 32K lines of SIMD code in an existing Rust port, producing a memory-safe video decoder that runs 2.7x faster than that port with identical output.
Youtube thumbnail

Gemini 4 Argon Benchmarks: Where It Leads and Where It Trails

Google's announcement compares Argon with OpenAI's GPT-6 Astra and Anthropic's Claude Fable 5.1 and Claude Opus 5.5. These are Google's own reported scores, with the highest score in each row in bold:

BenchmarkGemini 4 ArgonGPT-6 AstraClaude Fable 5.1Claude Opus 5.5
Vals Index (knowledge work)68.9%63.1%65.8%67.0%
AutomationBench51.3%41.4%31.4%42.5%
Vals Finance Agent v265.4%53.5%58.9%58.6%
Harvey's Legal Agent Benchmark19.6%5.4%6.7%3.8%
DeepSWE v1.1 (agentic coding)77.9%74.1%67.4%74.2%
FrontierSWE v255.0%65.5%56.3%62.3%
Vibe Code Bench91.9%89.6%90.3%90.3%
Terminal-bench 4.057.4%58.2%57.9%66.4%
PostTrainBench (ML engineering)45.3%44.3%40.2%49.3%
Terminal-Bench Science 0.157.6%68.1%52.6%63.3%
LABBench 288.8%85.4%68.6%73.1%
RiemannBench76.0%72.0%65.6%69.6%
GraphWalks, up to 128K tokens99.7%98.7%91.4%90.6%
GraphWalks, 256K to 1M tokens84.2%71.8%65.0%66.8%
Agent's Last Exam (computer use, pass rate)39.5%34.2%Not reported38.2%
OSWorld-2.0 (offline subset)69.2%72.6%Not reportedNot reported
Chartography71.6%71.0%46.2%66.3%
LVBench (long video)91.7%87.5%79.7%83.7%
CWE-bench v1 (cybersecurity)68.0%68.0%58.0%67.0%

Argon's widest leads come in knowledge work and very long context:

  • Legal work - 19.6% on Harvey's Legal Agent Benchmark, against 6.7% for the next-best model.
  • Business automation - 51.3% on AutomationBench, Zapier's end-to-end business workflow benchmark, 8.8 points ahead of Claude Opus 5.5.
  • Very long context - 84.2% on GraphWalks between 256K and 1M tokens, against 71.8% for GPT-6 Astra.
  • Finance research - 65.4% on Vals Finance Agent v2, against 58.9% for Claude Fable 5.1.

It trails on several of the coding and computer-use rows that matter most to agent builders:

  • Terminal tasks - 57.4% on Terminal-bench 4.0, behind Claude Opus 5.5 at 66.4%.
  • Frontier software engineering - 55.0% on FrontierSWE v2, behind GPT-6 Astra at 65.5% and Claude Opus 5.5 at 62.3%.
  • Computer use - 69.2% on the OSWorld-2.0 offline subset, behind GPT-6 Astra at 72.6%.
  • Science and ML engineering - behind GPT-6 Astra on Terminal-Bench Science 0.1 and behind Claude Opus 5.5 on PostTrainBench.

Argon tops DeepSWE v1.1, Google's headline long-horizon coding result, yet a terminal-heavy workload may still run better on another model, so compare on your own tasks. For how Anthropic's model behaves in agent work, see Claude Opus 5.5. Microsoft AI's next model, MAI-Base-2, has finished pre-training but cannot be called yet.

How Does Gemini 4 Argon's 1M-Token Limit Work?

Argon has one 1,048,576-token limit shared by input, thinking, and output, according to Google's latest model guide. By default, a response stops at 64K output tokens. Past that, long decode generates the response in hops of up to 64K tokens, each continuing where the previous one stopped.

Google's long decode guide makes it opt-in: on the Interactions API you set background to true and max_output_tokens above 65,536, and a background interaction can then run for up to 24 hours.

Google's two guides spell out what that means for an agent or eval harness built on Argon:

  • Hop billing - each hop resends your prompt plus everything generated so far as input, so a 1,000-token prompt that produces 200,000 tokens takes four hops and counts 397,216 input tokens. Implicit caching discounts that repeated input by 95%.
  • Thinking-only hops - you pay for every hop, including hops that contain only thinking, and an interaction that uses all of max_output_tokens can end with empty text because the model was still thinking.
  • Harness defaults - a low max_output_tokens value set by an agent framework or evaluation tool, such as 8,192, stops the model before it finishes thinking, so check the value your harness actually sends.
  • When to skip it - Google recommends long decode for hard reasoning and for evaluation tools that stop at the output limit, but not for agents that call tools in many short turns, where large outputs fill the shared window fast.
  • Timeouts - hard prompts can take several minutes, so use background execution or set HTTP client timeouts to at least 15 minutes.
  • Grading - long outputs such as whole-library migrations need graders that check the complete result, like a full diff and a build, rather than spot checks of a few sections.

How Strong Is Gemini 4 Argon at Cybersecurity?

Google trained Argon for cybersecurity defense: it can autonomously find, validate, and patch critical software vulnerabilities, and it ties GPT-6 Astra for first place on CWE-bench v1, which measures vulnerability remediation.

For trusted defenders and its own teams, Google is releasing Argon without cyber guardrails. Wiz is already using it in its Scan for Good initiative, where the model found a critical vulnerability exposing sensitive personal information in healthcare software used by hospitals worldwide, a risk earlier frontier models had missed.

Access runs through the Fairwind Program, which gives governments, critical infrastructure operators, and core technology platforms early access to frontier models for defense. Google DeepMind says it works with over 650 partners, and a set of them can use Argon inside CodeMender, its code security agent.

  • Dual-use tasks only - partners may run authorized threat simulation, reverse engineering, and malware analysis for defensive and academic research.
  • Security controls - partners agree to user-level authentication and phishing-resistant MFA, and may grant Argon access only to internal cybersecurity, incident response, or penetration testing teams.
  • No resale - partner organizations cannot share, redistribute, or sell access to the model.

At general availability, Google began expanding Fairwind to hundreds of additional partners. Defenders who want Argon without cyber guardrails apply to the program, and Google admits them on a rolling basis.

What Safeguards Ship With Gemini 4 Argon?

Google's two launch posts list these safeguards:

  • Misuse - Argon is designed to refuse cyber and CBRN attack requests while preserving legitimate dual-use research, and Google monitors the model's internal activations to spot misuse.
  • Prompt injection - after automated AI red teaming and adversarial training, Argon leads Gray Swan's Indirect Prompt Injection benchmark, with attack success cut to 0.7% across 15 adaptive attempts.
  • Shortcuts - Google trained Argon to respect system boundaries, report negative findings honestly, and pause for human confirmation before irreversible steps, and shortcut attempts on coding benchmarks dropped to 2%.
  • Misalignment - monitors watch Argon's chain of thought and actions and stop execution when the model steps beyond the user's intentions.
  • Sandboxes - Google is isolating and sealing its sandboxed environments before high-risk training or evaluations begin.

Those controls protect Google's own deployment. Your agent still needs its own checks: prompt injection testing covers the attacks, and the OpenAI Hugging Face incident shows why a record an agent can write is weak evidence of what it did.

Note

Note: Google's monitors watch what Argon does inside Google's systems. For agents you build that call tools and change files, TestMu AI Agent Assurance grades each run against what actually changed. Try TestMu AI free!

Where Can You Use Gemini 4 Argon, and What Does It Cost?

Google's launch post lists where each audience can use Argon:

AudienceWhere Argon is available
DevelopersPaid tier of the Gemini API in Google AI Studio and Android Studio. Google AI Ultra subscribers can also try it in AI Studio and Google Antigravity.
EnterprisesGemini for business, the Gemini API, and Google Antigravity.
ConsumersGemini app and AI Mode in Search, for Google AI Pro and Ultra subscribers.
Cybersecurity defendersFairwind Program, which grants access without cyber guardrails on a rolling basis.
ScientistsApplications for 6 months of free access to Google Antigravity's highest usage tier, offered to 10,000 scientists subject to eligibility, capacity, and selection.

For how Argon performs as a coding agent next to Gemini 3.8 Flash on the same build tasks, see the Google Antigravity review.

Google's API documentation says Argon is not available on the Gemini API free tier or on the Google AI Pro plan, and it lists API prices on two schedules:

  • Through 30 November 2026 - $2 per million input tokens, $10 per million output tokens including thinking tokens, and $0.10 per million cached input tokens.
  • From 1 December 2026 - $4 per million input tokens, $20 per million output tokens, and $0.20 per million cached input tokens.
  • Implicit caching only - explicit context caching is not supported on Argon, and implicit caching discounts repeated input by 95% on cache hits.

How Should You Test Agents Built on Gemini 4 Argon?

Argon leads Google's table on DeepSWE v1.1, and its score there still leaves roughly one long-horizon coding task in five unsolved. Code an Argon agent writes needs checking before it ships, the same as code from any other model.

  • Benchmark on your own tasks - run your evaluation suite on Argon and on your current model, weighting the terminal and computer-use tasks where Google's table shows Argon trailing.
  • Repeat each eval case - Argon ignores temperature, top_p, and top_k, so you cannot pin sampling for a repeatable run. Run each case several times, score the pass rate, and test your workload at the low, medium, and high thinking levels.
  • Test long decode paths - check that your client joins every hop of a long response, handles an interaction that runs out of max_output_tokens, and sends a max_output_tokens value high enough that the model does not stop mid-thought.
  • Rerun prompt-injection tests - Google reports leading robustness on Gray Swan's benchmark, but your agent's tools and data differ, so attack your own setup.
  • Check refusals on security tasks - the generally available model keeps the cyber guardrails that Fairwind access removes, so test vulnerability-research prompts against the version you will actually get.
  • Verify what coding agents ship - check the running app, not only the diff and the unit tests.

Kane CLI covers that last check. It validates rendered UI in a real Chrome browser from a plain-English objective, runs from your terminal, a coding agent's loop, or CI, and returns a pass or fail backed by an evidence pack. The same pattern with Google's coding agent is shown in Gemini CLI with Kane CLI.

For chat, voice, and phone agents moving to a new model, Agent Testing scores conversations on 9 quality metrics, including hallucination, and its testing agents include a Security Researcher persona that probes for prompt injection and data exfiltration. The Agent Testing CLI runs those evaluations from a terminal or a CI pipeline.

Kane CLI - Testing Agent in Your Terminal

Getting Started With Gemini 4 Argon

Point your eval suite at the gemini-4-argon model ID on the paid Gemini API tier, and rerun the tasks your agent fails today on both Argon and your current model while the introductory price applies. Keep the UI checks in CI with Kane CLI remote execution so every model switch is tested the same way.

Testing an AI agent? TestMu AI's Agent Testing validates chat, voice and IVR agents before your customers talk to them.

To run your first test, start free, or
for a full walkthrough of the product.

Author

...

Chaitanya Sharma

Blogs: 23

  • Linkedin

Chaitanya Sharma is an AI Product Manager at TestMu AI, where he builds agentic AI capabilities focused on computer vision and multi-modality, moving testing beyond static script execution toward autonomous, agent-driven workflows. Before TestMu AI he shipped 135+ features at Sprinklr for a no-code community and website builder used by Fortune 500 enterprises including Dell, Samsung, and Polestar. At Policybazaar he led the zero-to-one launch of a digital lending and insurance marketplace embedded in Bahrain's dominant payments app, building a risk-intelligence engine that compressed loan-approval times by 80%. He explored machine learning and NLP through research at the University of Cambridge, and holds a B.Tech from Delhi Technological University.

Reviewer

...

Sirajuddin Khan

Reviewer

  • Linkedin

Sirajuddin Khan is Vice President of Product Management at TestMu AI, where he drives the company's agentic AI product strategy, building a suite of autonomous agents that includes Agentic Browsers and Agentic Visual Testing and shifting the unit of work from test execution to autonomous outcomes. One of the company's earliest product leaders, he has owned the roadmap for the high-performance execution cloud and grew the cross-browser testing products from early adoption to market leadership. He brings over a decade of experience across SaaS, B2B, and eCommerce, with earlier product roles at Wydr and ShopClues, where his catalog and search work cut delivery SLAs and lifted seller activity. Sirajuddin holds an MBA in Information Technology from Sikkim Manipal University and a B.Tech in Computer Science Engineering from Maharshi Dayanand University.

Add to Google preferred sources

Summarise with AI

Copied to Clipboard!
...

3000+ Browsers. One Platform.

See exactly how your site performs everywhere.

Try it free
...

Write Tests in Plain English with KaneAI

Create, debug, and evolve tests using natural language.

Try for free

Gemini 4 Argon FAQs

Did you find this page helpful?

More Related Blogs

TestMu AI forEnterprise

Get access to solutions built on Enterprise
grade security, privacy, & compliance

  • Advanced access controls
  • Advanced data retention rules
  • Advanced Local Testing
  • Premium Support options
  • Early access to beta features
  • Private Slack Channel
  • Unlimited Manual Accessibility DevTools Tests