77.9%
Gemini Argon for long-horizon engineering
Reported on DeepSWE v1.1 for multi-hour code migration, debugging, and algorithm optimization tasks.
Unofficial tracker · Last verified October 1, 2026
Google DeepMind’s first reported frontier model in the Gemini 4 line is built for long-horizon reasoning, mixed media, and workflows that run far beyond a short chat turn. We track access, pricing, and independent results daily, with sources linked for every important claim.
Looking for Google’s newer model for cybersecurity defense and long-horizon engineering?
Explore Gemini 4 CarbonAccess status
As of October 1, 2026, reported access is limited to Google’s internal teams and vetted defenders in the Fairwind Program. Paid API customers and Google AI Ultra subscribers may be prioritized next, but Google has not confirmed a public launch date.
Last checked October 1, 2026 · We update this section when Google publishes new information. Source: blog.google.
If Gemini 4 Argon opens to external users, access will likely depend on more than a sign-up button. Expect model-region availability, quota limits, safety review, and terms for workloads that use long context or automated tools. We will record the first verifiable public route here, including whether Gemini Argon appears in Google AI Studio, an API catalog, or a paid plan.
Core capabilities
The current picture is less about casual chat and more about long-running, high-context work where a model has to reason, inspect, revise, and complete a sequence of connected tasks.
77.9%
Reported on DeepSWE v1.1 for multi-hour code migration, debugging, and algorithm optimization tasks.
68.9%
Reported leadership on the Vals Index, with a focus on legal, finance, research, and end-to-end automation.
68.0%
Reported on CWE-bench v1, tied first among the compared systems for finding, validating, and patching vulnerabilities.
Model overview
Gemini 4 Argon is the reported first frontier model in Google DeepMind’s Gemini 4 series. It is designed for deep reasoning and very long workflows rather than short chat turns.
According to the current reports, Gemini Argon has been used internally for quantum algorithm optimization, datacenter memory optimization, and large-scale codebase migration. It is also described as handling long video, charts, and mixed media inputs with a stronger safety and alignment layer.
The headline specification is output length: up to 1,000,000 output tokens, compared with 64K in the previous generation. Gemini 4 Argon is not publicly available yet; this page combines official announcements and independent evaluations, and labels unverified figures as reported.
That context changes the practical question from whether Gemini Argon can answer a prompt to whether it can stay coherent across an entire project. A long codebase, a legal matter with many exhibits, or a research workflow with charts and video can be kept in one working session instead of repeatedly summarized. The benefit remains a hypothesis until public access and reproducible evaluations arrive.
For teams watching Gemini 4 Argon, the most important signals are not a single viral score or screenshot. Look for a public model name, documented limits, a stable API surface, and evaluation details that explain the prompt, tools, latency, and sampling method. Those details determine whether a reported Gemini Argon result can be repeated in a real product. Until then, this tracker separates confirmed announcements from informed reporting so readers can make decisions without treating early claims as a guarantee.
Commercial details
The timeline below separates reported access milestones from what Google has not announced. Pricing is introductory reporting, not a final public price sheet.
| Model | Input / 1M | Output / 1M | Max output | Availability |
|---|---|---|---|---|
| Gemini 4 Argon | $2 intro → $4 | $10 intro → $20 | 1,000,000 output tokens | Fairwind + internal; public API pending |
| GPT-6 Astra | $10 | $50 | ~128k | Public |
| Claude Opus 5.5 | $4 | $20 | ~128k+ | Public |
| Claude Fable 5.1 | $10 | $50 | High | Public |
Introductory pricing as reported; Google has not published final public pricing. Verify these figures against official sources before relying on them. A practical Gemini Argon cost estimate should also account for cached input, tool calls, output length, concurrency, and any enterprise terms that are not visible in an early headline price.
Independent evaluations
Knowledge work, long-horizon coding, cyber defense, and cost efficiency. Data is shown as reported and may change as evaluators publish new harnesses and results.
Benchmarks help define where Gemini 4 Argon may be useful, but they are not a substitute for testing your own workload. Compare the same prompts, context size, tool permissions, and success criteria before choosing a model for production. We keep missing values as “—” rather than filling the table with guesses, and we will revise the Gemini Argon comparison when a source publishes a reproducible result.
| Category | Benchmark | Gemini Argon | GPT-6 Astra | Claude Fable 5.1 | Claude Opus 5.5 |
|---|---|---|---|---|---|
| Knowledge work | Vals Index | 68.9% | — | — | — |
| Long-horizon coding | DeepSWE v1.1 | 77.9% | — | — | — |
| Cyber defense | CWE-bench v1 | 68.0% | — | — | — |
| Intelligence index | Artificial Analysis | 53 High | 53 High | ~58 | — |
| Hallucination rate | Independent checks | ~15% | — | — | — |
Gemini Argon is reported at High 53, level with GPT-6 Astra and behind Claude Opus 5.5 at roughly 58.
Early Gemini Argon results show strong Text Arena signals and notable cost efficiency, while the sample remains small.
A benchmark score is not a universal ranking. Gemini Argon results depend on the evaluator, prompt, tool access, latency target, and whether the task rewards persistence or short answers.
Gemini Argon may lead in knowledge work, long context, coding, and cyber defense; results vary by task, prompt, and harness.
Questions and answers
Short answers about Gemini 4 Argon access, pricing, capabilities, and what this independent tracker does and does not claim.
Gemini Argon is the reported first frontier model in Google DeepMind’s Gemini 4 series. It is described as a system for deep reasoning, mixed media, and long workflows, with a headline capacity of up to 1M output tokens.
Gemini 4 Argon is not publicly available according to the current tracker. As of October 1, 2026, reported access is limited to internal Google teams and vetted Fairwind Program defenders; paid API and Google AI Ultra access have not been announced.
Reported introductory Gemini Argon pricing is $2 per million input tokens and $10 per million output tokens, with a possible move to $4 and $20. Google has not published final public pricing, so treat these figures as as-reported estimates.
The reported Gemini 4 Argon differences are a much longer output limit, stronger coding and cyber-defense results, broader mixed-media workflows, and a staged release with additional safety and alignment monitoring.
Watch official Google and DeepMind announcements, check Google AI Studio, and follow this Gemini Argon tracker for changes. Teams preparing an API key, evaluation harness, and representative test suite will be ready if access opens.
No. This is an independent Gemini 4 Argon tracker. It is not affiliated with, endorsed by, or operated by Google, and it links to official and independent sources for every reported update.
Stay informed
One email when Gemini 4 Argon public API or Google AI Ultra availability is confirmed. No newsletter and no spam.