Google has introduced Gemini 4 Argon, its first Gemini 4 model, focusing on coding, research, cybersecurity and long-running tasks. The model can generate up to 1 million tokens in a single response and is being rolled out gradually through a restricted safety program.
Google has introduced Gemini 4 Argon, the first model in its Gemini 4 generation, with a clear focus on tasks that require more than a quick answer.
The company is positioning Argon around software engineering, business research, legal and financial work, long-running projects, multimodal understanding and cybersecurity.
Google announced Gemini 4 Argon on September 30, 2026, at an important moment for the technology industry.
Google is competing with companies such as OpenAI and Anthropic, both of which have recently introduced new models aimed at coding, professional work, and long-running computer tasks.
Google's response with Argon is notable not only for its performance claims, but also for how far the company has pushed the model's output in a single run.
Argon will launch at an introductory price of $2 per million input tokens and $10 per million output tokens, with cached input tokens priced at 95% off the input token price.
In a major architectural shift, Gemini 4 Argon can generate up to 1 million tokens in a single response, up from 64k in previous versions.
Combined with the Gemini API's Long Decode Continuation feature, it can navigate complex, hundreds-of-thousands-of-tokens-long reasoning trajectories without timing out.
Argon can natively detect, verify, and fix critical software vulnerabilities.
In internal tests, it successfully flagged a live cross-hacker vulnerability in global healthcare infrastructure that other frontier models completely missed.
It is also highly resilient to prompt injections and autonomous misalignment.
Instead of just outputting short script snippets, the model handles large-scale codebase migrations, complex algorithm optimisations, and porting massive legacy C/C++ projects directly into Rust.
On the AA-Omniscience benchmark, Argon posted a 15% hallucination rate, the lowest among all top-tier frontier models.
It prioritises acknowledging when it doesn't know an answer over confidently guessing.
On the LVBench, it hit a state-of-the-art score of 91.7%, showcasing an elite ability to analyse data across hours-long video streams or complex stacks of visual business charts.
Loading post from https://x.com/Google/status/2105388143902175529…
Gemini 4 Argon, GPT-6 Astra, and Claude Opus 5.5 represent the peak tier of frontier AI, each engineered with a distinct competitive edge.
Gemini 4 Argon leads in raw scale and specialised technical domains, boasting an unprecedented 1 million output token limit and a dominant 77.9% score on the DeepSWE v1.1 coding benchmark. However, it slightly lags in rapid terminal tool navigation.
GPT-6 Astra serves as the primary benchmark standard on the Artificial Analysis Intelligence Index, matching Argon closely on cybersecurity tracking while slightly trailing both competitors in long-horizon business logic automation.
Claude Opus 5.5 remains the undisputed king of agentic tool utilisation, outclassing both rival models with a 66.4% score on Terminal-Bench 4.0, making it exceptionally adept at executing multi-step terminal workflows despite sitting just behind Argon in macroeconomic knowledge work.
Gemini 4 Argon is not yet broadly available to the public or standard developers, as Google is executing a highly restricted, multi-phase safety rollout.
Because the model has advanced autonomous cybersecurity capabilities, Google heavily manages its distribution to mitigate risk.
Access is strictly limited to select cybersecurity defenders via Google's invitation-only Google Fairwind Program.
During this phase, Google is also completing the U.S. government's voluntary pre-release model check processes to verify the model's safety guardrails.
The rollout will expand to paid API customers and subscribers to the premium Google AI Ultra tiers, priced at $100 per month and $200 per month.
A wider public release for standard developers and consumers will follow "as soon as possible" once safety and alignment testing are fully finalised.
Google has not yet committed to a specific public calendar date.
Gemini 4 Argon is still in its early stages.
Its headline specifications are impressive, and Google's benchmark results show strong performance across several areas.
Still, independent testing and broader access will ultimately provide a much clearer picture of how it compares with competing systems in everyday use.
