On October 1 at 5:03 AM (KST), Google unveiled Gemini 4 Argon, its next-generation frontier model. It emphasizes writing long code and carrying multi-step professional work through to the end, and its output limit grew from 64K to 1 million tokens. It took first place on Vals and Arena, and its launch-period API price is $2 per million input tokens and $10 per million output tokens. Inside Google, it is already being used to cut data center memory and migrate large amounts of code.
Source: X @sundarpichai
Artificial Analysis: the same 53 points at a lower cost per task
On the Artificial Analysis Intelligence Index, Gemini 4 Argon scored 53, the same as GPT-6 Astra and Claude Fable 5.1. It fell short of Claude Opus 5.5's 58. Cost per task was $1.99 for Argon, $3.26 for Astra, and $7.63 for Fable 5.1. Argon's $1.99 is based on the 50% launch discount price, and at list price it is $3.98.
Rather than setting a new record in top-tier reasoning, it came in at a similar score to competing models at a lower cost.
Source: X @ArtificialAnlys
First place on Vals' real-world evaluation
On the Vals Index, which covers financial research, building Excel financial models, coding, and legal and tax work, Argon was first at 68.90%. Sonnet 5.5 scored 67.04% and Opus 5.5 scored 66.97%. It was also first at 65.40% on Finance Agent v2, which does multi-step financial research, and second at 91.91% on Vibe Code Bench, which tests app-building ability.
Argon's strengths show in gathering scattered material into a deliverable and turning ideas into working apps.
Source: X @ValsAI
Floor plan reconstruction and Arena user preference
Andon Labs' Blueprint-Bench 2 is an evaluation where the model reconstructs an apartment floor plan from photos taken of the rooms in sequence. The scores for reconstructing the connections between rooms were 0.544 for Argon, 0.512 for Opus 5.5, and 0.497 for Astra. The human score was 0.59.
Source: X @andonlabs
It is also first in Arena user preference evaluations. In the default text overall ranking, which adjusts for style differences, Argon High took first place with 1,525 points. The result is based on 4,942 votes. Opus 5.5 High scored 1,504. It also did well on spatial understanding and on the answers people actually prefer.
Source: X @arena
Gemini 4 Argon API pricing
| Pricing tier | Input (per million tokens) | Output (per million tokens) | Cached input |
|---|---|---|---|
| Launch period | $2 | $10 | 95% off the input rate ($0.10) |
| After the launch period | $4 | $20 | 95% off the input rate ($0.20) |
The end date of the launch pricing has not been announced. Vals' cost per test at the standard rate was $15.68 for Argon and $32.14 for Opus 5.5. Argon had the higher overall score at about half the cost.
Who can use it first
For now, vetted cyber defense partners in the Fairwind program and Google's internal teams use it first. Google plans to expand to paid API customers and Google AI Ultra subscribers afterward, but has not announced specific dates. While it gathers feedback from early partners, it is strengthening misuse prevention, model behavior monitoring, and runtime security.
Source: X @demishassabis
Vending-Bench 2: running a vending business for a year
In Andon Labs' Vending-Bench 2, a model starts with $500 and runs a virtual vending machine business for a year, finding suppliers, negotiating by email, and managing inventory and pricing. It also handles refunds and delivery problems. Averaged over 6 runs, Argon's ending balance was $13,718, in third place. Astra scored $15,515 and GPT-6 Sol $14,428, and Argon's standard error was $3,100.
It is a result that lets you compare even the ability to hold on to the same goal for a long time and keep making judgments. See the Vending-Bench 2 results
Cyber defense: from finding vulnerabilities to fixing them
According to Google, Wiz used Argon to find a serious vulnerability in medical software used in hospitals around the world, a problem earlier frontier models missed. On CWE-bench v1, a vulnerability fixing evaluation, it was also tied for first at 68%. Grok 4.7 and GPT-6 Astra were also at 68%.
Google is giving vetted defense partners and internal teams a model with the cyber guardrails removed. By putting the model to work on real defense work first, Google is strengthening the safeguards for the general release.
Development work where Claude leads
On Vals Terminal-Bench 4.0, which performs terminal tasks, Argon scored 57.58%. Sonnet 5.5 scored 62.63% even when counting as failures 4 tasks that it refused and another model handled instead. On ProgramBench, which reproduces the full behavior of an existing program, the number of programs that passed all hidden tests was 5 out of 200 for Argon and 37 for Opus 5.5.
The models' strengths split between adding features to a new app and restoring the behavior of existing software.
Inside Google: 300TiB of data center memory freed up
Inside Google, multiple Argon agents analyzed performance and memory usage records across servers, found wasted parts, optimized them, and deployed the changes. As a result, Google says it saved more than 300TiB of data center memory. Google estimates that applying this approach more broadly could save a total of 500TiB to 1PiB.
From C/C++ to Rust, up to an 800,000-line kernel
Argon agents were also put to work migrating Google's C/C++ codebases to Rust. The targets were the regular expression library re2, the video decoder libgav1, and the Fuchsia Zircon kernel, which is over 800,000 lines. The kernel is the core of the operating system, the foundation programs use for memory and hardware. Important migration work goes through automated and manual audits, emulation testing, and code review.
libgav1: 32,000 lines of SIMD code replaced, 2.7x faster
In the video decoder libgav1, it replaced 32,000 lines of SIMD code in the existing Rust port. It repeated a cycle of measuring performance, analyzing compiler output, and then changing the code. In the end, it produced safe Rust code that the compiler can auto-vectorize, and while outputting the same video, it was 2.7 times faster than the existing Rust port.
1 million tokens of output
Argon's output limit grew from the previous 64K to 1 million tokens, about 15.6 times. Separate from the input limit, the model can sustain longer reasoning and generation along a single task path. Google explains that having room to generate hundreds of thousands of tokens helps in working through complex problems deeply.
The Gemini API also gained a Long Decode Continuation feature, which pauses a long response and continues generating in the next call. It can reason up to 1 million output tokens without a request timeout.
Source: X @Google
When access opens up to outside developers, they will be able to check such long tasks on their own codebases.