Gemini 4 Argon launches with 1M-token output cap and $2/$10 intro API pricing

What was announced

According to a dev.to article, Google announced Gemini 4 Argon on September 30, 2026, positioning it as a frontier model for complex software engineering, knowledge work, and defensive cybersecurity. The headline specification is not a leaderboard score but the output limit: Google says Argon can generate up to 1 million tokens in a single response, a sharp rise from the 64K ceiling cited for the previous setup.

Access is deliberately narrow for now. Google is rolling the model out first to trusted cyber defenders and testers through what the report calls the Fairwind Program, with paid API customers and Google AI Ultra subscribers to follow. The announcement does not specify a general availability date, rate limits, latency figures, or a complete input context window, so those remain open questions rather than confirmed capabilities.

Why a million output tokens is unusual

Most model comparisons revolve around input context, meaning how much a model can read. Google is instead emphasizing how much a model can produce. A million-token output budget would, in principle, let an agent reason, draft, inspect intermediate work, and deliver a long multi-step result in one continuous run without pausing to hand off state.

The dev.to piece cautions against treating the ceiling as a default target. Very long generations increase elapsed time, cost, and the surface area for errors, so production use should involve explicit budgets, streamed partial results, checkpointed agent state, and tests or human review before any code change is applied. The expanded cap is most compelling when a task genuinely benefits from a single long trajectory.

Coding results and an internal Rust rewrite

Google reports 77.9% on DeepSWE v1.1 and describes Argon as state of the art on that long-horizon software engineering evaluation. The company also points to internal work on migrations and optimization, including an Argon-assisted Rust rewrite of 32,000 lines of SIMD code in libgav1. Google claims the resulting safe Rust decoder ran 2.7× faster than the existing Rust port while producing identical video output. The dev.to article stresses that this is a Google-reported internal case study, not an independently reproduced result, and that DeepSWE v1.1 and CWE-bench v1 differ from SWE-bench Verified and CWE-bench v0, so their percentages are not interchangeable with those better-known benchmarks.

The GPT-6.1 Sol comparison remains unsettled

The article frames Argon against OpenAI's GPT-6.1 Sol but declines to declare a winner, because no apples-to-apples public evaluation exists. The two represent different promises: Argon's long execution horizon and low announced introductory rates versus Sol's established position in coding workflows. Until matching model cards, identical benchmark runs, and broad public API availability exist, teams cannot make a confident migration decision on numbers alone.

Capabilities beyond code

Google also claims leadership on the Vals Index, an economic-impact-weighted evaluation spanning finance, coding, legal, and tax, along with 51.3% on AutomationBench and 91.7% on LVBench for long-video understanding. On security, the company cites a 68% tie for first place on CWE-bench v1 and says Argon can find, validate, and patch vulnerabilities. The launch describes safeguards including misuse protections, prompt-injection testing, misalignment monitoring, and hardened evaluation environments.

Pricing and access

Announced introductory pricing is $2 per million input tokens and $10 per million output tokens, with cached input discounted 95%. As dev.to notes, a million-token output ceiling can make output consumption substantial, and end-to-end agent costs also include retries, tool calls, and verification steps. For now the pricing is a signal, not a guarantee that any developer can provision the model today.

Why it matters

For coding and enterprise AI users, Argon shifts competition onto a new axis: output capacity rather than input context. If Google's claims hold up under independent evaluation, agents could plausibly take on larger migrations, multi-file changes, and long debugging chains in a single pass, at token prices that undercut typical frontier-model rates. But every number so far is publisher-reported, the rollout is staged, and the headline comparison with GPT-6.1 Sol has no controlled head-to-head data behind it. The practical move for teams is to wait for broad API access, then test Argon and incumbent models on their own repositories with equivalent prompts, tool access, and budgets, measuring accepted task completion and total token spend rather than trusting benchmark scores alone.