Gemini 4 Argon undercuts rival flagships fivefold, but developers can't buy it yet

Google, OpenAI and Anthropic each shipped a top-tier model within about a month of the others, and a pricing comparison published on dev.to highlights why developers are paying attention: Gemini 4 Argon entered at $2 per million input tokens, one fifth of the $10 rate charged by OpenAI's GPT-6 Astra and Anthropic's Claude Fable 5.1. The catch is access. The cheapest flagship of the three is not yet on general sale.

A fivefold gap, with an asterisk

According to the dev.to roundup, whose figures come from each vendor's own announcements, Gemini 4 Argon (launched by Google on September 30, 2026) carries an introductory rate of $2 per million input tokens and $10 per million output tokens, rising later to $4 and $20. GPT-6 Astra (September 3, 2026) and Claude Fable 5.1 both sit at $10 and $50.

The roundup models a modest production workload of 10 million input and 2 million output tokens per month. At list prices that costs $40 on Argon's introductory rate, $80 once the price rises, and $200 on either Astra or Fable 5.1.

Two qualifications matter. The $2 tier is not exclusive to Google: GPT-6.1 Sol and Claude Sonnet 5.5 are also priced at $2 and $10, and both can be called today. And once Argon's introductory window closes, it costs the same as Claude Opus 5.5 at $4 and $20.

Access, not price, decides today

Access to Argon is currently limited to trusted cyber defenders through a programme the roundup names Fairwind. Google says a paid API and the Google AI Ultra tier are next, but gives no date. By contrast, Astra is available now through ChatGPT paid plans and the API, Sol is reachable via the API, and all three Anthropic models are generally available.

That leads to the roundup's practical guidance. Teams that need to ship this week should compare Sol and Sonnet 5.5 at the $2 tier, or Opus 5.5 at $4 if they need more capability. Those who want the top tier immediately can choose Astra or Fable 5.1 at $10 and $50. Anyone planning around Argon should budget at the $4 and $20 price rather than the introductory one.

What each vendor claims

Google positions Argon for long, complex work such as coding, legal and finance analysis, and cybersecurity defence, with output expanded to 1 million tokens, up from 64K. Its self-reported results include a record 77.9 percent on DeepSWE v1.1, first place on AutomationBench at 51.3 percent, 91.7 percent on LVBench and a tied-first 68 percent on CWE-bench.

OpenAI markets Astra as a model that can carry out computer-based tasks on a user's behalf, and claims state-of-the-art results on FrontierMath Tier 4, ARC-AGI 3 and TerminalBench 4.0. It also says the much cheaper Sol matches Astra on DeepSWE.

Anthropic describes Fable 5.1 as its most capable generally available model, built for long-running work, and says Opus 5.5 performs at Fable's level on most tasks while costing 40 percent less to run than Opus 5 and running more than 30 percent faster.

The benchmarks do not settle it

Every score cited is the vendor's own result on its own setup, the roundup stresses. Different harnesses, prompts and tool access can shift a score significantly, so a Google number and an OpenAI number cannot be ranked against each other. The author's recommendation is to treat these figures as claims about where each model is strong, and to run a real evaluation: 20 to 50 tasks drawn from your actual product, scored identically across every candidate model.

Why it matters

A fivefold price gap at the top of the model market would normally redraw developer budgets overnight, but this one is partly illusory. The durable $2 tier already exists at OpenAI and Anthropic, Argon's own price doubles after its introductory window, and the model remains unavailable to most buyers for now. For engineering teams, the deciding variables are availability and the post-introductory price, not launch headlines. Until independent benchmarks appear, vendor-reported scores should carry little weight in the decision.