Gemini 4 Argon Launch Puts Google Back in the Frontier Model Race

A two-step launch for Google's new flagship

Google's first Gemini 4 model arrived in two stages. On September 30, 2026, the company announced Gemini 4 Argon. It said the model delivers frontier performance on real-world software engineering, knowledge work and cybersecurity defense, and that it can produce up to 1 million output tokens, a limit Google calls industry-leading.4 At first only a select group of cyber partners could use it, through Google's Fairwind Program. Google said the model was trained specifically for defensive security work and can find, validate and patch critical software vulnerabilities on its own.5 Thirteen days later, on October 13, Google made Argon available worldwide through the API in AI Studio and the Gemini platform. Google AI Ultra subscribers got it in Antigravity, and Ultra and Pro subscribers got it in the Gemini app and in AI Mode in Search.2 The model ID is gemini-4-argon, and the Google Cloud documentation lists it as generally available, with retirement no earlier than October 2027.18

TechCrunch's launch-day headline called Argon Google's most powerful model yet.5 That claim lasted about two weeks. On October 28, Google announced a second Gemini 4 model, Carbon, which outscores Argon on the Artificial Analysis Intelligence Index v5.0, 59 to 50. Carbon is limited to Fairwind cyber defenders for now and has no public release date.3

The pricing is the bigger story

The benchmark claims got the headlines, but the price may matter more. Argon launched at an introductory $2 per million input tokens and $10 per million output tokens, with cached input 95% off. Through November 30, 2026 that means $0.10 per million cached tokens. On December 1, standard rates of $4 and $20 take effect.115 The Verge put the introductory price at roughly a fifth of what OpenAI's GPT-6 Astra and Anthropic's Fable 5 models cost.6 VentureBeat reached the same conclusion from OpenAI's published Astra rate of $10 input and $50 output. It also pointed out that Argon's standard price is the same as Claude Opus 5.5's base API price.15

The reporting lines up on the price numbers but differs on what they mean. The Verge warned that cheaper tokens don't automatically mean cheaper work, because models use different numbers of tokens to finish the same task.6 Third-party testing supports that warning. Vals measured Argon's cost on its index below that of the Claude models just behind it. But the cost rose sharply on long agentic jobs, reaching $193.78 per test on CUA-bench.14 One comparison against Gemini 3.1 Pro found that Argon already costs more per finished task than its predecessor, even before the introductory discount ends.16 In other words, Argon is cheap per token, not necessarily per job. Teams running long autonomous coding sessions should measure their own costs rather than rely on the rate card.

Benchmarks: a wide lead, not a clean sweep

Google claims that Argon beats GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5 across coding, cybersecurity, science, finance and law.6 Its own comparison table is more mixed than that summary suggests. One breakdown of the 19 published rows found that Argon led 13 outright and tied one. It lost on FrontierSWE v2, where it scored 55.0% to GPT-6 Astra's 65.5%, and on Terminal-bench 4.0, where Opus 5.5 led 66.4% to 57.4%.19 VentureBeat described the same picture: Argon tops or ties more categories than any rival, but doesn't win them all.15

Argon's clearest wins are in enterprise-style work. On Harvey's Legal Agent Benchmark it scored 19.6%, against 5.4% for GPT-6 Astra and 3.8% for Opus 5.5.15 It also posted large leads in long-context reasoning between 256,000 and 1 million tokens, and on the Vals Finance Agent benchmark.19 When Vals first ranked Argon, it put the model first on its index at 68.90%, ahead of Claude Sonnet 5.5, Opus 5.5 and Fable 5.1.14 By late October the same page showed Argon fourth of 30, at 69.24%.14 That drop came from newer entrants, not a change in Argon's score.

Independent composite scores are less flattering. One leaderboard summary put Argon at about 53 on the Artificial Analysis index, roughly level with Fable 5.1 and slightly behind Opus 5.5, Sonnet 5.5 and Meta's Muse Wonder.18 A mid-October snapshot ranked Argon sixth at 52.6, behind all three leading Claude models, GPT-6 Astra and Moonshot's open-weight Kimi K3.1 Pro.21 Google's later Carbon materials list Argon at 50 on the updated v5.0 index, compared with 57 for Astra and 58 for Opus 5.5.20 The different numbers come from different index versions, but they all show the same thing. On composite measures, Argon is a frontier contender, not the clear leader. Its strongest case is in specific enterprise and long-context tasks.

The specs also vary slightly depending on where you look. Google describes a 1 million-token limit shared between input and output through the Gemini API.2 Vals lists a 1 million-token context window and a 262,000-token maximum output under its test settings.14 Another comparison says the full 1 million-token output requires Google's long-decode mode.16

Why the staged rollout matters

Argon's release followed a pattern that has become standard for frontier models in 2026. Labs now restrict early access to vetted security users before a public launch.6 Google said that releasing capabilities at this level safely requires a phased approach. It also said it is taking part in the U.S. government's voluntary program for pre-release model access.7 Alongside the public model, Google is giving a small set of trusted defenders a version of Argon without cyber guardrails.1 Cloud documentation lists it as a separate model ID, gemini-4-argon-cyber-permissive.8 Carbon is following the same path, starting with cyber defenders and expanding to trusted partners, with no date set for general access.17

The Verge framed the launch as Google's first major frontier release since Gemini 3 in late 2025. It said that in the meantime, strong releases from Anthropic and OpenAI had left Google well behind, and internal upheaval had raised doubts about whether it could catch up.6 That context explains the aggressive pricing. A lab that has fallen behind needs developers to try its model, and an introductory discount is the quickest way to get them to.

A crowded autumn for model releases

Argon arrived during an unusually busy stretch. Anthropic released its Claude 5.5 family in about a month: Opus 5.5 on September 22, Sonnet 5.5 on September 28, and Haiku 5.5 on October 7. Haiku 5.5 costs $0.10 input and $0.50 output for prompts up to 100,000 tokens.25 Sonnet 5.5 now powers Claude's free tier, and one tracker measured it ahead of Opus 5.5 on Terminal-Bench 4.0.21 Anthropic later previewed Claude Duet, an agent app that runs in the background and asks for approval before important actions.24

OpenAI launched GPT-6 Astra on September 3 and made GPT-6 the default in ChatGPT on October 7.22 It announced GPT-6.1 Sol at DevDay, but no GPT-6.1 Astra followed. According to a Wall Street Journal report cited by one tracker, that release was cancelled after internal safety tests.21 OpenAI then added GPT-6.1 Terra at $1 input and $5 output. Release dates for Terra differ: one source says October 15, while others say October 26.232228

The open-weight side may matter even more for Google's pricing. Moonshot's Kimi K3.1 Pro scored 55 on the Intelligence Index, the highest reading yet for an open-weight model. It costs the same $2/$10 as Argon's introductory price, with full weights due November 4.21 Meta now runs two tracks: closed Muse Wonder models at $4/$20, and an open-weight model, Muse Spark Open.2124 Mistral previewed Mistral Large 4, a model with roughly one trillion parameters, and has promised open weights after safety testing.30 One roundup argued that the big labs spent October making existing model families cheaper and easier to access rather than setting new capability records.25

The takeaway

Argon shows Google catching up rather than taking the lead. It is a real frontier model, with clear wins in legal, finance and long-context work and broad availability on Google's own platforms. But independent composite scores put it level with or slightly behind Anthropic's best, and its token-price advantage partly disappears on long agentic tasks. Within a month, Google released a stronger model and sharply undercut its rivals on price, after almost a year without a new flagship. Whether that changes the race depends on two things: how quickly Carbon becomes available to the public, and whether developers stay once Argon's price doubles in December.