Gemini 4 Argon runs million-token code migrations without chat

Gemini 4 Argon changes software development pipelines by replacing conversational prompting with massive, unattended execution runs. As woshipm reports, Google lifted the maximum output ceiling from 64,000 tokens to 1 million tokens. Context limits match across inputs and outputs. According to ifanr, Google set introductory pricing at $2 per million input tokens and $10 per million output tokens.

A 95% discount on cached input tokens makes long autonomous refactoring economically viable. Unattended batch execution replaces the chat interface.

Benchmark numbers recorded by ifanr illustrate the operational divide. In an API test conducted by AIHubmix, Gemini 4 Argon finished a pixel-art scene task in 3 minutes for $0.50, while GPT-6 Astra took 25 minutes and cost $11. In AI/ML API's 3D scene benchmark, Argon incurred $4.22 compared to $23.82 for GPT-6 Astra. Rubens Soto's 3D animation test showed a parallel gap, with Argon costing $3.83 per generation against $16.36 for Claude Opus 5.5.

Pricing million-token outputs over interactive chat

Conversational prompting fails when a complex task requires continuous code changes across multiple directories. According to woshipm, Gemini 4 Argon expands the single-output ceiling from 64,000 tokens to 1 million tokens, aligning output volume directly with total input capacity. ifanr highlights that this unified 1 million token limit eliminates the need to slice refactoring jobs into iterative chat sessions. Developers no longer babysit turns. Instead, teams can submit a repository specification and let the system emit complete updates in an unattended batch run.

Caching economics accelerate that shift. woshipm notes that cached input tokens receive a 95% discount down to $0.10 per million tokens, beating the 90% discount on Gemini 3.8 Flash and penalizing teams that query models through repetitive chat.

Base API pricing reinforces the transition toward large autonomous traces. woshipm records an introductory 50% discount setting rates at $2 per million input tokens and $10 per million output tokens, before standard pricing rises to $4 and $20. Generation benchmarks illustrate the resulting operational divide. According to ifanr, an AIHubmix test saw Argon complete a pixel-art scene task in 3 minutes for $0.50, compared to 25 minutes and $11 for GPT-6 Astra. Rubens Soto logged a 3D animation run where Argon cost $3.83 against $16.36 for Claude Opus 5.5, while an AI/ML API benchmark measured $4.22 against $23.82 for GPT-6 Astra.

Benchmark sweeps against restricted developer access

Google's published evaluations position Gemini 4 Argon as an undisputed frontrunner. In 18 benchmark tests released by Google, Argon took first place exclusively in 12 evaluations and tied in 1, while GPT-6 Astra secured 3 wins and Claude Opus 5.5 claimed 2, according to woshipm. A wider suite of 19 sub-benchmark tests pitted Argon against GPT-6 Astra, Claude Fable 5.1, and Claude Opus 5.5. Argon led in 14 of those assessments, tying Astra on CWE-bench v1. On the DeepSWE v1.1 software engineering benchmark, Argon posted 77.9%, outperforming Astra's 74.1% and Opus 5.5's 74.2%.

Independent metrics tell a tighter story. Argon scored 53 points on the Artificial Analysis Intelligence Index, matching GPT-6 Astra while exceeding GPT-6.1 Sol by 1 point, according to woshipm. Even with top placement on the Vals Index measuring performance in key economic sectors as reported by ifanr, woshipm author Cang He argues that Claude Opus 5.5 retains an unmatched user experience that benchmark scores do not fully capture.

Official claims of a global arrival run directly into reality. While ifanr notes a global launch following six months of development and testing by over 100,000 Google users, appinn reported Gemini 4 Argon is currently unavailable for use.

That contradiction stems from a gated release schedule. According to woshipm, Google is releasing Gemini 4 Argon in phases. Access is initially restricted to trusted cybersecurity defenders, starting with more than 650 trusted partners in the Fairwind program alongside the US government's pre-release evaluation. Rollouts will only then reach paid API users and Google AI Ultra subscribers. Broader developer access through Google AI Studio, Android Studio, the Gemini API, Search AI mode, and Google Antigravity follows this staggered security pipeline.

Execution latency and step monitoring during large refactors

Unattended migrations across enterprise codebases test execution throughput and syntactic fidelity. According to ifanr, Gemini 4 Argon assisted Google's Chrome team in refactoring 6 million lines of legacy code from a 40 million-line codebase over several weeks. A parallel initiative converted C/C++ systems to Rust. As woshipm reported, this covered re2, libgav1, and more than 800,000 lines in the Fuchsia Zircon kernel. Within libgav1, the model rewrote 32,000 lines of SIMD code into safe, auto-vectorizable Rust, producing a 2.7-fold speed increase over the initial translation.

The model translates structural optimizations into direct execution gains. Ifanr noted that Argon raised the end-to-end inference speed of the open-source Gemma 4 E2B model by 2.85 times without loss of quality. In broader software tasks, woshipm reported that Argon scored 77.9% on the DeepSWE v1.1 benchmark.

Antigravity integration shows the resulting operational stability. Internal adoption increased Google engineers' agent code commits by more than 35% while decreasing code rollbacks by nearly one-third, according to ifanr.

Multi-hour traces fail when models take unverified shortcuts or absorb malicious context. Step monitoring counters this degradation. Ifanr recorded that real-time inference step monitoring dropped Argon's unauthorized shortcut rate in programming tests to 2%. Defensive containment held under pressure. In Gray Swan's indirect prompt injection test over 15 attempts, woshipm documented a 0.7% attack success rate, outperforming Claude Opus 5.5 at 1.0% and GPT-6 Astra at 8.5%. On the AA-Omniscience benchmark, Argon posted a 15% hallucination rate, the lowest among models scoring above 45 on the intelligence index.

Frontier AI Models: Benchmark Performance, Execution Cost, and Output Capacity

Why Antigravity hosts Anthropic models alongside Argon

Developer environments rarely rely on a single execution style. In Google's AI programming tool Antigravity, Gemini 4 Argon shares space with Claude Opus 5.5 Medium and Claude Sonnet 5.5 Medium, according to appinn. While ifanr notes that Argon is accessible through Google AI Studio, Android Studio, the Gemini API, and Google AI Pro or Ultra subscriptions in Search AI mode, Antigravity operates as an integrated workbench. Google also announced six months of free top-tier Antigravity access for 10,000 scientists globally. The environment accommodates distinct ways of writing software.

Different engineering tasks reward different model profiles. Raw automated throughput requires one kind of engine, while human-in-the-loop debugging demands another.

Benchmark evaluations highlight that functional divide. On the AutomationBench-AA benchmark tracked by woshipm, Gemini 4 Argon scored 77.5%, outpacing Claude Sonnet 5.5 at 71.3%. Yet benchmark scores do not fully capture developer utility. As woshipm author Cang He argues, Claude Opus 5.5 still offers an unmatched user experience. Released on September 22, Opus 5.5 targets complex tasks and extended agent workflows where interactive feedback keeps programmers oriented during deep architectural updates.

Claude Sonnet 5.5 occupies a complementary lane. Released on September 28, it targets everyday work, bug fixing, document edits, and spreadsheet operations, according to appinn. Speed is its primary draw. Appinn reports that Sonnet 5.5 delivers a speed increase of more than 30% compared to Sonnet 5. Pairing Argon's unattended batch refactoring with Anthropic's rapid interactive responses prevents developers from forcing every software problem through the same pipeline.

Workload Routing Framework: Gemini 4 Argon vs. Frontier Alternates

• Massive codebase migrations, multi-hour autonomous refactoring, or extended agentic code synthesis Route to Gemini 4 Argon via the Gemini API or Google Antigravity. The model provides a 1 million token maximum output limit (up from 64,000 tokens) and scored 77.9% on the DeepSWE v1.1 software engineering benchmark (outperforming GPT-6 Astra at 74.1% and Claude Opus 5.5 at 74.2%). Google validated this pattern internally by refactoring 6 million lines of legacy Chrome code, migrating over 800,000 lines of the Fuchsia Zircon kernel to Rust, and lowering rollback rates by nearly one-third.
• High-volume generation or complex multi-step rendering under strict API expenditure caps Choose Gemini 4 Argon over GPT-6 Astra or Claude Opus 5.5 to exploit aggressive context and token pricing. During its promotional period, Argon is priced at $2 per million input tokens and $10 per million output tokens ($4 and $20 standard), with cached inputs discounted by 95% to $0.10 per million tokens. In empirical generation tests, Argon finished a 3D animation run at $3.83 (versus $16.36 for Claude Opus 5.5) and a 3D scene task at $4.22 (versus $23.82 for GPT-6 Astra).
• Autonomous tool execution exposed to untrusted web scraping or external third-party prompt injection vectors Deploy Gemini 4 Argon with real-time inference step monitoring. In Gray Swan indirect prompt injection testing across 15 attempts, Argon held attack success to 0.7% (compared to 1.0% for Claude Opus 5.5 and 8.5% for GPT-6 Astra) while reducing unauthorized shortcuts on programming benchmarks to 2%.
• Zero-budget querying, ad-hoc conversational triage, or non-paying enterprise seats Do not architect workflows around Gemini Pro or Argon for unpaid users. Starting October 9, 2026, Google routes unpaid Gemini traffic automatically to Flash-Lite, removing Flash and Pro entirely from the free tier. Maintain lightweight triage pipelines strictly within Flash-Lite's bounds or migrate to paid API/subscription tiers for complex reasoning.
• Everyday interactive developer tasks, rapid bug fixing, and office document manipulation Use Claude Sonnet 5.5 (available inside Google Antigravity). Sonnet 5.5 delivers an inference speed increase of more than 30% compared to Sonnet 5 and is explicitly optimized for everyday operations, bug fixes, and spreadsheet/document edits rather than multi-hour unattended batch generation.

Structuring pipelines around rolling compute limits

Starting October 9, 2026, free Gemini users will only have access to the Flash-Lite model, with Flash and Pro removed from the free tier entirely, according to appinn. As ifanr reports, Google's updated access policy routes requests from unpaid users automatically, assigning most requests to Flash-Lite. This change cleanly divides cheap, interactive triage from sustained code refactoring. Back in May, Gemini shifted to consuming quota based on computational volume rather than simple interaction counters.

This dynamic compute quota refreshes every 5 hours and remains tied to weekly limits, according to ifanr. That rolling window stops engineers from executing unbounded batch jobs without paying for dedicated capacity.

Gemini 4 Argon represents Google DeepMind's first in-house developed model above the Flash tier released in over seven months, according to woshipm. Moving complex codebases through such heavy architectures requires navigating Google's revised permission tiers. Appinn notes that Google AI Plus subscribers receive twice the standard Gemini usage quota, granting access to Flash-Lite and Flash while withholding Pro. The rollout timing for these permission changes differs for Plus subscribers, who will be notified by email.

Heavy refactoring pipelines demand higher service levels to avoid stalling mid-migration. According to appinn, Google AI Pro provides four times the standard quota, while Google AI Ultra delivers 5 times or 20 times the usage quota of Google AI Pro. Subscribers on both the Pro and Ultra tiers retain access to Flash-Lite, Flash, and Pro, alongside access to Deep Think. By budgeting around the 5-hour reset intervals, teams can coordinate automated long-trace runs without exhausting their weekly capacity.

Engineering teams planning unattended refactors can evaluate feasibility against Google's internal runs. As reported by woshipm, autonomous agents converted the Fuchsia Zircon kernel with over 800,000 lines of code and rewrote 32,000 lines of SIMD code in libgav1 into safe Rust, yielding a 2.7-fold speed increase. On Zapier's AutomationBench workflow test, the system led with 51.3% against 42.5% for Claude Opus 5.5.

Access remains tightly staged. Google is routing initial availability through more than 650 trusted partners in its Fairwind cybersecurity program and the US government's pre-release evaluation before opening access to paid API users.

Security validation is the remaining gating factor for unattended pipelines. In Gray Swan's indirect prompt injection test over 15 attempts, Gemini 4 Argon logged an attack success rate of 0.7%, compared to 8.5% for GPT-6 Astra.

For readers outside China

• Availability: Google launched Gemini 4 Argon globally across Google AI Studio, Android Studio, the Gemini API, Search AI mode, and Google Antigravity. However, rollout follows a gated phased schedule: early access is restricted to over 650 trusted security partners in Google's Fairwind program and US government pre-release evaluators before reaching general paid API users and Google AI Ultra subscribers. One source notes the model is currently unavailable for use during initial rollout. Consumer app access is strictly tiered: starting October 9, 2026, unpaid accounts only have access to Flash-Lite (with Flash and Pro removed), Google AI Plus accounts get Flash-Lite and Flash (Pro excluded), and full Pro/Deep Think access requires Google AI Pro or Google AI Ultra.
• Pricing: Introductory API pricing for Gemini 4 Argon is set at $2 per million input tokens and $10 per million output tokens (a 50% discount), with cached input tokens discounted by 95% down to $0.10 per million tokens. Standard post-promotional pricing will adjust to $4 per million input tokens and $20 per million output tokens. Web application access is metered through dynamic compute quotas resetting every 5 hours against weekly caps: Google AI Plus grants 2 times standard quota, Google AI Pro grants 4 times standard quota, and Google AI Ultra grants 5 times or 20 times the quota of Google AI Pro. Subscription pricing amounts for the Plus, Pro, and Ultra tiers are not disclosed in sources. Google also granted 10,000 scientists globally six months of free top-tier Google Antigravity access.
• Closest Western equivalents: OpenAI GPT-6 Astra; OpenAI GPT-6.1 Sol; Anthropic Claude Opus 5.5; Anthropic Claude Sonnet 5.5; Anthropic Claude Fable 5.1
• Data residency: Not disclosed in sources. The source material provides no details regarding data storage regions, jurisdiction-specific hosting, or residency guarantees for Gemini 4 Argon or Google Antigravity processing.

Sources

• ifanr ‌‌⁤‌⁢‬⁣⁣⁢⁢⁡⁡⁢⁡‌‬⁤⁢‌⁡⁤⁡⁣⁡⁤‌⁢⁤⁤‌⁢‬⁣⁡刚刚，Gemini 4 发布！性价比「斩杀」GPT-6 和 Opus 5.5，重回御三家 https://ifanr.com/1684241
• appinn Gemini 免费版大缩水，10月9号起只剩 Flash-Lite https://appinn.com/gemini-free-flash-lite
• woshipm 还得是谷歌，Gemini 4 发布，猛的离谱。。。 https://woshipm.com/ai/6473026.html
• woshipm Gemini 4 Argon 发布 https://woshipm.com/ai/6473057.html