Google launches Gemini 4 Argon globally with introductory API pricing of $2/$10 per million tokens

Google has made Gemini 4 Argon globally available on October 13, 2026, opening access to its latest frontier AI model for developers, enterprises, researchers, and paying consumers. The release introduces temporary API pricing of $2 per million input tokens and $10 per million output tokens, with both rates scheduled to double after November 30. The model combines advanced reasoning, multimodal understanding, long-running agent capabilities, and specialized performance across software engineering, finance, law, cybersecurity, and scientific research.

The announcement marks the transition from limited access introduced on September 30 to a broader commercial rollout. Developers can use the model through the Gemini API and Google AI Studio, enterprises through Google Cloud, and eligible Google AI Pro and Ultra subscribers through the Gemini app and AI Mode in Search.

Argon introduces up to one million tokens of unified context and generation capacity, configurable reasoning levels and continuation mechanisms for extended tasks. Google positions these features around sustained work, including codebase optimization, complex document analysis and multi-step enterprise operations.

Its performance claims are substantial but require careful interpretation. Argon leads some professional-work evaluations, while competing frontier models remain stronger on selected coding and general-intelligence benchmarks. Google's demonstrations in robotics, cybersecurity and research illustrate its breadth without establishing production reliability across every environment.

··········

GEMINI 4 ARGON IS NOW AVAILABLE ACROSS GOOGLE'S CONSUMER, DEVELOPER AND ENTERPRISE PLATFORMS.

The October 13 launch expands access to a model previously available through restricted testing, while preserving different eligibility requirements across Google's products.

Google originally introduced Gemini 4 Argon on September 30, initially emphasizing defensive cybersecurity and selected early-access programs. The broader release now makes the model commercially available through the paid Gemini API. The API identifier is gemini-4-argon, and supported development routes include Google AI Studio, Google Cloud and Google Antigravity.

Consumer access is available through the Gemini app and AI Mode in Search for eligible Google AI Pro and Ultra subscribers, subject to product-specific limits. Access to the model does not guarantee identical controls across these interfaces: API users may configure reasoning and continuation options unavailable in consumer products.

Google also plans to offer 10,000 eligible scientists six months of access to the highest Antigravity usage tier, subject to selection and onboarding capacity.

........

Platform	Availability	Primary use
Gemini API	Paid API tier	Application development and automated workflows
Google AI Studio	Supported developer access	Testing, prototyping and API integration
Google Cloud	Enterprise availability	Managed business applications
Gemini app	Eligible AI Pro and Ultra subscribers	General-purpose reasoning
AI Mode in Search	Eligible subscribers	Search-assisted research
Google Antigravity	Eligible Ultra and enterprise access	Autonomous coding agents

........

··········

INTRODUCTORY API PRICING ENDS NOVEMBER 30, WITH INPUT AND OUTPUT RATES DOUBLING IN DECEMBER.

Google is offering a temporary 50% discount relative to Argon's announced standard public API prices. Until November 30, developers pay $2 per million input tokens and $10 per million output tokens. From December 1, the published rates rise to $4 and $20 respectively.

Cached input receives a 95% discount during the introductory period, bringing its effective price to $0.10 per million cached input tokens under the stated arrangement. Caching is most useful for repeated reference documents or shared instructions, but actual savings depend on eligibility and reuse.

........

Pricing component	Through November 30	From December 1
Input, per 1M tokens	$2.00	$4.00
Output, per 1M tokens	$10.00	$20.00
Cached input, per 1M tokens	$0.10	Applicable post-promotion pricing
Introductory discount	50%	None

........

Consider an application processing 100 million input tokens and generating 20 million output tokens monthly. At introductory prices, input costs $200 and output costs $200, for $400 per month. Under the December prices, identical consumption costs $800. These calculations exclude caching, tool charges and infrastructure.

........

Monthly usage	Introductory cost	Standard cost
10M input + 2M output	$40	$80
100M input + 20M output	$400	$800
500M input + 100M output	$2,000	$4,000
1B input + 200M output	$4,000	$8,000

........

A lower token price does not automatically imply a lower cost per completed task. Difficult workflows may require retries, additional reasoning or human intervention, so organizations should measure reliability and actual token consumption during the promotional period and recalculate budgets at December rates.

··········

THE ONE-MILLION-TOKEN CAPACITY IS DESIGNED FOR LONG-RUNNING REASONING AND AGENT WORKFLOWS.

Argon introduces an unusually large unified input-output token budget, with supported API mechanisms for extended decoding and continuation. It does not mean that every consumer interface generates a million visible words or that ordinary requests routinely consume the entire capacity.

The model supports configurable thinking levels: Low for straightforward tasks, Medium as the default balance of reasoning and cost, and High for more demanding problems. Higher reasoning effort can improve difficult-task performance but does not guarantee better answers in every situation.

For example, a software agent may inspect repository files, identify dependencies, modify code, run tests, diagnose failures and repeat the cycle. Extended reasoning and continuation can help sustain this workflow, but reliability still depends on execution infrastructure, context management, permissions and error recovery. A larger token budget does not independently solve those engineering problems.

··········

ARGON LEADS PROFESSIONAL-WORK BENCHMARKS WHILE COMPETITORS REMAIN STRONGER IN OTHER EVALUATIONS.

Argon's competitive position varies by benchmark and task category. Google highlights a first-place position on the Vals Index, which measures economically relevant professional work. In October 12 results, Argon scored 68.90%, ahead of Claude Sonnet 5.5 at 67.04% and Claude Opus 5.5 at 66.97%.

Argon also reportedly scored 65.40% on Finance Agent v2. Such results are relevant to accounting and financial-analysis workflows, but they do not establish suitability for unsupervised financial decisions. Calculations, source documents, assumptions and applicable rules still require verification.

........

Benchmark	Gemini 4 Argon	Selected competitor
Vals Index	68.90%	Claude Sonnet 5.5: 67.04%
Finance Agent v2	65.40%	First among evaluated models
Harvey Legal Agent	19.6%	GPT-6 Astra: 5.4%
Zapier AutomationBench	51.3%	Claude Opus 5.5: 42.5%
DeepSWE v1.1	77.9%	Claude Opus 5.5: 74.2%
Terminal-Bench 4.0	57.4%	Claude Opus 5.5: 66.4%

........

The independent Artificial Analysis Intelligence Index presents a more balanced picture. As reported on October 13, Argon scores 53, matching GPT-6 Astra and Claude Fable 5.1 but trailing Claude Opus 5.5, Claude Sonnet 5.5 and Meta Muse Wonder Full.

........

Model	Intelligence Index
Claude Opus 5.5	58
Claude Sonnet 5.5	56
Meta Muse Wonder Full	56
Gemini 4 Argon	53
GPT-6 Astra	53
Claude Fable 5.1	53
Meta Muse Wonder	52

........

Some comparisons originate from Google's own evaluations and should not be presented as independently reproduced. Scores also vary with configuration and benchmark methodology. Argon is highly competitive on professional workflows, but does not lead every major frontier-model evaluation.

··········

GOOGLE REPORTS PRODUCTIVITY GAINS FROM LONG-RUNNING SOFTWARE ENGINEERING AGENTS.

Google is presenting Argon as an engineering system capable of investigating repositories, executing changes and evaluating results over extended periods.

In one demonstration, Argon worked for roughly 11 hours on a Python codebase of approximately 200,000 lines. Google reports that its accepted changes reduced in-memory session preparation time by about 40% for long-running sessions and accelerated cold package imports by 21%. Two optimizations were accepted and four rejected, illustrating the continuing importance of testing and human review.

Google also reports more than 100,000 monthly internal Antigravity users. Over a four-week period, engineering teams recorded a 35% increase in agentic code submissions alongside a 30% reduction in post-submission rollbacks. These company-reported operational figures do not establish that unrelated organizations will experience identical improvements.

A separate internal optimization reportedly raised Gemma 4 E2B inference throughput from roughly 37,900 to more than 108,000 tokens per second on an NVIDIA A100. For businesses, useful metrics include accepted changes, regression rates, time to resolution, maintenance burden and total inference and supervision costs. More generated code does not necessarily mean better software.

··········

ROBOTICS, SCIENTIFIC RESEARCH AND CYBERSECURITY EXTEND ARGON BEYOND BUSINESS APPLICATIONS.

Google emphasizes Argon's combination of visual reasoning, physical understanding, code generation and sustained analytical work.

In robotics, Google reports an 87% score on an Embodied Reasoning benchmark. In a controlled demonstration, Argon analyzed video of a human playing cornhole, generated simulation code and contributed to a physical trial with Apptronik's Apollo humanoid robot, which completed seven consecutive throws. This does not establish general-purpose robotic autonomy across unfamiliar environments.

Google also describes applications in quantum computing, fusion, life sciences and enzyme engineering. These projects suggest that frontier models can assist hypothesis generation and computational analysis, but scientific conclusions still require independent validation and experimental evidence.

In cybersecurity, Google reports a 0.7% attack-success rate across 15 adaptive attempts on the Gray Swan Indirect Prompt Injection benchmark. A low result under these test conditions is encouraging but does not establish immunity to novel attacks or tool-based exploitation. Google is expanding its Fairwind Program to provide eligible defensive-security researchers access under controlled conditions.

··········

THE COMMERCIAL COMPARISON WITH OPENAI, ANTHROPIC AND META DEPENDS ON COST PER COMPLETED TASK.

Argon's temporary pricing is competitive with several frontier models, but the December increase materially changes deployment economics.

At $2 input and $10 output per million tokens, Argon matches reported standard API rates for GPT-6.1 Sol and Claude Sonnet 5.5. From December, its $4/$20 pricing moves into the reported price category of Claude Opus 5.5 and Meta Muse Wonder.

........

Model	Input / 1M tokens	Output / 1M tokens
Gemini 4 Argon — introductory	$2	$10
GPT-6.1 Sol	$2	$10
Claude Sonnet 5.5	$2	$10
Gemini 4 Argon — December	$4	$20
Claude Opus 5.5	$4	$20
Meta Muse Wonder	$4	$20
GPT-6 Astra	$10	$50

........

The comparison excludes caching differences, tool charges, infrastructure and service-specific discounts. A model costing $0.04 per attempt with 80% independent success has an expected model cost of approximately $0.05 per successful completion when retried until success. A model costing $0.06 with 95% success costs approximately $0.063 per completion, although it may reduce latency and supervision.

For consequential tasks, the financial cost of errors may outweigh token-price differences. Organizations should therefore compare task completion, accuracy, auditability and recovery requirements, rather than choosing solely on list prices.

··········

ARGON'S LONG-TERM POSITION WILL BE DETERMINED BY RELIABILITY, DEPLOYMENT COST AND REAL-WORLD PERFORMANCE.

The October 13 release makes Gemini 4 Argon commercially accessible with strong professional benchmarks, extended reasoning capacity and aggressive temporary pricing. Its million-token capacity, configurable thinking and integration with Google's development infrastructure offer meaningful options for complex coding, document analysis, research and agent workflows.

The evidence remains mixed: Argon leads the reported Vals Index but trails several competitors on broader intelligence and terminal-coding evaluations. Enterprise outcomes will also depend on permissions, retrieval quality, monitoring, tool reliability and human oversight.

Organizations adopting Argon before November 30 should test workloads using both introductory and December rates. The relevant questions are whether the model completes difficult work more reliably, reduces intervention, produces measurable improvements and justifies inference costs under normal production conditions.

Gemini 4 Argon's global launch strengthens Google's position in frontier AI, but its durable competitive advantage will depend on verified task outcomes and sustainable economics after the promotional period ends.

··········

FOLLOW US FOR MORE.

DATA STUDIOS

datastudios.org