Google made Gemini 4 Argon available to developers, businesses and paying Gemini subscribers on 13 October 2026. The announcement, by Koray Kavukcuoglu, SVP of Google DeepMind, followed a limited release to security teams on 30 September.
The short answer: Argon is one of the strongest models available for business knowledge work and very long documents, and it is cheap per token during its introductory period. Independent testing places it level with OpenAI’s GPT-6 Astra rather than ahead of the whole field. It trails on terminal-based coding, and it writes more tokens per task than its rivals, which reduces the price advantage.
If you tested Google’s earlier models, our article on Gemini 3.5 and action-first AI covers what to check before handing more work to an AI agent. The same checks apply to Argon.
What Google released
Argon is Google’s new top-tier reasoning model. It accepts text and image input, produces text, and has a 1 million token context window, according to Artificial Analysis.
The main technical change is the output limit. A single response can now run to 1 million tokens, up from 64,000 on earlier Gemini models. Google says this lets the model work through long, multi-step tasks in one run.
It is available through:
- The Gemini API in AI Studio and Google’s enterprise Gemini platform (paid tier)
- Google Antigravity, Google’s agentic coding tool, for Google AI Ultra subscribers and businesses
- The Gemini app and AI Mode in Search, for Google AI Pro and Ultra subscribers
- The Fairwind Program, which gives approved security teams a version without cyber guardrails
| Pricing per 1M tokens | Introductory (until 30 Nov 2026) | Standard (from 1 Dec 2026) |
|---|---|---|
| Input | $2 | $4 |
| Cached input | $0.10 | $0.20 |
| Output | $10 | $20 |
Cached input is priced at 95% off the input rate, which helps when the same instructions or documents are sent with every request.
Google also showed Argon working on a physical task. In a robotics test described in the launch post, Argon watched a video of a person playing cornhole, mapped the scene and wrote the simulation code for the throwing movement. Google says that when the result was transferred to Apptronik’s Apollo humanoid robot, it landed seven shots in a row. Google DeepMind published this demonstration video on its YouTube channel:
Where the evidence supports Google’s claims
Argon’s clearest lead is in business knowledge work, and most of those results were scored by outside organisations rather than Google. Google’s comparison table sets Argon against GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5. The rows below are the ones scored by third parties, plus Google’s long-document test.
| Benchmark (who runs it) | What it tests | Argon | Next best |
|---|---|---|---|
| Vals Index (Vals AI) | Finance, coding, legal and tax tasks, weighted by share of US GDP | 68.9% | 67.0% (Opus 5.5) |
| AutomationBench (Zapier) | Business tasks completed end to end | 51.3% | 42.5% (Opus 5.5) |
| Vals Finance Agent v2 (Vals AI) | Multi-step financial research | 65.4% | 58.9% (Fable 5.1) |
| Harvey Legal Agent Benchmark (Vals AI) | Legal research and drafting | 19.6% | 6.7% (Fable 5.1) |
| GraphWalks 256K–1M (Google) | Finding information across very long inputs | 84.2% | 71.8% (Astra) |
Scores as published by Google on 30 September 2026, compiled by Emergent.
The Vals Index lead holds on Vals AI’s own leaderboard, updated 29 September, where Argon scored 68.90% against 67.04% for Claude Sonnet 5.5, per OrcaRouter. The margin is about two points, so the lead is real but small.
The legal result needs context. Argon nearly triples the next model’s score, but 19.6% is still a low absolute figure. None of these models can handle legal work without supervision.
Two results matter for anyone building tools that face customers or read outside content:
- Fewer made-up answers. On Artificial Analysis’s Omniscience test, Argon gave a wrong answer instead of declining 15% of the time, against 51% for GPT-6 Astra. On the same test it also answered correctly less often than Astra, so the low rate means it declines more often when unsure.
- Resistance to prompt injection. On Gray Swan’s indirect prompt injection benchmark, attacks succeeded 0.7% of the time across 15 attempts. Claude Opus 5.5 and Fable 5.1 scored 1.0%, and GPT-6 Astra 8.5%. This test covers hidden instructions in web pages, emails or files that an agent reads.
Where it falls short
Argon is not the leader across every domain, and Google’s own table shows this.
Coding results are split. Google’s headline coding figure is 77.9% on DeepSWE v1.1, a test of long software engineering tasks. Google ran Argon’s score in its own setup and took rival scores from leaderboards and system cards. On two other coding tests in the same table, Argon finished last of four:
| Benchmark | What it tests | Argon | Leader |
|---|---|---|---|
| FrontierSWE v2 (Proximal) | Hard software engineering problems | 55.0% | 65.5% (GPT-6 Astra) |
| Terminal-Bench 4.0 | Working through tasks in a command line | 57.4% | 66.4% (Claude Opus 5.5) |
Artificial Analysis ran Terminal-Bench 4.0 independently and got a similar result for Argon, 57%. Argon also trails GPT-6 Astra on OSWorld 2.0, a test of controlling desktop software.
Independent rankings put it level with GPT-6 Astra. The Artificial Analysis Intelligence Index (v4.3.2) combines a set of reasoning, knowledge and coding tests run by Artificial Analysis itself. Argon scored 53, the same as GPT-6 Astra and Claude Fable 5.1, and five points below Claude Opus 5.5 at 58.
It uses a lot of tokens. On that index, Artificial Analysis recorded about 62,000 output tokens per task for Argon against about 27,000 for GPT-6 Astra, according to eesel AI’s pricing analysis. Output tokens are the expensive part of the bill, so a lower price per token does not guarantee a lower price per job.
We covered a similar trade-off in Claude Sonnet 5 effort levels, which explains how to test cost per task before raising an agent budget.
| Cost per Intelligence Index task | Argon | GPT-6 Astra | Claude Opus 5.5 |
|---|---|---|---|
| At introductory pricing | $1.99 | $3.26 | $5.98 |
| At standard pricing (from 1 Dec 2026) | about $3.98 | $3.26 | $5.98 |
Artificial Analysis also measured Argon at 48.9 output tokens per second, against 92.2 for Claude Opus 5.5. Combined with the high token count, long tasks will take noticeably longer to complete.
Some headline results are not like-for-like. Google’s evaluation notes, as summarised by Emergent, show that on LVBench, its long-video test, Argon received one frame per second of video while rival models received a fixed 300 to 800 frames per video. Several case studies in the announcement, such as refactoring part of Chrome’s codebase, come from Google’s own teams and have not been reproduced outside Google.
What you can use it for
Argon is best suited to work that involves large amounts of reading and several steps of analysis. Based on where it scores well, these are the uses worth testing:
| Use | Why Argon fits | Supporting result |
|---|---|---|
| Reviewing long documents such as contracts, policies, tender packs or archives | Holds accuracy across very long inputs | 84.2% on GraphWalks 256K–1M; 1M token context |
| Financial and reporting analysis from exports or spreadsheets | Leads on multi-step financial research | 65.4% on Vals Finance Agent v2 |
| Multi-step business workflows, such as processing enquiries or updating records across tools | Leads on end-to-end business tasks | 51.3% on Zapier’s AutomationBench; 77.5% on Artificial Analysis’s version |
| Assistants that answer from company policies or product information | Less likely to give a confident wrong answer | 15% hallucination rate on AA-Omniscience |
| Agents that read web pages, emails or uploaded files | Lowest prompt injection success rate in Google’s comparison | 0.7% on Gray Swan IPI |
| Turning a written brief into a working prototype or internal tool | Top score, though the top three models are within 1.6 points | 91.9% on Vibe Code Bench |
| Large code refactors and migrations | Long output limit suits big changes in one run | 77.9% on DeepSWE (Google-run) |
| Analysing long video, such as recorded calls or training footage | Strong long-video results, with the frame-rate caveat above | 91.7% on LVBench (Google-run) |
For most people, the easiest way to try it is the Gemini app on a Google AI Pro or Ultra plan. Developers and agencies can test it through AI Studio at the introductory API rate until 30 November 2026.
Other models are likely to be a better choice for:
- Coding agents that work mainly in a terminal, where Claude Opus 5.5 leads Terminal-Bench 4.0 by nine points
- Tasks where response speed matters, given Argon’s lower output speed
- High-volume, simple tasks such as tagging or short summaries, where a smaller and cheaper model will do the job at a fraction of the cost
Verdict
Argon meets Google’s claims for business knowledge work, long documents and resistance to prompt injection, and most of that evidence comes from third parties. It does not meet the claim of outperforming other frontier models across domains. Independent indexes rank it level with GPT-6 Astra and behind Claude Opus 5.5, and it trails on terminal-based coding.
The introductory pricing makes it worth testing before 30 November 2026. Test it on your own documents and workflows, and measure the cost per finished task, since Argon’s high token use can cancel out its lower price per token.
Sources
- Gemini 4 Argon: Built to solve your hardest problems, Google, 13 October 2026
- Gemini 4 Argon: our next era of frontier intelligence, Google, 30 September 2026
- Gemini 4 Argon benchmarks: every score, its source, and what it means, Emergent, 1 October 2026
- Gemini 4 Argon (High) intelligence, performance and price analysis, Artificial Analysis
- Gemini 4 Argon pricing, eesel AI
- Gemini 4 Argon vs Claude Opus 5.5, OrcaRouter
- Vals Index, Vals AI
Author
