Gemini 4 Argon
In one sentence
The lower-priced model that can also read video, audio, and PDFs.
Best for
A lot of text at a low price, long documents, video, audio, and PDFs.
Visit site (opens in a new tab)
Details Google AI Studio
Gemini 4 vs ChatGPT Astra vs Claude Fable
Three current top models, in plain words. Gemini 4 Argon is from Google, GPT-6 Astra from OpenAI, and Claude Fable 5.1 from Anthropic. This page shows price, speed, what each one is good at, and where you can use it. It does not pick an overall winner.
Words we use
A token is a small piece of text the AI reads or writes. In English that is about three quarters of a word on average. Each vendor counts a little differently. Anthropic says 1 million tokens is about 555,000 words or 2.5 million characters on its current counter. At about 500 words a printed page, that is roughly 1,100 pages. That is an estimate. Google and OpenAI do not publish a word count. The size is a similar order, and that word count is not published.
The context window is how much the AI can keep in view at once. That includes your question, the documents, and its answer. All three are about 1 million tokens. That is roughly a full bookshelf of more than 1,000 pages. In the ChatGPT app it is less. Plus has 256,000 tokens and Pro has 400,000. That is on chatgpt.com/pricing.
Cached input means text the AI already read recently, for example the same long document again. It is billed much less per 1 million tokens. Gemini is $0.10 on the introductory price, Astra is $1, and Fable is $0.25.
In short
GPT-6 Astra
In one sentence
The model for computer use and the longest tool list, at a higher price.
Best for
When the model should use a computer, browse, and call many tools.
Visit site (opens in a new tab)
Details OpenAI API
Claude Fable 5.1
In one sentence
The model for long coding and office work, with cheap repeat reads of the same text.
Best for
Long coding runs and office work when the same text is read again and again.
Visit site (opens in a new tab)
Details Anthropic API
At a glance
| Criterion | Gemini 4 Argon | GPT-6 Astra | Claude Fable 5.1 | Winner |
|---|---|---|---|---|
| Price | Gemini 4 Argon Winner | |||
| Long documents and media | Gemini 4 Argon Winner | |||
| Coding | No clear winner | |||
| Computer use and agents | GPT-6 Astra Winner | |||
| Cloud availability | Claude Fable 5.1 Winner |
How we rated: published prices and counts only, not taste. A price star is 5 times the lowest input price divided by that model's price, rounded, through 2026-11-30. Long documents give one star each for text, image, extra media, output above 128,000 tokens, and a price that does not jump inside the window. Coding stays at 4 because the vendors' tables disagree. Computer use and agents count tool groups plus a comparable OSWorld figure, and a preview loses one star. Cloud counts the named platforms, at most 5.
Winners: price and long documents go to Gemini 4 Argon, computer use and agents to GPT-6 Astra, cloud to Claude Fable 5.1. Astra is close on cloud, with the API, Azure, and Bedrock. Coding has no clear winner. The basis is figures each vendor reported itself. There is no overall winner.
What an everyday text costs
These are API estimates for input only, using about 45,000 tokens for a 50-page report of about 25,000 words, with Anthropic's ratio. That report costs about 9 cents on Gemini, and about 18 cents from 2026-12-01. On Astra and Fable it is about 45 cents. A reply of about one page, roughly 700 tokens, adds about 0.7 cents on Gemini and about 3.5 cents on Astra and Fable. In the app you pay a flat subscription, not per token. None of the three models is on a free plan.
Speed
There is no speed test of all three against each other. Anthropic calls Fable 5.1 slower inside its own lineup. Google warns of occasional slowness or a timeout. OpenAI offers a Fast mode: up to twice the speed at twice the price. In one OpenAI test, Astra finished OSWorld tasks in about 40 minutes, and an older OpenAI model took about 75 minutes. That is not a race against Gemini or Claude.
Free or paid, and where
Gemini 4 Argon
Pro and Ultra, not Free, not AI Plus. Ultra US: $99.99 or $199.99. Pro US price: not published.
GPT-6 Astra
Plus yes. Free and Go no. Pro with Astra, 3 tiers. App context: Plus 256,000, Pro 400,000.
Claude Fable 5.1
Free no. Pro $20 ($17 yearly), extra credits only. Max from $100: up to 50 percent of the week.
Strengths and weaknesses
Gemini 4 Argon
Strengths
- Lowest list price per token, including after 2026-11-30.
- It is the only one that reads video, audio, and PDFs.
- Output can go up to about 1 million tokens with long decode.
Weaknesses
- Computer use is a preview.
- Google only. Not on AWS or Azure.
- Cache is automatic only. Hits are not guaranteed. Google warns of occasional slowness.
GPT-6 Astra
Strengths
- The longest list of hosted tools, including computer, shell, and MCP.
- On Azure and Amazon Bedrock, not only at OpenAI.
- Strong on computer use in its own tests. OSWorld 2.0 offline, partial, 72.6 percent.
Weaknesses
- Gemini's introductory price is about five times lower.
- Above 272,000 input tokens the whole request costs more.
- Input is text and image only. No video, no audio, and no PDF as its own type.
Claude Fable 5.1
Strengths
- One price across the whole 1 million token window.
- Reading again from cache: $0.25 per million tokens.
- Five named platforms, including AWS, Google Cloud, and Microsoft.
Weaknesses
- The same list price as Astra: $10 and $50 per million tokens.
- Input is text and images only. Anthropic calls it slower in its own lineup.
- On the Pro subscription only through extra credits, not inside the normal limits.
Which one for whom
A small budget, a lot of text
Gemini 4 Argon. Through 2026-11-30 a token costs about one fifth of Astra and Fable, and after that about two fifths. A short request with 100,000 tokens in and 10,000 out costs about $0.30, and about $0.60 from December. Astra and Fable cost about $1.50 for that request.
Very long documents, video, audio, PDFs
Gemini 4 Argon. Only this model takes video, audio, and PDF as input, and output can get very long with long decode. On Google's own table it is highest on GraphWalks from 256,000 to 1 million, at 84.2. Fable keeps one price across its 1 million window. Astra gets about twice as expensive above 272,000 input tokens.
Coding
A close race, with no clear winner. Each table puts a different model first. Astra is strong on Terminal-Bench and FrontierSWE in its own tests, and it has Codex. Fable has CursorBench 73.4 on Anthropic's post and is described for long unattended runs, with Claude Code. Argon has DeepSWE 77.9 on Google's table, and Antigravity. The settings differ.
Computer use and agents
GPT-6 Astra, when the model should operate the computer and use many hosted tools. OSWorld 2.0 offline, partial, is 72.6 at OpenAI and on Google's table. Fable fits long agent runs because cache reads cost $0.25. On Argon, computer use is a preview.
Full specifications
The marked cell is the lowest price, the longest output, the most input types, the newest knowledge cutoff, or the most platforms in that row. Where a row has no clear best value, nothing is marked.
| Criterion | Gemini 4 Argon | GPT-6 Astra | Claude Fable 5.1 |
|---|---|---|---|
| Input price, per 1M | $2 through 2026-11-30, then $4. | $10. Above 272,000 input tokens: $20 for the whole request. | $10 across the whole window. |
| Output price, per 1M | $10 through 2026-11-30, then $20. Thinking tokens count. | $50. Above 272,000 input tokens: $75 for the whole request. | $50. US-only inference: 1.1 times. |
| Cached input, per 1M | $0.10 through 2026-11-30, then $0.20. Automatic only. No write price. Hits are not guaranteed. | Reads $1, writes $12.50. Above 272,000: reads $2, writes $25. | Reads $0.25. Writes $12.50 for 5 minutes and $20 for 1 hour. |
| Context window | 1,048,576 tokens for input and output together. | 1,050,000 tokens. At most 922,000 of those are input. | 1 million tokens, one price for the whole window. |
| Max output | 65,536 tokens as the default on synchronous requests. Up to 1,048,576 with long decode. | 128,000 tokens. | 128,000 tokens on the synchronous Messages API. |
| Input and output | Input: text, image, video, audio, PDF. Output: text. | Input: text and image. Output: text. | Input: text and images. Output: text. |
| Knowledge cutoff | March 2026. | April 30, 2026. | June 2026. |
| Plans in the app | Pro and Ultra, not Free, not AI Plus. Ultra US: $99.99 or $199.99. Pro US price: not published. | Plus yes. Free and Go no. Pro with Astra, 3 tiers. App context: Plus 256,000, Pro 400,000. | Free no. Pro $20 ($17 yearly), extra credits only. Max from $100: up to 50 percent of the week. |
| Cloud | Google only. Not AWS, not Azure. | OpenAI API, Azure, Amazon Bedrock. | Claude API, Bedrock, Google Cloud, Microsoft Foundry, Claude Platform on AWS. |
| Reasoning control | Thinking: low, medium, high. | reasoning.effort: low, medium, high, xhigh, max. | Adaptive thinking always on. Default high. High in Claude Code, medium in Cowork and on Claude.ai. Per-message effort is a beta. |
| API limits | No public RPM for this model alone. Spend cap per 10 minutes: tier 1 $10, tier 2 $50, tier 3 $200. Batch queue: 5 million, 500 million, and 1 billion tokens by tier. | Build: 5,000 requests a minute and 1 million tokens a minute. Launch: 10,000 and 4 million. Grow: 15,000 and 40 million. | Not published. |
| Architecture | Published: sparse MoE transformer. | Not published. | Not published. |
| Retirement | Not published. | Not published. | Not sooner than 2027-09-01. |
Tools
These are API tools, not buttons in the app.
Gemini 4 Argon
- Code execution: yes.
- Computer use: yes, in preview.
- File search, function calling, structured outputs, URL context: yes.
- Google Search and Google Maps grounding: 5,000 free requests a month shared across Gemini 3 and newer, then $14 per 1,000.
- Agentic video understanding: yes.
- Not included: Live API, image generation, audio generation, explicit cache. Implicit cache only.
- MCP is not listed on the model page. In Google demos, Blender MCP runs through Antigravity.
- Managed agents in the Gemini API and Antigravity Agent: compute is not billed during the preview.
GPT-6 Astra
- On the Responses API: web search, file search, image generation, code interpreter, hosted shell, apply patch, skills, computer use, MCP, tool search.
- Endpoints: Responses, Chat Completions, Batch. Not included: Realtime, Assistants, fine-tuning.
- Codex: notes across context windows. Earlier windows are searchable. That part is experimental.
Claude Fable 5.1
- Web search, web fetch, code execution, computer use, browser use, bash, text editor, memory tool, tool search, MCP connector, advisor tool.
- Code execution: $0.05 per container hour. 50 free hours a day per organization.
- Managed Agents: $0.08 per session hour.
- Forced tool use returns an error on Fable 5.1.
Vendor benchmarks
Each vendor ran its own tests with its own settings. Numbers from different tables are not directly comparable. Example: Terminal-Bench 4.0 for Astra is 58.2 on Google's table and 57.9 on OpenAI's table. OSWorld for Fable: Anthropic reports 77.9 partial on a different task release. Google and OpenAI do not publish a number for that.
Google's table
| Criterion | Gemini 4 Argon | GPT-6 Astra | Claude Fable 5.1 |
|---|---|---|---|
| Vals Index | 68.9% | 63.1% | 65.8% |
| AutomationBench | 51.3% | 41.4% | 31.4% |
| Vals Finance Agent v2 | 65.4% | 53.5% | 58.9% |
| Harvey Legal Agent | 19.6% | 5.4% | 6.7% |
| DeepSWE v1.1 | 77.9% | 74.1% | 67.4% |
| FrontierSWE v2 | 55.0% | 65.5% | 56.3% |
| Terminal-bench 4.0 | 57.4% | 58.2% | 57.9% |
| Terminal-Bench Science 0.1 | 57.6% | 68.1% | 52.6% |
| GraphWalks 256k-1M BFS F1 | 84.2% | 71.8% | 65.0% |
| OSWorld-2.0 offline partial | 69.2% | 72.6% | Not published. |
| LVBench | 91.7% | 87.5% | 79.7% |
| CWE-bench v1 | 68.0% | 68.0% | 58.0% |
Google's model card, October 2026. Google also ran the other two models for this table. Blank cells were not reported on the card.
https://deepmind.google/models/model-cards/gemini-4-argon/
OpenAI's table
| Criterion | Gemini 4 Argon | GPT-6 Astra | Claude Fable 5.1 |
|---|---|---|---|
| Terminal-Bench 4.0 | Not published. | 57.9% | 55.8% |
| DeepSWE v1.1 | Not published. | 74.1% | 67.4% |
| GPQA Diamond | Not published. | 96.0% | 93.7% |
| HLE with tools | Not published. | 57.2% | 63.8% |
| ARC-AGI-2 | Not published. | 95.0% | 90.0% |
| ARC-AGI-3 | Not published. | 99.9% | Not published. |
| FrontierMath Tier 4 | Not published. | 97.6% | 87.8% |
| OSWorld 2.0 offline partial | Not published. | 72.6% | Not published. |
| BrowseComp | Not published. | 91.5% | 87.4% |
| MRCR v2 8-needle 512K-1M | Not published. | 96.3% | Not published. |
OpenAI's launch post, OpenAI's own run, effort max. The table columns are Astra, then GPT-5.6 Sol, then Fable 5.1. This table shows the first and the third figure. On HLE with tools the third figure is 63.8 percent. The second figure in that row is 65.0 percent and belongs to GPT-5.6 Sol. Anthropic itself reports 65.0 percent for Fable 5.1 with tools. The settings differ. Some Fable scores on the OpenAI page come from Mythos, footnote 17, named for ScreenSpot-Pro and ExploitGym, not for these rows. Blank cells: not given in this pairing.
https://openai.com/index/gpt-6-astra/
Anthropic's table
| Criterion | Gemini 4 Argon | GPT-6 Astra | Claude Fable 5.1 |
|---|---|---|---|
| Terminal-Bench 4.0 | Not published. | Not published. | 55.8% |
| Terminal-Bench 4.0 Mythos 5.1 | Not published. | Not published. | 60.9% |
| Terminal-Bench-Science 0.1 | Not published. | Not published. | 52.6% |
| OSWorld 2.0 partial | Not published. | Not published. | 77.9% |
| OSWorld 2.0 strict | Not published. | Not published. | 41.7% |
| HLE no tools | Not published. | Not published. | 60.9% |
| HLE with tools | Not published. | Not published. | 65.0% |
| AutomationBench | Not published. | Not published. | 31.4% |
| CursorBench 3.2.0 | Not published. | Not published. | 73.4% |
| GDPval-AA v2 | Not published. | Not published. | 1853 |
Anthropic's launch post, with production safeguards on. Terminal-Bench 4.0 also lists Mythos 5.1 at 60.9. OSWorld 2.0 is the August 2026 task release. Anthropic says these numbers are not directly comparable with earlier published OSWorld 2.0 results, so no competitor is shown there. GDPval-AA v2 is an Elo-like score, not a percent.
https://www.anthropic.com/claude-fable-and-mythos-5-1
Three cost examples
| Criterion | Gemini 4 Argon | GPT-6 Astra | Claude Fable 5.1 |
|---|---|---|---|
| Short chat or code: 100,000 in, 10,000 out | $0.30. From December $0.60. Math: 0.1×2 + 0.01×10. | $1.50. Math: 0.1×10 + 0.01×50. | $1.50. Math: 0.1×10 + 0.01×50. |
| One document: 500,000 in, 10,000 out, one request | $1.10. From December $2.20. Math: 0.5×2 + 0.01×10. | $10.75. Above 272,000 the surcharge covers the whole request. Math: 0.5×20 + 0.01×75. | $5.50. Math: 0.5×10 + 0.01×50. |
| Cache-heavy agent: 900,000 cache reads, 100,000 new, 50,000 out. Every request under 272,000. | $0.79. From December $1.58. Math: 0.9×0.10 + 0.1×2 + 0.05×10. The automatic cache is not a sure hit. | $4.65. The 100,000 new tokens are counted as a cache write. Math: 0.9×1 + 0.1×12.50 + 0.05×50. | $3.98 (the math comes to 3.975). The 100,000 new tokens are counted as a 5-minute write. Math: 0.9×0.25 + 0.1×12.50 + 0.05×50. |
Standard API list prices in US dollars, with no tool fees and no tax. Batch and Flex are half price on all three (Gemini Batch and Flex, OpenAI Batch and Flex, Anthropic Batch). Gemini Priority on the introductory price: $3.60 input and $18 output. OpenAI Fast: double. Anthropic US-only: 1.1 times. Gemini's cache is automatic, and hits are not guaranteed. These are estimates from list prices. A real bill depends on cache hits and thinking tokens. On Gemini the output price includes thinking tokens.
Consumer plans
Gemini 4 Argon
Gemini app: Google AI Pro and Ultra, not Free and not AI Plus. Ultra in the US: $99.99 a month for 5 times Pro, or $199.99 for 20 times Pro. The fetched page does not show a US price for Pro. Limits follow compute, every 5 hours up to a weekly limit, and AI credits can extend them. Antigravity with Argon is on Ultra. The API: not on the free tier and not through Google AI Pro in AI Studio.
GPT-6 Astra
ChatGPT pricing table: Astra is yes on Plus, and no on Free and Go. Plus also includes Astra in ChatGPT Work and Codex. Pro: Pro reasoning powered by GPT-6 Astra, three usage tiers. Business and Enterprise: yes. On Enterprise an admin turns the model on, and it is off by default. Context in the app, for reasoning: Plus 256,000, Pro 400,000. That is smaller than the API window.
Claude Fable 5.1
Claude: Free is no. Pro is $20 a month, or $17 if billed yearly. Fable there is only through extra usage credits, not inside the normal limits. Max from $100 a month, 5 times or 20 times: Fable up to 50 percent of the weekly limits. Team, premium seats: up to 50 percent of the weekly limits. Enterprise: yes, $20 per seat plus usage at API rates.
Cloud platforms
Gemini 4 Argon
Gemini API, AI Studio, Gemini for business, and the Gemini Enterprise Agent Platform. Prices there can differ. Not on AWS or Azure. The checked pages do not list it there, so: Google only.
GPT-6 Astra
OpenAI API, Microsoft Azure, and Amazon Bedrock. Plus the ChatGPT app.
Claude Fable 5.1
Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry, and Claude Platform on AWS. That is five named platforms.
Safety, privacy, and enterprise
Gemini 4 Argon
Paid API: not used to improve the products. The free tier is, but Argon is not on it. Frontier Safety Framework: reaching the biological CBRN CCL and the cyber CCL cannot be ruled out, and safeguards are applied. The Fairwind Program gives cyber defenders access without the cyber guardrails. Gray Swan, IPI attack success: 0.7 percent. Enterprise: VPC-SC, CMEK, and data residency boundaries (Gemini Enterprise).
GPT-6 Astra
Astra meets the Critical cyber threshold under the Preparedness Framework. It refuses advanced cyber tasks such as proof-of-concept exploits. Daybreak is meant to expand that. In the API a flagged task stops. Zero Data Retention for eligible API customers. ChatGPT Enterprise: data residency in the US, EU, UK, JP, CA, KR, SG, IN, AU, and UAE. SOC 2 Type 2. EKM.
Claude Fable 5.1
Mythos 5.1 is the same model with looser safeguards, only for verified programs. Fable 5.1 can find vulnerabilities but does not write exploits. Some dual-use cyber and life-science tasks are sent to Opus models. Zero Data Retention for eligible customers until Enterprise Frontier Safeguards (EFS, storage in the customer's cloud) rolls out from fall 2026. An invisible watermark on outputs (EU AI Act code). Team and Enterprise: no training by default. Enterprise: custom retention, HIPAA-ready, SCIM, audit logs. US-only inference: 1.1 times.
FAQ
Which model is best?
None of them for everything. On price and on video, audio, and PDFs, Gemini 4 Argon is ahead. On computer use and tools, GPT-6 Astra is ahead. On cloud platforms, Claude Fable 5.1 is ahead. On coding the tables disagree. There is no overall winner.
What do the three cost on the API?
Per 1 million tokens, standard: Gemini is $2 in and $10 out through 2026-11-30, then $4 and $20. Astra and Fable are $10 and $50. Above 272,000 input tokens Astra costs more for the whole request: $20 and $75. A report of about 50 pages is roughly 9 cents on Gemini and 45 cents on the others. The apps bill through the subscription.
Is one of them free?
No. Gemini 4 Argon is not on Free and not on AI Plus. GPT-6 Astra is not on Free and not on Go. Claude Fable 5.1 is not on Free. The apps have paid plans. The API is paid per token.
What is a token?
A small piece of text the AI reads or writes. In English, about three quarters of a word. Anthropic puts 1 million tokens at about 555,000 words. A 50-page report is then roughly 45,000 tokens.
Which one is faster?
That is not published. There is no shared speed test. Anthropic calls Fable slower in its own lineup. Google warns of occasional slowness. OpenAI sells a Fast mode at up to twice the speed for twice the price.
Which one for a long PDF?
Gemini 4 Argon. It takes PDF, video, and audio as input. Astra and Fable list text and image as input. For very long text, Fable keeps one price across 1 million tokens. Astra gets more expensive above 272,000 input tokens.
Official sources
Gemini 4 Argon
- https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-argon/
- https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/
- https://ai.google.dev/gemini-api/docs/pricing
- https://deepmind.google/models/model-cards/gemini-4-argon/
- https://ai.google.dev/gemini-api/docs/models/gemini-4-argon
- https://ai.google.dev/gemini-api/docs/rate-limits
- https://gemini.google/subscriptions/
GPT-6 Astra
- https://openai.com/index/gpt-6-astra/
- https://developers.openai.com/api/docs/models/gpt-6-astra
- https://developers.openai.com/api/docs/pricing
- https://help.openai.com/en/articles/20001354-gpt-6-and-other-models-in-chatgpt
- https://chatgpt.com/pricing
Claude Fable 5.1
- https://www.anthropic.com/claude-fable-and-mythos-5-1
- https://www.anthropic.com/claude/fable
- https://platform.claude.com/docs/en/models/fable-5-1/overview
- https://platform.claude.com/docs/en/agents-and-tools/tool-use/overview
- https://www.anthropic.com/pricing