Online AI compared: Claude, GPT and Gemini (October 2026)

Online AI models change every few weeks, and so do their prices. This comparison collects what each provider's official pages say as of 26 October 2026. We include no benchmarks of our own and no third-party figures: when a fact cannot be confirmed from the official source, we leave it out.

A note on names

"Codex" is OpenAI's agentic coding tool, not a versioned model. The model OpenAI introduced at DevDay on 29 September, and which you can use in Codex and ChatGPT Work, is called GPT-6.1 Sol. "Argon" belongs to Google: Gemini 4 Argon. "Fable" belongs to Anthropic: Claude Fable 5.1.

The models side by side

List prices per million tokens, in US dollars, before tax, pay-as-you-go API. Input and output are priced separately.

Model	Provider	Input / output	Context	What the official source says
Claude Opus 5.5 (claude-opus-5-5)	Anthropic	$4 / $20	1M tokens (max output 128K)	Released 22 September 2026. Built for long-running agentic coding and knowledge work. Adaptive thinking is always on.
Claude Sonnet 5.5	Anthropic	$2 / $10	1M tokens	The middle tier of the family, with "fast" latency according to Anthropic.
Claude Fable 5.1	Anthropic	$10 / $50	1M tokens	The highest-priced model in the family, with "slower" latency according to Anthropic.
Claude Haiku 5.5	Anthropic	from $0.10 / $0.50	1M tokens	The fastest in the family.
GPT-6.1 Sol (gpt-6.1-sol)	OpenAI	$2 / $10 (cached input: $0.10)	1,050,000 tokens (max output 128,000)	Introduced 29 September 2026. Text and image input, text output. Data residency in the US and the EU. Above 272K input tokens the price goes up.
Gemini 4 Argon (gemini-4-argon)	Google	$2 / $10 through 30 November 2026; from 1 December, $4 / $20	not stated on the pricing page	Paid tier only: not on the free tier. Output includes thinking tokens.

Sources: Anthropic, Claude Opus 5.5, OpenAI, GPT-6.1 Sol and Google, Gemini API pricing. Prices change: always check the provider's page before you budget.

What to know about each

Anthropic: Claude Opus 5.5 and the family

The 5.5 family has four steps (Haiku, Sonnet, Opus and Fable) with the same 1M-token context and very different prices, so you can use the small one for bulk work and the big one only where it matters. It is also offered through Amazon Bedrock, Microsoft Foundry and Google Cloud, which helps if your company already buys from one of those clouds. One technical detail of Opus 5.5: thinking cannot be switched off and is steered with the effort parameter.

OpenAI: GPT-6.1 Sol

This is the model OpenAI has put in Codex and ChatGPT Work since 29 September. Its official page lists data residency in the US and the EU, which matters if you process personal data; confirm in your contract which region you actually use. OpenAI says it comes close to its GPT-6 Astra model at one fifth of the price; that is the company's own claim and we have not measured it.

Google: Gemini 4 Argon

Google announced it on 30 September and it is now in the paid API. The launch price is temporary and doubles on 1 December: if you are budgeting a project, use the December price. We covered it in Gemini 4 Argon vs Claude Opus 5.5.

How to choose without chasing the hype

• Start from the use case, not the model. Define one measurable task (summarising customer records, sorting email, answering questions about documents) and test two or three models on your own examples.
• Do the sums with your volume. Within one family, the cheapest and the dearest model differ by a factor of a hundred in input price. For simple bulk tasks the small one is usually enough; keep the big one for hard work.
• Check where data is processed. Region, contract and retention depend on the platform you use. Get it in writing.
• Don't marry one provider. Models and prices change every quarter; design your automation so you can switch.

When you don't need the cloud: the private alternative

If what you handle must not leave the company, or usage is constant and high, an open model in your own office can pay off. You don't beat the best paid model at everything, but your documents travel to nobody. We compared the tools for setting it up in Ollama, AnythingLLM and LM Studio, and we explain the criteria in IA local para empresas (in Spanish).

And if you would rather not build it yourself, ForgeNEX sets up your office server: hardware, network, backups and setup, so private AI works with your data inside (indicative price from €3,690 + VAT in October 2026; final quote after a call; the configurator is in Spanish). Our AI agents for businesses use models hosted in the EU; tell us about your case.