← Blog October 12, 2026

The AI price war of October 2026: time to recalculate your AI costs

OpenAI, Anthropic, Google and Mistral have lowered their prices or released cheap new models in two weeks. What changed, where the fine print is, and what you can do with it.

Stapels munt-achtige fiches van verschillende hoogtes die kleiner worden, als symbool voor dalende AI-prijzen

Anyone who has followed AI news in recent weeks noticed something striking: major providers are lowering their prices simultaneously. Between late September and early October 2026, OpenAI, Anthropic, Google and Mistral all came out with cheaper models or significant price cuts. For entrepreneurs using or considering AI, this is the moment to do the math again. Automations that weren't feasible half a year ago might be possible now. But the headlines don't tell the whole story.

What changed in two weeks

OpenAI GPT-6.1 Sol (September 29)

OpenAI released GPT-6.1 Sol for $2 per million input tokens and $10 per million output tokens. That's exactly one-fifth the price of GPT-6 Astra, OpenAI's most powerful model. According to OpenAI, Sol performs close to Astra on agentic programming, computer use and professional work. A separate Astra 6.1 release was cancelled. Sol is available via the API, ChatGPT Work and Codex; it wasn't yet in regular ChatGPT chat as of October 6.

Anthropic Claude Haiku 5.5 (October 7)

Anthropic released Claude Haiku 5.5, starting at $0.10 per million input tokens and $0.50 per million output tokens, matching OpenAI's GPT-6 Luna. The model is designed for repetitive work that makes AI expensive at scale: summarizing documents, classifying information, querying databases and performing smaller tasks for heavier agents. It runs on Anthropic's own platform and via AWS, Google Cloud and Microsoft Azure. Important for those building on it: Anthropic commits not to phase out the model before October 7, 2027. At the same time, cache reads for Sonnet 5.5 became 50% cheaper.

Google Nano Banana 2.1 (around October 7)

Image generation is also getting cheaper. Google's Nano Banana 2.1, which runs on Gemini 3.6 Flash, roughly halves the price: a 1K image drops from $0.067 to $0.0336, a 4K image from $0.151 to $0.0756. These figures come from a news roundup; check them against Google's own announcement before relying on them.

Mistral Large 4 (October 6, preview)

European Mistral presented a preview of Mistral Large 4, trained on and running from its own European infrastructure. It supports more than 160 languages, including all official EU languages. The preview API can be tested via Mistral Studio; weights follow later this month. Sources disagree on the price: one mentions figures, another reports that Mistral hasn't published prices for this preview yet.

Read the fine print

A 90% price reduction sounds spectacular, but reality is more nuanced:

Magnifying glass over blank grid sheet as image for reading the fine print

Thresholds for long prompts. With Haiku 5.5, the low price applies up to 100K tokens; above that you pay $0.50/$2.50, a 50% reduction rather than 90%. With GPT-6.1 Sol, prompts over 272K input tokens are charged for the entire request at 2x the input and cache rate and 1.5x the output rate.
Average savings are lower than the headline. Anthropic itself estimates that workloads work out about 75% cheaper than on Haiku 4.5, not 90%.
How you call it matters. Haiku's Batch API gives 50% discount; at OpenAI, Batch and Flex halve the price, while Fast mode doubles it. Smart caching and running non-urgent tasks in batches can make more difference than the model choice itself.

What this means for your business

The tasks that SMBs often perform in large numbers — summarizing emails and documents, sorting support tickets, simple classification — now cost a fraction of what they did a few weeks ago. Three practical steps:

Recalculate your business case. Take an idea you previously thought was too expensive and compare it again to the new rates, including your typical prompt length.
Distribute work across models. Let a cheap model do the bulk work and only use a heavier model where it really matters.
Test European options. If you value data sovereignty and EU hosting, Mistral Large 4 is a serious candidate. Since it's still a preview: test it, but don't yet put production work on it. Self-hosting isn't realistic for small businesses; the weights alone take up about 525 GB on INT4.

Low prices are no guarantee

There's also a dissenting voice. In Dutch financial media, investor Ray Dalio (IEX, October 9) warns that rapid growth in AI investment lending and rising interest rates are bringing markets closer to the point where the AI bubble could burst. What's cheap today doesn't have to be tomorrow. So build in a way that lets you switch models or providers without overhauling your entire application: a thin abstraction layer between your software and the model, and keep your prompts and tests decoupled from one vendor.

At Silvar we use AI daily, for our own development and in the document processing, customer service and data disclosure we build for clients. Curious whether an automation you previously wrote off would be worthwhile with these new prices, or how to set up your solution vendor-independently? We're happy to do the math with you.