A bank’s customer-service team fields thousands of routine questions a day. Where is my card? Why was I charged twice? Each answer artificial intelligence writes carries a small cost, and at that volume, the costs add up fast.
This week, that math changed. The price of a basic AI answer fell, and companies gained more choice over where their AI runs.
We’d love to be your preferred source for news.
Please add us to your preferred sources list so our news, data and interviews show up in your feed. Thanks!
Anthropic Drives Down the Price of Small Models
- Anthropic cut prices on its fastest model, Claude Haiku 5.5, which launched Wednesday (Oct. 7). Anthropic says it costs one-tenth the price of its predecessor, Claude Haiku 4.5, for most requests and about 75% less on average. The model targets high-volume work such as customer support, summaries and data lookups. The release follows the September debut of Opus 5.5, Anthropic’s flagship model, which costs 40% less to run than Opus 5, PYMNTS reported.
- Businesses can choose how hard the model thinks across five effort levels, trading speed for depth. Financial research firm Rogo already uses it for quick lookups. Alex Wang of Rogo said in Anthropic’s launch announcement that the model “goes into the 10-K and pulls the segment revenue line the deck needs.”
- Anthropic said the model beats OpenAI’s comparable small model, GPT-6 Luna. In Anthropic’s own tests, Haiku 5.5 outscored Luna on every benchmark both models were tested on. On a test of how well AI operates computer software, it scored 72.4% to Luna’s 48.9%.
Mistral, Reflection and Aleph Alpha Expand Open Model Options
- Mistral introduced a flagship model companies will be able to host themselves. Mistral Large 4 launched as a public preview on Tuesday (Oct. 6). It has 1.05 trillion parameters, the internal settings a model learns in training. More parameters usually mean more capability and higher costs. Once Mistral releases the model’s open weights later this month, firms can run it on their own servers and keep sensitive data in-house. It reads images and text and can take in a full contract or loan file at once. Mistral is offering it at half price during the launch period.
- A U.S. startup took aim at China’s lead in open models. Reflection AI unveiled Beam on Monday (Oct. 5) for coding and automated business tasks. The model has 501 billion parameters but uses only 23 billion at a time, which keeps it cheaper to run. Reflection says it matches Z.ai’s GLM-5.2, a leading Chinese open model, on reasoning while using one-third to one-quarter of the computing power. The most powerful open models are almost all Chinese, Axios reported. Cost is already pushing software companies toward open-weight models, including Chinese ones, PYMNTS reported. Nvidia-backed Reflection could offer a cheaper alternative to Anthropic, OpenAI and Google. Reflection plans to release Beam’s weights under a free, open-source license later this month.
- A German AI company built a model for regulated work. Aleph Alpha released Kolibri on Saturday (Oct. 3) for compliance-heavy tasks in English and German. It has 78 billion parameters but switches on only 3 billion at a time, which keeps costs low. It can process very long documents in one pass. Companies can use and modify it for free.
Amazon and Reka Push AI Into Voice and Video
- Amazon sped up voice AI. Amazon Web Services (AWS) made Nova 2.5 Sonic available on Monday. The model listens and speaks directly, without first converting speech to text and back. That lets AI agents hold live phone conversations with customers. AWS says this version reasons better, follows instructions more closely and responds with less delay. It also pulls from other business tools more accurately. That accuracy is what an agent needs to check an account balance or reschedule a payment mid-call.
- Reka combined text, images and video. The startup released an early version of Rho-1 on Monday. Reka said the single model reads and creates text, images and video, and can act on its own. It is a research preview, not yet a commercial product.
The Big Picture: The cost of an AI answer keeps falling. The harder question for financial firms is which model to trust with their customers, and where it should run.
For all PYMNTS AI coverage, subscribe to the daily AI Newsletter.
Add as Preferred Source
