OpenAI halves prices, Claude gets cheaper: Sept. 23

The bill for automated AI work just dropped, and any business paying by usage should check its model settings today. OpenAI halved the prices of two new GPT-6 models, while Anthropic cut the cost of its latest Claude model. Google also moved more shopping transactions from retailers' sites into its own AI results.

OpenAI halves the price of everyday GPT-6 work

OpenAI launched GPT-6 Sol and Luna on September 22 with usage prices 50% below their GPT-5.6 equivalents. Usage pricing is the separate bill charged when software sends automated requests to an AI model. It isn't included in a regular ChatGPT subscription.

Sol now costs $2 per million input tokens and $10 per million output tokens. A token is a small unit of text used to measure what the model takes in and sends back. Luna costs $0.10 per million input tokens and $0.50 per million output tokens. Both are available through OpenAI's developer service, while access inside ChatGPT Work and Codex is rolling out to paid users.

The immediate winner is a business already paying for AI inside customer support, document processing, or internal software. A lower list price won't fix a wasteful process, though. Teams using the ChatGPT app alone may see new model choices, but they won't receive a 50% subscription discount. Our ChatGPT pricing breakdown explains why app seats and automated usage belong in separate budget lines.

Anthropic cuts Claude's premium rate

Anthropic released Claude Opus 5.5 on September 22 at 20% lower input and output prices than Opus 5. The new model costs $4 per million input tokens and $20 per million output tokens. Anthropic says typical work costs 40% less because repeated instructions and background material are also cheaper to reuse.

That repeated material is called a cache. Think of it as letting the model refer back to text it has already processed instead of charging to read it from scratch each time. Opus 5.5 cache reads cost $0.20 per million tokens, down from $0.50. Anthropic also says the model writes responses more than 30% faster and has raised five-hour usage limits for several paid plans.

This is useful for teams running long research, analysis, or software tasks. It matters less if Claude is used for a few short drafts each week. Compare the new usage bill with a less expensive Claude model before moving every process to Opus. The difference between subscriptions and usage charges is mapped out in our Claude pricing guide.

Google brings checkout into AI answers

Google is telling some Shopify merchants that eligible products can now be bought inside AI Mode and Gemini, according to Merchant Center notices reported September 22. For affected stores, eligible items are included automatically unless the merchant turns the feature off in Shopify under Sales channels > Agentic.

The purchase happens on Google's screen through Google Pay, but the retailer remains responsible for the sale, shipping, returns, and customer support. Google's official merchant guidance says the program is limited to selected merchants and eligible products, with availability across the United States, Canada, and Australia. Customers may encounter it in Google Search's AI Mode and Gemini.

Retailers should treat this as a new storefront, not a small checkout tweak. Product titles, prices, stock status, brand assets, and policies in Merchant Center can now shape a sale that never reaches the usual website checkout. That also changes measurement because fewer shoppers will visit the store's checkout pages.

Check the setting before Google chooses the path for you. Merchants that keep native checkout enabled should place a test order, confirm tax and shipping behavior, and separate Google AI purchases in weekly reporting. Our guide to AI shopping discovery covers the wider shift from website visits to purchases inside assistants.

Worth watching

• OpenAI also introduced a dashboard for reused instructions and background text, with discounts of up to 90% on eligible cached input. This mainly matters to businesses running repeated automated tasks.
• Google's AI checkout is still limited. Watch whether access expands beyond selected retailers before changing a broader ecommerce plan.
• Two premium model launches in one afternoon make task-level cost testing more valuable than loyalty to a single AI provider.