OpenAI and Anthropic ship new models on the same Monday, and both cut prices

Sol is OpenAI's mid-priced model for recurring work like code review, debugging and data analysis. It costs $2 per million input tokens and $10 per million output tokens, half the GPT-5.6 Sol rate. Luna, the high-volume tier, costs $0.10 and $0.50. OpenAI told VentureBeat these are "permanent prices, not promotional or introductory pricing." Both are rolling out in the API, ChatGPT and Codex, and Luna is available to free users on desktop.

Anthropic priced Opus 5.5 at $4 per million input tokens and $20 per million output tokens, down from $5 and $25 for Opus 5, with cache reads 60% cheaper. Anthropic says it costs about 40% less on typical workloads and generates output about 30% faster. It also raised five-hour usage limits on its Pro, Max, Team and Enterprise plans.

The same-day timing shows how crowded releases have become. Google shipped Gemini 3.8 Flash three weeks earlier.

What it means for your costs

If you don't build on these APIs yourself, the short version is that the AI inside the tools you pay for just got cheaper to run, and some of that should reach you as more generous limits. Anthropic's higher usage limits on its paid plans are the first example.

If you do build on them, here's what the cuts look like on a real job. Take a feature that makes 100,000 calls a month, each using about 1,500 input tokens and 400 output tokens. That's 150 million input tokens and 40 million output tokens a month (FER calculations from the list prices above):

• Opus 5 at $5 and $25: $750 plus $1,000, so $1,750 a month.
• Opus 5.5 at $4 and $20: $600 plus $800, so $1,400 a month.
• GPT-6 Sol at $2 and $10: $300 plus $400, so $700 a month.
• GPT-6 Luna at $0.10 and $0.50: $15 plus $20, so $35 a month.

The top and bottom of that list are fifty times apart. That's why "which model?" is now a pricing decision as much as a technical one.

What to do this week

Don't switch on launch day. Do this instead:

1. Pick the two or three features where you spend most on model calls.
2. Pull 50 real examples of each from your logs, along with the output you'd count as good.
3. Run them through the new models and score the results the same way you score the current one.
4. Compare cost per good answer, not cost per token. A cheaper model that needs two attempts isn't cheaper.
5. Move the work where the new model is as good and cheaper. Leave the rest alone.

If you're an agency building AI features for clients, the same steps give you a tidy story for your next client review: "We reran the numbers after September's price cuts, and here's what we've saved you."

Disclosure: this publication's research agent runs on Claude Opus 5.5. We've reported its launch from Anthropic's published figures only.