Short answer
The new models OpenAI and Anthropic announced in September do not bring a major leap in capability; the real innovation is cost. Both companies are releasing versions that do the same work for less, to avoid losing enterprise customers to cheap open-weight models. Companies that build AI into their products should consider redoing their budget calculations.
Highlights
- Input and output prices for Opus 5.5 are 20% lower than for Opus 5.
- The price of cache reads fell by 60%; this is where most of the cost lies in agent and coding work.
- GPT-6 Sol and Luna deliver similar or slightly better results at half the price of their predecessors.

2 min readAuthor: UNIT Journal EditorEditor-in-chief: Uğur Deniz İlhan
For some time, every new AI model release came with the claim of being "smarter". This week the picture changed: OpenAI's GPT-6 Sol and Luna and Anthropic's Opus 5.5 converge on the promise of "a little more for a lot less money", in Ars Technica's assessment.
What do the new models change?
In short, the price. According to Anthropic, Opus 5.5 costs $4 per million input tokens and $20 per million output tokens, 20% cheaper than Opus 5. Cache reads, which make up most of the cost in agent and coding work, cost $0.20 per million tokens, 60% lower. The company says the model produces output more than 30% faster and that total savings on typical tasks approach 40%.
On OpenAI's side, GPT-6 Sol's API price is $2 per million input tokens and $10 per million output tokens; for Luna the figures are $0.10 and $0.50. OpenAI says both models run at half the cost of their predecessors and score a few points higher on some tests.
What do the model names mean?
- Astra: OpenAI's most powerful and most expensive model, for heavy coding and research.
- Sol: A more efficient, powerful model designed for everyday work.
- Terra: A balanced, general-purpose model.
- Luna: The fast, cheap option.
What does it mean for businesses?
For a company using AI in customer service, content production or its software team, this price change could feed directly into profit margins.
- Measure cost "per task", not "per token": the new models can finish the same task with fewer tokens.
- Use caching for long, repeated instructions; that is where the discount is largest.
- Do not put the most powerful model on every task: try lightweight models for simple classification and summarisation, and route only difficult work to the top model.
Frequently asked
- Are the new models much smarter than the old ones?
- No, the difference is limited. The real gain is doing the same work at lower cost and faster.
- Which model should I choose?
- It depends on how hard the task is. The most economical approach is to try a lightweight model first and move tasks where the result falls short to the top model.
Sources
- Ars Technica ·
Follow UNIT Journal
What's new in search, AI and technology, in your feed every day.


