From Ars Technica · · 2m
AI models get cheaper, not smarter
OpenAI and Anthropic release new models focused on cutting costs rather than breaking new ground.
In brief
OpenAI and Anthropic release new models focused on cutting costs rather than breaking new ground. The AI industry is shifting from raw capability races to practical, cost-effective deployment as enterprises prioritize predictable, affordable solutions over frontier performance. Originally reported by Ars Technica.
Two AI giants chase efficiency
Anthropic released Opus 5.5 while OpenAI announced GPT-6 Sol and Luna. Both companies are prioritizing lower costs over new capabilities as enterprises seek cheaper alternatives to expensive frontier models.
Anthropic's Opus 5.5 cuts prices sharply
Input and output tokens cost 20% less than Opus 5, while cache reads drop 60%. The model also generates output 30% faster and uses fewer tokens per task, delivering roughly 40% savings on typical workloads.
“Cache reads are $0.20 per million tokens, 60% less than Opus 5.”
Opus 5.5 matches Astra on some tasks
Benchmarks show Opus 5.5 performing better at coding and knowledge work than OpenAI's GPT-6 Astra in some cases, though improvements are modest. This helps Anthropic catch up after Astra's recent release.
OpenAI's Sol and Luna cost half as much
GPT-6 Sol costs $2 per million input tokens and $10 per million output tokens. Luna is even cheaper at $0.10 and $0.50. Both are a few percentage points more capable than predecessors while cutting costs in half.
Model names map to different tiers
OpenAI's Astra is the most powerful and expensive. Sol is the daily driver. Terra is balanced for general use. Luna is fast and cheap. They roughly correspond to Anthropic's Fable, Opus, Sonnet, and Haiku, though the mapping isn't perfect.
Enterprises want predictable costs
Companies are less focused on frontier performance and more focused on operationalizing AI cost-effectively. Model orchestration and routing are now as important as raw capability for real-world deployment.
The frontier is still moving forward
Today's models are more capable and aligned than those from months ago. But the real bottleneck is no longer raw AI performance. It's how organizations contextualize and operationalize these tools with human expertise.
A natural slowdown is emerging
As AI leaders call for safety-focused slowdowns, market forces are creating their own brake. Current models are good enough for many tasks, but require sophisticated human oversight and reasonable costs to be useful at scale.
The one thing
The AI industry is shifting from raw capability races to practical, cost-effective deployment as enterprises prioritize predictable, affordable solutions over frontier performance.
