Language models have progressed over eight years; costs fall while task-targeted models rise

Summary

The article chronologically traces eight years of language model evolution, highlighting the dramatic drop in costs and the emergence of highly specialized task-focused agents.

Why it matters

Understanding this shift is crucial for developers and businesses because it reshapes the pricing and distribution of AI services. The collapse of the capability-cost curve shows that powerful models are now accessible at a fraction of their previous expense, paving the way for the democratization of advanced AI capabilities.

Highlights

• Coding ability has surged nearly sixfold per year since the end of 2024, marking a massive leap.
• OpenAI's GPT‑5.6 Luna now delivers flagship performance at just $1–$6 per token, vastly outpacing older models.
• Specialized agents dominate specific domains: Claude Opus 5 excels in front‑end coding, Claude Fable 5 shines in warehouse‑level tasks, and GPT‑5.6 Sol leads in terminal environments.

Critical take

The article relies heavily on proprietary benchmarks and cost structures, which may limit its universal applicability across different platforms or real‑world scenarios.