If you've felt whiplash trying to figure out whether AI is getting cheaper or bleeding your budget dry, congratulations: you're paying attention. The Register just published a sharp piece of reporting on the great split AI token costs. Commodity inference prices are racing toward free, while frontier models quietly moon-shot in the other direction. The piece draws extensively on commentary from Larridin CTO Ameya Kanitkar.
The headline describes the situation vividly: "AI is becoming a bargain hunter's market, with a few luxury models on top." The article makes three key points.
1. Commodity inference pricing is collapsing, while frontier model pricing rises sharply.
The article cites math from AI engineer Aman Panjwani: GPT-4-class output ran about $20 per million tokens in late 2022. Today, equivalent capability goes for roughly $0.40, a drop of more than 50x in just four years. The article cites the introduction of DeepSeek's R1, which landed in January 2025 at roughly $2 or less per million input/output tokens, when OpenAI's o1-preview was charging more than $10 for the same. Today, high-end models are doubling and tripling in price as commodity token pricing drops.
2. Nobody budgeted for "agentic," and now the bill has arrived.
Larridin CTO Kanitkar focused on “pay as you go” pricing. He told The Register that six months ago, AI cost conversations were focused on the cost of LLM subscriptions, ranging from $20-$100/month per user for LLM subscriptions. But as the new year gained momentum, so did a move toward per-token pricing and agentic AI, which adds a new layer of Ops to the Dev work that models had already started doing. Result: "On average we have seen the cost go up about 10x between January and now, especially in engineering ops," Ameya said. Ameya puts token spend at 10-20% of an engineer's fully loaded labor cost these days; $2,000 to $4,000 a month on top of a $200K salary, and rising quickly. That's not a rounding error anymore.
3. Most of the new spending isn't buying much.
Kanitcar then cites Larridin research to show where the money’s going: 15-30% of AI users account for more than half of total spend, and that spending frequently does not correlate with better output. Larridin found that only the first 35-40% of token burn was tied to rapid productivity increases; as costs rise, gains fall. The Larridin CTO suggests that usage can be safely capped at that point with no loss of productivity.
The picture is still shifting. Open-weight models such as Kimi 2.6/2.7 and GLM 5.2 are closing in on frontier quality at a fraction of the price, but nearly half of enterprise AI spend still flows to Anthropic's high-end Opus model anyway. Bargain-hunting for tokens only works if you know what you're actually spending and where it's going, the exact problem Larridin exists to solve. That's the difference between guesstimating your AI budget, and managing it in the dark, vs. making AI a cost-effective tool in a company’s arsenal.
Read the full article on The Register: AI is becoming a bargain hunter's market, with a few luxury models on top