August reshuffled LLM prices in a way worth watching, not because a model dropped that changes everything, but because the cost per token is settling downward and the tiers are reorganising. On 10 August Anthropic made Claude Sonnet 5's introductory price permanent, 2 dollars per million input tokens and 10 output. On 12 August it updated Opus 5, released 24 July at 5 and 25 dollars, with faster inference. On 6 August OpenAI restructured its lineup around durable capability tiers (Sol, Terra, Luna) and launched ChatGPT Work, an agentic system for long-running projects. Google shipped Gemini 3.7 Flash on 13 August, Alibaba Qwen3.8 Max on the 2nd. If your code still picks the model you had in June, you are paying old prices.
What actually changes if you run a feature in production
The lesson is not "use the cheapest model". It is that the price gap between a frontier model and a mid-class one is now wide enough that picking the model by task, rather than by habit, is the single largest optimisation you have. Opus 5 at 5 and 25 dollars against Sonnet 5 at 2 and 10 is a factor between two and a half and five: classifying a ticket, extracting fields from a document, tagging content do not need the flagship.
Task tiering, concretely
- Deterministic, high-volume tasks: classification, extraction, tagging, short rewriting go on the mid-tier or flash model. This is where you spend without noticing.
- Reasoning and long generation: complex synthesis, code, multi-step analysis justify the frontier model, but only on that stretch of the flow.
- The model id stays configuration, not code: as we wrote in July on vendor dependency, the model should be read from a variable and hidden behind one interface. Changing tier or vendor should cost a deploy.
- Your own eval suite: twenty real cases from your domain with expected output. Without it you never know whether dropping a tier degrades quality or just the bill.
Verdict
LLM prices change every month, and August proved it with four announcements in two weeks. Anyone who wired a single model into the code pays twice: on the bill today and in refactoring when they have to change it. Put an abstract interface in front of the model, keep twenty tests that tell you whether a lower tier holds, and review tiering every quarter as you review cloud spend. Half a day of work that pays for itself on the first invoice.