On 30 July 2026, OpenAI cut the API price of its two cheaper GPT-5.6 tiers: Luna dropped 80% to $0.20 per million input tokens and $1.20 per million output tokens, Terra fell 20% to $2/$12, while flagship Sol pricing held. OpenAI’s own engineering writeup credits the cut to real infrastructure gains, not a loss-leading discount — rewritten GPU inference kernels cut serving costs 20%, and a redesigned speculative-decoding draft model lifted token-generation efficiency by more than 15%. Subscription pricing for ChatGPT and Codex is unchanged; usage-based credit consumption for Luna and Terra just got cheaper.
This is one data point in a pattern that’s been running all year: Anthropic, Google and OpenAI have each cut or restructured pricing on their coding-relevant models multiple times since June, and every cut so far has come from genuine efficiency work rather than a race to the bottom on margin. For anyone running AI-assisted development at production scale — agentic coding pipelines, high-volume code review, automated test generation — the unit cost of the underlying model calls has been falling faster than most build budgets get revised to reflect it.
So what
If your cost estimates for an AI-assisted build still reflect model pricing from even three or four months ago, they’re probably overstating the ongoing running cost. That doesn’t change the core commissioning decision — which tools and which team can actually ship a reliable product — but it does mean the “AI usage costs” line in a build budget deserves a fresh look before you assume it’s a fixed constraint. If you’re planning a build that leans on AI-assisted development and want realistic numbers rather than stale pricing assumptions, see our AI-assisted development work or get in touch to talk through the approach.