An 80% cut three weeks after launch isn't generosity. It's a floor being set by DeepSeek

Read the announcement and you get a story about engineering excellence. Read the calendar and you get a different one. GPT-5.6 launched on 9 July. On 24 July, Anthropic shipped Claude Opus 5 at $5/$25. Google had Gemini 3.6 Flash out at $1.50/$7.50. DeepSeek V4 was serving at $0.14 input and $0.28 output while posting 80.6% on SWE-bench Verified. On 30 July, OpenAI cut its cheapest tier by 80%.

If this were purely an efficiency pass-through, you would expect it to show up across the family. It didn't. Luna fell 80%, Terra 20%, and Sol — the flagship, the highest-revenue line, the one with no cheap open-weight equivalent — held flat at $5/$30 and got a speed option instead.

That distribution is a map of where competition actually bites. At the bottom, where DeepSeek and the open-weight models live, OpenAI cannot hold a premium and has stopped trying. In the middle, it trims. At the top, where the buyer is an enterprise that will not run frontier legal or engineering work on a Chinese open-weight model, the price does not move — because it does not have to.

The efficiency work itself is not spin, and the detail is more interesting than the discount. OpenAI ran GPT-5.6 inside its own Codex agent against production traffic to rewrite GPU kernels, improve routing heuristics and optimise speculative decoding — reporting around 20% lower serving costs from the kernels and 15% better token-generation efficiency from the decoding changes.

Take that seriously. Not because it is 'AI improving itself' — it isn't, in the sense people mean — but because it means inference cost is no longer purely a function of hardware generations. A lab that can point a competent coding model at its own serving stack every quarter gets a compounding cost advantage over one that can't. That is a real moat, and it is the one OpenAI should be shouting about rather than the discount.

Practically: re-run your cost model. Luna at $0.20/$1.20 changes which workloads justify a frontier model at all, and the subscription arithmetic changed too — OpenAI says the lower rates flow through to how usage is counted in Codex and ChatGPT Work, so the same seat goes about five times further on cheap-tier work.

Strategically: notice the argument OpenAI started making around this cut and has made louder ever since — that you should judge cost per completed task, not cost per token. That framing is correct on the merits for agentic work, where a smarter model finishing in fewer turns genuinely wins. It is also, conveniently, the only framing in which a $10/$50 flagship like GPT-6 Astra looks cheap next to open weights at a twentieth of the price. Both things are true. Keep both in view when the next price announcement lands.

There is a second-order effect worth naming, because it shapes what gets built. When the cheap tier is genuinely capable and costs a fifth of what it did last month, the rational architecture stops being 'pick one model' and becomes 'route between them'. Cheap tier drafts, filters, classifies and does the eighty percent of calls that were never hard; the expensive tier gets invoked only where it demonstrably earns its keep. That is exactly the pattern OpenAI's own harness work is built around, and it is why the routing heuristics got optimisation attention alongside the kernels. The model that wins a workload in 2026 is increasingly the router, not the frontier system it occasionally calls.

One caution for anyone modelling this forward. A cut this steep, this fast, is not a promise about future pricing — it is a response to a specific competitive moment, and the tier that did not move shows where the pricing power still sits. Build on the assumption that cheap-tier costs keep falling and frontier-tier costs do not, and you will be right more often than the reverse.