OpenAI Just Slashed GPT-5.6 Luna by 80% — And the AI Cost Curve Just Bent Again

On July 30, 2026, OpenAI did something that would have been unimaginable twelve months ago: it cut the input price of its fastest GPT-5.6 tier by four-fifths. GPT-5.6 Luna dropped from $1 to $0.20 per million input tokens and from $6 to $1.20 per million output tokens. The mid-tier GPT-5.6 Terra got a smaller but still aggressive 20% trim, landing at $2 input and $12 output. The flagship GPT-5.6 Sol held its $5/$30 sticker price but picked up a new Fast mode that runs 2.5x standard speed for 2x the standard cost.

The cuts redrew the GPT-5.6 price graph in a single afternoon. Luna now costs roughly 4% of Sol's input rate — one twenty-fifth, down from one-fifth before. The gap between the cheapest and most expensive tier in the same product family is now wider than the gap between Luna and the open-weight competition. That is not an accident. It is a strategy.

The numbers, not the narrative

OpenAI's post frames the change as "advancing the price-performance frontier," and the benchmarks back the framing. Luna scores 66.6 on BenchLM's composite index, ranks #25 of 214 models in early August 2026, and posts its strongest score in Coding at #6 — ahead of last-gen flagships and roughly even with GLM-5.2 (max) and Gemini 3.5 Flash, at a fraction of the per-token cost.

The per-task economics tell the cleaner story. Artificial Analysis's Intelligence Index puts a GPT-5.6 Luna (max) task at about $0.21, Terra (max) at $0.55, Sol (max) at roughly $1.00. A 5x spread on a task-equivalent basis — and the cheap tier is still beating frontier competitors on raw benchmarks.

Meanwhile, Sol took new state of the art on the Artificial Analysis Coding Agent Index at 80 — 2.8 points above Fable 5 — using less than half the output tokens, less than half the wall-clock time, and about one-third the cost of the previous leader. The flagship got faster and cheaper. The mid-tier got cheaper. The cheap tier got much, much cheaper.

What it actually means for builders

The price chart is now a routing problem. For an application sending a million requests a day, the difference between routing everything to Sol and routing 80% to Luna is the difference between a unit-economics-positive product and a fundraising-dependent one. The hard part is no longer whether to afford a frontier model. It is whether you can route the right request to the right tier without losing accuracy where it matters.

A few patterns are emerging:

• Luna for the long tail. Classification, extraction, summarization, RAG, code completion, simple tool use — structured work over contextualized input. The bar is "good enough," and Luna cleared that in 2025.


• Terra for everyday reasoning. Multi-step agent loops that need coherent planning without paying frontier prices. Terra is now the sensible default for most agentic workflows.


• Sol for the hard 5%. Open-ended generation, ambiguous reasoning, anything where a wrong answer costs more than a slow one. The new Fast mode makes Sol viable for latency-sensitive use cases that previously could not afford it.


• Mix-and-match pipelines. Teams that built a single-model abstraction layer two years ago are now rebuilding it as a router — a thin layer that classifies the request, picks a tier, calls the right model, and verifies the result. The router itself is becoming the product.

The competitive context

The cut lands in a market already sliding. Qwen3.8-Max's open-weight preview arrived in mid-July at 2.4T parameters with ~95B active per token — frontier-class reasoning at inference economics that closed most of the gap with closed APIs. Claude Sonnet 5 went GA the same week at $2/$10 per million tokens through August 31, reverting to $3/$15. Anthropic's framing: frontier performance no longer has to cost frontier prices.

OpenAI's response was not to undercut the flagship. It was to flatten the curve across the whole family and turn Luna into the default for high-volume production traffic. Sol stays premium. Luna becomes the workhorse. Terra sits where most real workloads live.

It is a more sophisticated move than a price war. It forces every competitor to defend three price points, and gives OpenAI a credible answer to the open-weight narrative without giving up the high-margin tier. Whether it holds depends on whether Luna's quality stays above the open-weight floor as traffic scales — and whether the routing layer becomes a moat or a commodity.

The practical advice is short: stop optimizing for the cheapest model. Start optimizing for the best tier-routing strategy. The cheapest call in 2026 is the one that landed on Luna instead of Sol. The most expensive mistake is the one that landed on Luna when it needed Sol. That distinction is now the whole game.