OpenAI Slashes GPT-5.6 Luna Pricing by 80% as It Chases the Cost-Per-Task Frontier

OpenAI is cutting the price of its cheapest GPT-5.6 model, Luna, by 80 percent, while trimming the flagship Terra tier by 20 percent. Both changes take effect July 30, 2026, and apply across ChatGPT Work, Codex, and the OpenAI API.

The new numbers are stark. Luna now costs $0.20 per million input tokens and $1.20 per million output tokens, down from its prior rate. Terra moves to $2 per million input tokens and $12 per million output tokens, still a premium tier but noticeably cheaper than before.

OpenAI frames the cuts around a "price-performance frontier" argument rather than a simple discount. The company estimates Luna now costs roughly 6 cents per task compared with frontier-class models from just a year ago, while running nearly nine times faster.

That efficiency claim extends to quality, not just cost. OpenAI says Luna outperforms its earlier Fable 5 model on professional-grade work, as measured by the Agents' Last Exam benchmark, at an estimated cost per task nearly 99 percent lower.

Where the Numbers Diverge

Reporting on the magnitude of these cuts isn't perfectly aligned. OpenAI's own materials and third-party coverage agree that Luna dropped 80 percent and Terra dropped 20 percent, but some accounts conflate or transpose which tier got which cut given how closely the announcements were bundled. The verified figures above reflect the pricing tables OpenAI published alongside the July 30 rollout.

Notably, GPT-5.6 Sol, the mid-tier model, saw no pricing change at all. Instead, OpenAI focused on throughput: a new Fast mode for Sol in the API delivers up to 2.5 times the speed of Standard processing, at double the price for those who need lower latency.

Sol also picked up quieter infrastructure gains. Speculative decoding improved its token generation speed by more than 15 percent, and optimized GPU software cut deployment costs by 20 percent, savings OpenAI appears to be routing into margin rather than passing on as a price cut for that specific tier.

Why This Matters

The timing and scale of the Luna cut invite comparison to aggressive pricing moves from Chinese model providers, a framing several outlets have leaned into directly. For developers building high-volume, low-margin agent workloads, an 80 percent reduction on the cheapest tier meaningfully changes unit economics, especially when paired with claims of near-99 percent lower cost per task against a prior-generation model.

For enterprise teams already running Terra in production, the 20 percent reduction is smaller but still consequential at scale, particularly for output-heavy workloads where the $12-per-million-token rate applies. Sol's unchanged base pricing, offset by the new Fast mode option, suggests OpenAI is segmenting its lineup more deliberately: Luna for cost-sensitive volume, Sol for latency-tunable general use, and Terra for premium reasoning work.