OpenAI cuts GPT-5.6 Luna price by 80%

OpenAI cut prices across two tiers of the GPT-5.6 family on 30 July 2026. Luna fell 80% to $0.20 per million input tokens and $1.20 per million output. Terra fell 20% to $2 and $12.

Flagship Sol pricing was unchanged. It received a faster inference mode at 2.5x speed for 2x cost, replacing a 1.5x option at the same premium. OpenAI said the Luna and Terra reductions also flow through to subscription usage counting in Codex and ChatGPT Work.

OpenAI attributed the reductions to efficiency across the models, the inference systems and the agentic harness. It described using GPT-5.6 within Codex to analyse production traffic, rewrite GPU kernels and optimise speculative decoding, reporting roughly 20% lower serving costs and 15% better token-generation efficiency.

The GPT-5.6 family launched on 9 July 2026. In the intervening three weeks Anthropic released Claude Opus 5 at $5/$25 and Google shipped Gemini 3.6 Flash at $1.50/$7.50. DeepSeek V4 was serving at $0.14/$0.28 with 80.6% on SWE-bench Verified.

The efficiency work spanned three layers: the models, the inference systems serving them, and the agentic harness connecting them to tools and context. Kernel rewrites accounted for roughly 20% of the serving-cost reduction; speculative decoding improvements added about 15% better token generation efficiency.

Pricing pressure across the market has been concentrated at the cheap tier, where open-weight and Chinese providers compete most directly. Frontier-tier pricing has held or risen: OpenAI priced GPT-6 Astra at $10/$50 on 3 September, and Anthropic's Claude Fable 5.1 lists at the same headline rates.

OpenAI has argued since this announcement that buyers should compare cost per completed task rather than cost per token, a framing that favours models finishing agentic work in fewer turns.