OpenAI slashes GPT-5.6 Luna price 80% in bet on cheaper AI

The News

OpenAI has sharply cut the cost of running its latest models, positioning cheaper intelligence rather than bigger models as the next front in the AI race. In a post published on 31 July titled "Building abundant intelligence", the company said it had reduced the price of GPT-5.6 Luna by 80 percent, taking it to $0.20 per million input tokens and $1.20 per million output tokens.

Its mid-tier GPT-5.6 Terra model was cut by 20 percent, to $2 and $12 per million input and output tokens respectively. A faster GPT-5.6 Sol setting now runs at up to 2.5 times the speed of standard processing at twice the price.

The reductions are underpinned by engineering gains. OpenAI said it had lowered end-to-end serving costs by 20 percent and improved token-generation efficiency by more than 15 percent. On the public ARC-AGI-3 reasoning benchmark, it raised Sol's score from 13.3 percent to 38.3 percent while using six times fewer output tokens.

The company now claims more than one billion active users and more than two million businesses on its platform, with agentic work routed through its Codex tool accounting for 99.8 percent of weekly output tokens.

Why It Matters

The message beneath the numbers is a strategic one: "When the cost of useful intelligence falls, more work becomes worth doing." OpenAI is arguing that the winners will be decided by price per useful outcome, not by parameter counts or leaderboard-topping demos.

That is a notable shift in tone. When GPT-4 arrived in March 2023, the story was raw capability and scale. Two years on, the frontier labs are competing on unit economics, echoing the classic cloud-computing playbook in which falling storage and compute prices unlocked wave after wave of new businesses. Cheaper tokens widen the pool of applications that make commercial sense, which in turn funds the next round of infrastructure.

For rivals, the pressure is obvious. Once a leading lab resets pricing this aggressively, the rest of the market tends to follow, compressing margins across the board and forcing a rethink of who actually captures the value.

Indian Angle

For India's developer and startup economy, the arithmetic is the headline. At roughly 88 rupees to the dollar, GPT-5.6 Luna's new input price works out to under 18 rupees per million tokens, a level at which token cost stops being the binding constraint for most consumer-facing products. That is meaningful for the thousands of Indian SaaS and fintech teams building copilots, support agents and document tools on foreign APIs.

It also sharpens the question facing domestic model builders such as Sarvam and Ola's Krutrim, both tied to the IndiaAI mission's compute ambitions. Their pitch has leaned on data sovereignty and Indian-language strength; competing on raw price against an 80 percent cut is far harder. Expect the debate over subsidised local compute and government procurement preference to intensify.

For enterprise buyers, from banks to business-process outsourcers, cheaper inference improves the case for automating high-volume workflows. But it also deepens dependence on a single overseas vendor, a concern MeitY and data-localisation advocates are likely to revisit.

FAQ

What exactly changed?

OpenAI cut GPT-5.6 Luna's price by 80 percent and Terra's by 20 percent, alongside efficiency work that lowered end-to-end serving costs by 20 percent and lifted token-generation efficiency by more than 15 percent.

How cheap is this for Indian developers?

Luna now costs $0.20 per million input tokens, roughly 18 rupees, making token spend a minor line item for many consumer apps built in India rather than a core cost worry.

What does it mean for Sarvam and Krutrim?

Domestic labs face stiffer price competition and will likely lean harder on Indian-language capability, data residency and government-backed compute rather than headline token pricing.

Where can I read the original announcement?

OpenAI published the details in its "Building abundant intelligence" post on 31 July 2026.

This story was reported by OpenAI. Read the full original coverage at OpenAI.