GPT-5.6 Gets Cheap: OpenAI Cuts Luna 80%
OpenAI dropped the price of GPT-5.6 on July 30, and the cut on the cheap end is aggressive. Luna, the most cost-efficient tier, is down 80% to $0.20 per million input tokens and $1.20 per million output. Terra, the middle tier, is down 20% to $2 and $12. Sol stays the flagship at $5 in, $30 out.
The framing from OpenAI is price-performance frontier, which is corporate for a price war. But the number that matters isn't the discount, it's what a cheap tier does for agents specifically. An agent loop doesn't ask one question, it burns tokens by the billion. The autonomous-business experiment making the rounds this week chewed through 320 million tokens in a single day of one agent running one business. At the old Luna price that day cost real money. At 80% off, always-on agents that plan, retry, and self-correct all night suddenly pencil out.
That's the actual story of frontier pricing in 2026. The headline model gets the benchmarks, but the cheap tier decides which agent products can exist. Twenty cents a million input is the kind of number that turns a clever demo into something you can leave running.
This also isn't happening in a vacuum. Gemini Flash keeps undercutting, Kimi K3 shipped open weights you can self-host, and now OpenAI is defending the floor, not the ceiling. Full pricing at openai.com/index/advancing-the-price-performance-frontier-with-gpt-5-6.
← Back to all articles
The framing from OpenAI is price-performance frontier, which is corporate for a price war. But the number that matters isn't the discount, it's what a cheap tier does for agents specifically. An agent loop doesn't ask one question, it burns tokens by the billion. The autonomous-business experiment making the rounds this week chewed through 320 million tokens in a single day of one agent running one business. At the old Luna price that day cost real money. At 80% off, always-on agents that plan, retry, and self-correct all night suddenly pencil out.
That's the actual story of frontier pricing in 2026. The headline model gets the benchmarks, but the cheap tier decides which agent products can exist. Twenty cents a million input is the kind of number that turns a clever demo into something you can leave running.
This also isn't happening in a vacuum. Gemini Flash keeps undercutting, Kimi K3 shipped open weights you can self-host, and now OpenAI is defending the floor, not the ceiling. Full pricing at openai.com/index/advancing-the-price-performance-frontier-with-gpt-5-6.
Comments