OpenAI Unveils GPT-6 Astra for Long-Horizon Agents

If you are picking a model for serious agent workloads, OpenAI just laid its cards on the table. The company officially unveiled GPT-6 Astra, a model built for long-horizon agentic work. It targets autonomous multi-agent pipelines, not casual chat.

What Happened

GPT-6 Astra is designed from the ground up for persistent reasoning, tool use, and autonomous multi-step execution. OpenAI is clearly aiming it at developers running agents that need to keep working without handing control back.

Per 1M tokens, pricing is:

• Input: $10.00
• Cached input: $1.00
• Cache writes: $12.50
• Output: $50.00

The $1.00 cached input rate is a 90% discount over fresh input. That is a strong hint that OpenAI wants developers to build stateful agents that reuse large system prompts and memory banks instead of resending full context every call.

Why It Matters

This puts Astra in the heavy frontier tier. The numbers only make sense for serious production agents that keep long, cached contexts alive. Quick single-turn chat completions are better served by leaner, cheaper models. If your agents rely on a deep memory bank, the cached input rate is what keeps the bill reasonable.