OpenAI Launches GPT-5.6: Three Models Named Sol, Terra, and Luna Replace a Single Flagship

by Geth Author

OpenAI released GPT-5.6 on July 9, 2026 — not as a single model, but as a three-tier family named Sol, Terra, and Luna. The move mirrors how cloud infrastructure has long worked: a flagship tier for demanding workloads, a mid-tier for everyday use, and a fast, cheap tier for high-volume pipelines. Applied to large language models, it's a meaningful structural shift.

Three Models, One Release

According to OpenAI's official announcement, the three models are:

  • Sol — OpenAI's new flagship. State-of-the-art results in coding, knowledge work, cybersecurity, and science. Runs at up to 750 tokens per second on Cerebras hardware. Priced at $5 input / $30 output per million tokens.
  • Terra — The balanced option. Positioned for everyday professional work where Sol's full capability isn't required. Priced at $2.50 / $15 per million tokens.
  • Luna — The speed and cost tier. OpenAI's fastest and cheapest model for high-throughput use cases. Priced at $1 / $6 per million tokens.

All three became publicly available on July 9th. Sol replaced GPT-5's preview as the default model in ChatGPT for subscribers.

What's New Under the Hood

GPT-5.6 introduces programmatic tool calling in the Responses API — structured, deterministic function invocation rather than the model deciding when and how to call tools based on prompt phrasing alone. This addresses one of the more persistent pain points for developers building agentic workflows, where tool invocation reliability has been inconsistent across prompt variations.

Prompt caching gets a meaningful upgrade: explicit cache breakpoints let developers pin which prefix of a prompt should be cached, paired with a 30-minute minimum cache life that makes long-context workload economics more predictable. Previously, cache hits were best-effort and sometimes required submitting the same prompt twice before activation.

Sol ships with what OpenAI describes as its most robust safety stack to date — strengthened guardrails specifically for high-risk activity, sensitive cybersecurity requests, and patterns of repeated misuse.

ChatGPT Work

Alongside GPT-5.6, OpenAI released ChatGPT Work — an agent built on Sol designed to carry out whole jobs autonomously. Early demos show it handling multi-step workflows: researching a topic, drafting a document, formatting it for a target output, and sending it — chaining tools without prompting between steps. This is a consumer-facing product aimed at knowledge workers rather than an API feature, and it marks OpenAI's most direct push into the agentic productivity space to date.

The Bigger Picture

The Sol/Terra/Luna naming is more than marketing. It reflects a deliberate move away from pretending one model fits all workloads — and positions OpenAI to compete more directly with the tiered pricing structures of Anthropic and Google. Whether Terra carves out a genuine niche or whether users simply default to Sol and Luna on either end will be the real test of the strategy over the coming months.