Seven New Frontier Models
Samuel Su
on August 13, 2026Three model families shipped new flagships within a few weeks of each other. All seven are now selectable on your agents and through the API, while the GPT-5.6 models are also available in workflow AI nodes. The short version: the frontier moved, the floor got cheaper, and the context window got much larger.
What's new
| Model | Context window | Credits / message |
|---|---|---|
| GPT 5.6 Sol | 1M tokens | 8 |
| GPT 5.6 Terra | 1M tokens | 3 |
| GPT 5.6 Luna | 1M tokens | 0.5 |
| Claude Opus 5 | 1M tokens | 8 |
| Claude Sonnet 5 | 1M tokens | 4 |
| Gemini 3.7 Flash | 1M tokens | 3 |
| Gemini 3.5 Flash Lite | 1M tokens | 0.5 |
GPT-5.6 arrives as three models rather than one. Sol is the flagship, for the hardest reasoning and agentic work. Terra sits in the middle and is the one most agents should be on — a full generation newer than GPT 5.4, at three credits per message. Luna is the volume model: half a credit per message, with the same million-token window as its siblings.
Claude Opus 5 is Anthropic's flagship for complex agentic work. Claude Sonnet 5 is the more interesting one for most people — a newer generation at the same four credits per message as Claude 4.6 Sonnet, so moving an agent across costs you nothing.
Gemini 3.7 Flash brings frontier reasoning at Flash speed. Gemini 3.5 Flash Lite is built for high-volume, latency-sensitive work at half a credit per message.
A million tokens, everywhere
Every model in this release has a context window of at least a million tokens. That is roughly a few thousand pages of documentation in a single conversation.
In practice it means better continuity for the people chatting with your agent: fewer repeated questions, smoother follow-ups, and more room to use the relevant conversation and knowledge context we retrieve for each answer. You still get the same retrieval and summarization safeguards, but larger-context models make long support conversations feel less fragmented.
Picking one
If you want a starting point rather than a comparison:
- Most agents → GPT 5.6 Terra at three credits, or Claude Sonnet 5 at four. Both are current-generation; Claude Sonnet 5 is a drop-in replacement for Claude 4.6 Sonnet at the same price.
- High volume, simple tasks → GPT 5.6 Luna or Gemini 3.5 Flash Lite at half a credit per message.
- Hardest reasoning → GPT 5.6 Sol or Claude Opus 5 at eight credits. Worth it when the answer matters more than the cost per message; usually not worth it for routine support replies.
Credits per message are shown next to every model in the picker, so you can see what a choice costs before you make it.
Switching
Open your agent, go to the model dropdown, pick one. The change takes effect on the next message — no retraining, no reindexing, nothing to migrate. If it isn't what you wanted, switch back the same way.
The new models are available on plans with advanced-model access.