OpenCode's September: GPT-6 support, and fewer timeouts on slow models

OpenCode is an open source coding agent that works with models from many providers. In September it shipped a small release most weeks, from 1.18.26 on September 1st to 1.18.33 on the 28th. Together they get it ready for OpenAI's GPT-6 models, and make slow or busy models fail less often.

Slide text

01 02 GPT-6 arrives

anomalyco/opencode · GitHub release v1.18.26 to GitHub release v1.18.33

OpenCode's September: GPT-6 support, and fewer timeouts

Versions 1.18.26 to 1.18.33 September 1 to 28, 2026

The first fix was simple. If you sign in to OpenCode with a ChatGPT plan instead of an API key, GPT-6 Astra didn't appear in your model list, because the filter didn't expect a whole number version like six. Since 1.18.29, on September 4th, it's there.

Then, on September 9th, GPT-6 models got their own system prompt. That's the standing instructions OpenCode gives a model before your request, like how to work and how to use the shell. Different model families get different ones. Any GPT-6 model, Astra, Sol or Luna, now gets one written for it.

OpenCode also sells access to models it has tested, and both of its services picked up new ones. Both are optional. Zen, which you pay for per request, now lists GPT-6 Sol, GPT-6 Luna and Claude Opus 5.5. And Go, a monthly plan from ten dollars that's built around open models, added Grok 4.7.

Now the reliability work.

The change most people will feel is a timeout. A model that takes a long time to start, like a big one on a busy provider, could hit OpenCode's limit and fail. Since September 2nd, OpenCode waits up to five minutes for a reply to begin, and five minutes between pieces of a streamed answer. You can switch the streaming limit off. And today's release makes models reached through Cloudflare's AI Gateway respect those limits too.

Several fixes are about specific model providers. Claude 5 sessions no longer fail after you change the prompt or the tools partway through. Gemini's thinking settings now match what each generation of the model actually supports. And on Amazon Bedrock, reasoning and replay are more reliable.

And a few smaller ones. Printing OpenCode's configuration for debugging now hides your keys and sensitive headers. If you run OpenCode inside an editor, a resumed session gets its model, effort and mode back. And if a shared remote configuration can't sign in, OpenCode now shows the error at startup and exits.

These were small releases: model support and fixes, with no new commands to learn. Every note is on OpenCode's releases page on GitHub. That's OpenCode's September: GPT-6 support, and fewer timeouts.