Correction (2026-09-19): this post describes the model line-up as it stood in July 2026. Vibely's default has since moved — builds now run on DeepSeek V4 Pro by default, with Claude Opus 5 and Sonnet 5, Gemini 3 Flash, GPT-5.6 and DeepSeek V4 Pro all selectable from the composer's model picker. The workflow described below is unchanged; the default model name is not. See pricing for what a turn costs on each.
Claude Opus 4.7 with a 1M-token context window is now the default planning model in Vibely. In our own use it finishes the same tasks in fewer turns than the previous default and recovers from long edit chains more reliably — we have not published a benchmark behind that, so treat it as our impression rather than a measured number.
What this unlocks
- Bigger projects, single conversation. 1M tokens means the agent can keep an entire mid-sized codebase in working memory without compaction artifacts.
- Sharper plans. Plan mode now produces concrete file lists and acceptance criteria instead of generic outlines.
- Better self-healing. When a tool call fails, Opus 4.7 reads the error, re-plans, and converges in one or two retries — rarely the full three-attempt cap.
How to choose your model
You do not have to think about this. Vibely's composer defaults to Auto, which picks the model for the task. If you want full control, the model picker in the composer lets you pin a specific model — useful for long debug sessions where consistency matters more than ceiling.
What did not change
The contract with the agent is the same: one loop, one history, real tools, streaming previews. We tune the model — you keep the workflow.