Nio v0.1.252

GPT-5.1 now reasons while it works in Nio. On our calculator benchmark it
went from an average of 70 to 98 through Nio, above its direct calls (91).

What changed

GPT-5.1 reasons in agent tasks. GPT-5.1 is the first OpenAI model whose
API default is no reasoning, and Nio didn't change it. Without reasoning it
often answered with a plan instead of making the changes. Nio now asks
GPT-5.1, and later GPT-5.x point releases, for medium reasoning in each
step of a task, the same as for Claude models, and low for small tasks.
GPT-5, GPT-5 mini and the o-series already reason by default and are
unchanged. Set NIO_OPENAI_AGENT_REASONING_EFFORT (none, low, medium
or high) to choose a level yourself.
More time for each reasoning reply. Those steps can take longer, so
each one now gets up to 6 minutes, as Claude replies do, instead of 2.
Before, slow replies were dropped and retried. Set
NIO_OPENAI_REASONING_TIMEOUT_MS to change the limit.
Cost. Reasoning is billed as output, so a GPT-5.1 task costs about twice
as much as before ($1.79 instead of $0.87 per run on the benchmark).
Benchmark on nioai.run. The GPT-5.1 row on the landing page shows the
new results.
October 21 results.

Windows PowerShell:

powershell -c "irm https://github.com/luisMan/Nio/releases/download/v0.1.252/nio-installer.ps1 | iex"

macOS and Linux:

curl --proto '=https' --tlsv1.2 -LsSf https://github.com/luisMan/Nio/releases/download/v0.1.252/nio-installer.sh | sh

If you have v0.1.195 or later, Nio offers this update when it starts, or
run nio update.

Each native archive and the MSI include a SHA-256 sidecar. Publishing this
release does not reset account token limits.