[FEATURE] Per-message model override without switching session model

Preflight Checklist

• [x] I have searched existing requests and this feature hasn't been requested yet
• [x] This is a single feature request (not multiple features)

Problem Statement

hen working in an Opus session, I frequently want to send a single low-complexity message using Haiku (e.g., "what's the jq command to extract all .name fields from an array?") without switching my session model. Currently, /model changes the model for the entire session, so I'd need to /model haiku → ask → /model opus every time. This creates friction and risks forgetting to switch back, accidentally burning Opus tokens on a simple question or vice versa — running a complex task on Haiku after forgetting to switch back.

The workaround is running claude -p --model haiku in a separate terminal, but that loses all session context and breaks the flow of working in a single pane.

Proposed Solution

A per-message model override that applies to a single response and automatically reverts. A few possible syntax approaches:
Option A — Inline flag (mirrors CLI conventions):
--model haiku what's the sed command to replace in-place on macOS?
Option B — Dedicated slash command:
/ask haiku what's the sed command to replace in-place on macOS?
Option C — Model tag prefix:
@haiku what's the sed command to replace in-place on macOS?
In all cases: the override applies only to that single message. The session model remains unchanged for all subsequent messages.

Alternative Solutions

Alternative Solutions:

Toggle twice: /model haiku → ask → /model opus. Works but adds friction and risks forgetting to switch back, which can mean running a complex task on Haiku or burning Opus tokens on trivial follow-ups.
Print mode in a second terminal: claude -p --model haiku "..." works but loses session context and requires switching panes.
Subagents with model frontmatter: Can define a Haiku-based agent, but this is designed for recurring task types, not ad-hoc one-off questions.

Priority

Low - Nice to have

Feature Category

Configuration and settings

Use Case Example

Use Case Example:

I'm in an Opus session working through a complex peer review allocation engine refactor across multiple files
I need to remember the exact sed in-place flag syntax on macOS (the -i '' quirk)
I type: --model haiku what's the sed command to replace in-place on macOS?
Haiku responds instantly with the answer
My next message automatically goes back to Opus and I continue the refactor

No model switching, no second terminal, no wasted tokens.

Additional Context

Related issues that show broader demand for more granular model control within sessions:

#20745 — Model setting changes globally across all sessions (regression)
#4937 — Add model selection support for custom commands
#25546 — Allow configuring model for built-in agents
#20392 — No model switching in web interface

The existing opusplan hybrid mode shows Anthropic already recognizes the value of multi-model workflows within a single session. This feature extends that philosophy to user-controlled, per-message granularity.