Why a demo gallery has a casting page
A prompt that worked in July can look broken in September for a boring reason: the model named in the write-up is no longer the one your reader has. Naming the cast makes a demo reproducible, and it keeps the gallery honest about the one thing that changes fastest.
This page is not a model-selection guide. It says which model to run a demo on and nothing more. For the general question of which Claude model to use, start at Claude Central. For scored comparisons, pricing analysis, and methodology, read Claude Reports.
The company
The house lead
Claude Opus 5
claude-opus-5
1M-token context, 128K max output, adaptive thinking
Default on Claude Max, strongest model on Claude Pro
Cast it for
- Any demo where you want one model to carry the whole show
- Artifact prototypes, vision audits, and data reveals in front of a live room
- Demos run on a Pro or Max seat with no API access
Wrong cast when: The task runs for hours without supervision, or the reveal depends on sustained multi-step tool use. Cast Fable 5.1 there instead.
Source: Claude Opus 5 overview
The long-horizon headliner
Claude Fable 5.1
claude-fable-5-1
1M-token context, 128K max output, always-on adaptive thinking
Generally available since September 1, 2026
Cast it for
- Long-horizon agent runs that plan, execute, and check their own work
- Demanding reasoning demos where Opus 5 at higher effort still falls short
- Agentic loops that reread the same context repeatedly
Wrong cast when: The demo is a 10-minute artifact reveal. The extra deliberation is invisible to the audience and slower on stage.
Source: Claude Fable 5.1 overview
The understudy who never misses a cue
Claude Sonnet 5
claude-sonnet-5
1M-token context, 128K max output
Speed and intelligence balance
Cast it for
- Rehearsing a prompt a dozen times before the real run
- Demos where response latency is part of the experience
- Long-context reads that do not need frontier reasoning
Wrong cast when: The reveal is the reasoning itself, such as a debugging loop or a multi-hour agent run.
Source: What's new in Claude Sonnet 5
The stagehand
Claude Haiku 4.5
claude-haiku-4-5
200K-token context, 64K max output
Fastest and cheapest in the lineup
Cast it for
- Dry runs and prompt-shape checks before you spend a frontier call
- Cheap classification or extraction steps inside a larger demo
Wrong cast when: The demo needs a 1M-token context window or frontier reasoning. Check the model overview for its current retirement date before building a demo on it.
Source: Claude model overview
What is cast where
Every demo page carries its own casting note. Most of them say Opus 5, because most demos are about prompt structure rather than model tier.
- Artifact Prototype — Opus 5
- PDF Evidence Table — Opus 5
- Vision UI Audit — Opus 5
- Data Visualizer — Opus 5
- Codebase Debugging — Opus 5 or Fable 5.1
- Browser Chore — Opus 5
- Research Brief — Opus 5
- Long Context — Fable 5.1 or Opus 5
- Long-Horizon Agent — Fable 5.1
- Effort Dial — Fable 5.1 and Opus 5
Two demos where the model is the point
In the long-horizon agent run, Fable 5.1 is not a nicer option, it is the subject: Anthropic positions it for exactly that kind of sustained agentic work. Inthe effort dial, the variable is the thinking-effort setting rather than the model name, and the demo only works because adaptive thinking is always on with different defaults per surface.
Everywhere else, swapping the model changes how fast the demo runs and how much it costs, not whether the reveal lands. That is worth saying plainly, because "use the biggest model" is the most common way a demo gets more expensive without getting better.
When to recheck this page
Recheck it whenever Anthropic ships a model, retires one, or changes a default effort level. The model overview lists current API IDs and retirement dates, and retirement is the reason a demo script eventually stops working rather than merely looking dated. Changes get logged on the freshness page.