TL;DR — the one rule
Start with Sonnet. Upgrade to Opus when the stakes warrant the depth. Drop to Haiku when speed matters more than judgment. That's the whole picker. The rest of this page tells you when each rule applies.
01The Three Models
Claude isn't one tool. It's a family of three, designed for three different jobs. The names sort by speed and depth — not by quality. A Haiku answer to a Haiku-shaped question is not a worse answer than an Opus answer. It's the right answer.
Opus 4.6
The careful one
Deep multi-step reasoning. The slowest of the three. The most expensive. Reserved for work where being wrong is costly — board memos, funder LOIs, complex analyses, anything where the reasoning trail matters.
Use for
- Drafting a Gates LOI or large federal proposal
- Board-facing strategy memos with multiple competing arguments
- Cross-document audits where logic has to hold
- The Adversarial Audit Skill on a high-stakes draft
- Anything signed by Preston or Eric that names a funder
Don't use for
- Routine emails to counselors or families
- Summarizing a single document
- Classifying or tagging items
- "Just because I want the best one"
Sonnet 4.6
The DC CAP default
Strong reasoning. Fast enough for everyday work. The model that should run most of your Projects, most of your Skills, and most of your one-off chats. Handles roughly 80% of DC CAP work well.
Use for
- Counselor and partner emails
- Quarterly partnership updates
- Briefing decks and weekly status writeups
- First drafts of grant narratives
- Voice scans on board memos (with thinking on)
- Most Projects and most Skills, by default
Don't use for
- The single biggest funder proposal of the year (upgrade to Opus)
- A 10-second classification task (downgrade to Haiku)
Haiku 4.5
The fast one
Built for speed and volume. The right tool when you have many small tasks and the cost of being slightly less polished is low. The wrong tool for anything that requires judgment, voice, or multi-step reasoning.
Use for
- Summarizing a long PDF down to bullet points
- Classifying a list of items (e.g., tagging emails by topic)
- Quick lookups and definitions
- Light copyediting where voice is already in place
- Pulling structured data out of unstructured text
Don't use for
- Anything signed by a person
- Drafting in DC CAP voice from scratch
- Multi-step analysis
- Verification of facts (it's faster, not more careful)
02How to Decide in 30 Seconds
Five questions, asked in order. The first "yes" gives you your model. If you reach the end without a yes, pick Sonnet.
Is this a quick lookup, classification, or summary?
If yes → Haiku 4.5. Examples: tag these 50 emails by topic. Pull every dollar figure out of this PDF. Summarize this 12-page brief in five bullets.
Is this a board-, funder-, or executive-facing artifact where being wrong is costly?
If yes → Opus 4.6. Examples: the Gates LOI. The Q3 strategy memo. The capstone audit. The board deck narrative.
Does this require multi-step reasoning across multiple documents or arguments?
If yes → Opus 4.6. Examples: comparing three peer organizations' theories of change. Reconciling conflicting data points across reports. Building a strategic argument that has to hold under cross-examination.
Is this everyday work — emails, drafts, briefings, partner comms, voice-aware writing?
If yes → Sonnet 4.6. This is most of what you do. Sonnet is the right answer roughly 80% of the time.
Still not sure?
Pick Sonnet. Then watch the answer. If it falls short on depth, upgrade to Opus and rerun. If it's overkill for the task, drop to Haiku next time. The picker improves with reps.
03Side-by-Side Tradeoffs
The same dimensions across all three models. Read the row, not the column — that's how you spot the right tradeoff for the task in front of you.
| Dimension | Opus 4.6 | Sonnet 4.6 | Haiku 4.5 |
|---|---|---|---|
| Strength | Deepest reasoning, longest planning horizon | Strong reasoning at everyday speed | Fastest response, lowest cost per task |
| Speed | Slowest | Balanced | Fastest |
| Voice fidelity | Highest with rich context | Strong with a Project loaded | Weak; not a voice model |
| Multi-step reasoning | Excellent | Good | Limited — use only for one-step tasks |
| Best with thinking on | Yes — that's the whole point | Yes, for high-stakes work | Rarely worth it |
| Cost relative to Sonnet | Notably higher | 1× (baseline) | Notably lower |
| When you're wrong | You over-spent on a simple task | You probably got it right | You under-served a task that needed judgment |
04Two Anti-Patterns to Avoid
Anti-pattern 1 — Defaulting to Opus by reflex
"I want the smartest one." Opus is slower and more expensive than Sonnet. Running every chat on Opus burns nonprofit-tier compute on tasks Sonnet would handle just as well. The cost shows up in two ways: slower responses for your team, and faster monthly burn rate on the org's compute budget.
The fix: Sonnet first. Upgrade only when you can name why the stakes earn the depth.
Anti-pattern 2 — Defaulting to Haiku because it's fast
Speed is the wrong variable when the work needs judgment. A Haiku-drafted donor letter or counselor email reads sloppy because Haiku isn't built for voice or nuance. Saving 20 seconds isn't worth shipping work that looks like AI-flavored slop.
The fix: Haiku only when the task is genuinely small and the output is genuinely structured (a list, a tag, a summary). For anything a human will read carefully, use Sonnet.
05Model Selection in Projects and Skills
Every Project has a default model. Every Skill runs under whatever model the chat is using when it triggers. As your team builds Projects and Skills this pilot, you're picking models for everyone who runs them — pick on purpose.
Quick guide
- Most Projects → Sonnet. The Project loads rich context, so Sonnet has what it needs to perform well. Upgrade the Project to Opus only if the work running through it is consistently high-stakes (a board-brief Project; an audit Project).
- Audit and verification Skills → Opus with thinking on. When the Skill exists to catch errors, depth and reasoning quality matter more than speed.
- Drafting and voice Skills → Sonnet. Voice rules in the SKILL.md give Sonnet enough scaffolding to perform at near-Opus quality on a fraction of the cost.
- Classification and summary Skills → Haiku. If the Skill exists to pull structure out of text, Haiku does it faster and cheaper than Sonnet.