Which model, and how much effort. Two defaults and a few rules for when to move off them.
From a lightning talk at the HGSC AI meeting, 30 July 2026. Questions: hale@bcm.edu
Two independent decisions, and people freeze because they collapse them into one:
- Where you run — Claude's own endpoints, AWS Bedrock, or our local GPU node. This is a data question, decided by data-classification policy, not by preference. That policy is actively being revised; the working group that owns it reports at the HGSC AI meeting. Go by them, not by this page.
- Which model, and how much effort — a task question. That's what's below.
Two dials, not one:
| Dial | What it controls |
|---|---|
/model |
how capable the model is |
/effort |
how long it's allowed to think |
Turn up effort before you turn up the model. It's the cheaper move and it fixes more than people expect.
/model → Default (Sonnet)
/effort → medium
Use that for weeks without thinking about it.
- it already failed once (don't re-prompt the same setup harder — change something)
- you don't know where the problem is
- you're designing, not editing
- you can't cheaply check whether the answer is right ← the one that matters most for us
If more thinking doesn't do it, then /model opus.
- renames, boilerplate, mechanical edits
- you know the answer, you just don't want to type it
/effort autoresets effort to the default.- Use the aliases (
opus,sonnet,haiku), not version numbers — versions roll. - Available models differ by machine and app version. The HPC install currently offers Default (Sonnet), Opus, and Haiku; laptop/desktop also offer Sonnet explicitly and Fable, and the desktop app can select the previous version of each.
xhighandmaxare not supported on every model; unsupported levels fall back.
Sonnet at medium is your own hands — what you use all day.
Opus, or xhigh effort, is hiring a consultant — sometimes for a genuinely two-minute gig, and that's fine.
The metaphor cuts both ways: you don't bill a consultant to rename a variable, and you also don't stay stuck for an hour to avoid calling one.
Running a model over your own data. Two places:
| Bedrock | Frontier quality, no GPU queue. Your data leaves the node. |
analysis-app1 (local GPU node) |
Nothing leaves the building. You own the quality problem. |
Honestly: just use gemma. Two reasons to move off it — it's code, or it isn't good enough.
| You're doing | Use |
|---|---|
| Text — summarize, extract, classify | gemma4:latest |
| Same, but quality matters more | gemma4:31b |
| Writing or reviewing code | qwen3-coder:latest |
| Thousands of trivial calls, need speed | llama3.2:latest |
| Small/fast coder option | qwen2.5-coder:latest |
Bigger = better and slower. That's the whole sizing rule.
Run ollama list on analysis-app1 to see what's actually loaded rather than trusting this table.
If you don't have a Claude account, or nothing may leave the node, Claude Code will run against local Ollama. It works. It is meaningfully worse — tool use and plan mode get shaky.
Point it at qwen3-coder:latest, not gemma. Driving Claude Code needs tool-use support, so the model recommended above for inference is the wrong one here. Two different jobs, two different answers.
export ANTHROPIC_BASE_URL=http://localhost:11434
export ANTHROPIC_AUTH_TOKEN=ollama
export ANTHROPIC_API_KEY=Needs a tool-use-capable model and ≥32K context.
export CLAUDE_CODE_USE_BEDROCK=1
export AWS_REGION=<region>
export ANTHROPIC_MODEL=<model ID or inference-profile ARN>Per-model overrides also exist: ANTHROPIC_DEFAULT_OPUS_MODEL, ANTHROPIC_DEFAULT_SONNET_MODEL, ANTHROPIC_DEFAULT_HAIKU_MODEL.
- Where am I allowed to run? Policy, not preference.
- Start at the default. Claude Code → Default (Sonnet),
/effort medium. Local inference →gemma4:latest. - Escalate on evidence, not on nerves. Failed once · don't know where to look · designing · can't check the answer → effort first, then model.
- Downshift for mechanical work. Don't bill the consultant for a rename.
Benchmarks and leaderboards, token pricing, parameter counts and quantization, context-window sizes, how reasoning works internally, fine-tuning and RAG, agent frameworks, cross-vendor comparisons, and the data-classification policy itself.
Also skipped: MAX_THINKING_TOKENS and the ultrathink prompt keyword. Both real, both a distraction, and their interaction isn't worth explaining before you've used /effort for a while.