Claude Opus 5.5: the top model for hard work
Introduction
Claude Opus 5.5 is the current top of the Claude line, above Sonnet 5.5 and Haiku 5.5. It wins on large refactors, tricky debugging, API design and careful security review. It costs roughly 40 times more than GPT 6 Luna on input, so never burn it on questions a free OpenCode model answers well. Rule: the free model does drafts, Luna does routine paid work, Opus decides quality on the last 10 percent.
Pricing and ids
| Variant | Console id | Price / 1M in+out |
|---|---|---|
| Opus 5.5 | opencode/claude-opus-5-5 | $4.00 / $20.00, cached read $0.20 |
| Sonnet 5.5, quality per dollar | opencode/claude-sonnet-5-5 | $2.00 / $10.00 |
| Haiku 5.5, cheap | opencode/claude-haiku-5-5 | $0.10 / $0.50 up to 100K |
Endpoint family:
https://opencode.ai/inference/v1/messages
Newer Opus and Sonnet revisions follow the same rule: newest Opus for max quality, newest Sonnet for quality per dollar. Verify prices before budgeting — Sonnet over 200K tokens costs more per 1M.
Where it runs
OpenCode Console serves Opus 5.5 under the id above, with pay-as-you-go credits, and that is the route these guides use. It is also sold directly by Anthropic and by aggregators such as OpenRouter, each under its own id and price, so compare the row for the route you actually use. Prices, rate limits and retention terms differ per route.
Use it in OpenCode
Pin Opus only where it matters — the reviewer agent — while the session default stays cheap, in opencode.jsonc at the project root:
vi opencode.jsonc
{ "$schema": "https://opencode.ai/config.json", "model": "opencode/gpt-6-luna", "agents": { "reviewer": { "description": "Final correctness and security pass", "mode": "subagent", "model": "opencode/claude-opus-5-5" } } }
Review command that always uses Opus in background, file .opencode/commands/review.md:
mkdir -p .opencode/commands
vi .opencode/commands/review.md
--- description: Review staged changes on Opus agent: reviewer model: opencode/claude-opus-5-5 subagent: true --- Review this diff for bugs, security and missing tests: !`git diff --cached`
Skills and agents fit
Opus is the only model to trust with long multi-file skills: policy plus rubric plus negative cases. Its skill is api-review, and only the reviewer agent may load it — see Gating skills per agent. Agent side, it runs reviewer: read-only, steps capped at 8, edit and shell denied, verdict ship, fix first, or block. Keep triggers tight even here — a bigger model makes bad triggers fail more eloquently, not less.
Example workflow
Final pass after Luna builds:
Use the reviewer subagent with @api-review on git diff --cached, verdict ship or fix first
Expected output: findings in severity order with file and line plus a minimal fix each, then one verdict line. Never edits files — the builder applies fixes on GPT 6 Luna and re-submits the diff. Drafts stay on the free OpenCode model.
Privacy and limits
Paid Console requests follow the provider data policy with limited retention for safety. Set monthly workspace and per-member limits, enable auto-reload deliberately, and review opencode stats --cost --models weekly. Session title generation uses cheap models and shows up in history separately.
Troubleshooting
- Too expensive: move drafts to the free model and routine builds to Luna; Opus reviews diffs only, never whole repos.
- Slow verdicts: cap steps at 8, scope to staged diff, deny shell so it cannot wander.
- Contradictory skills: allow only api-review for reviewer; overlapping checklists merge or gate by agent.
Official sources:
Article author: Arthur Isaev