Sonnet 5.5 vs Opus 5.5: the cost gap, and when Opus is worth it
One week apart, two models, exactly 2× apart on price. The interesting question is not which is better — it is which of your requests deserves the expensive one.
Short version: Sonnet 5.5 is the default. Route a request to Opus 5.5 when the cost of getting it wrong exceeds the token bill — and route it back down as soon as you can describe the failure mode.
The headline numbers
| Claude Sonnet 5.5 | Claude Opus 5.5 | |
|---|---|---|
| Released | 28 Sep 2026 | 22 Sep 2026 |
| Input / 1M tokens | $2 | $4 |
| Output / 1M tokens | $10 | $20 |
| Cached input read | $0.20 | $0.20 |
| Context window | 200K | 200K |
| Effort levels | low → max | low → max |
| Relative cost | 1× | 2× |
Confirm against the vendor pricing page before budgeting — this market re-prices monthly.
What 2× actually means on a bill
A 2× headline looks harmless until you multiply it by volume. Take a modest workload — 4,000 input tokens and 900 output tokens per call, 2,000 calls a day:
| Workload | Sonnet 5.5 / month | Opus 5.5 / month | Difference |
|---|---|---|---|
At low effort | $1,020 | $2,040 | +$1,020 |
At high effort | $4,800 | $9,600 | +$4,800 |
At xhigh effort | $11,280 | $22,560 | +$11,280 |
Note what dominates: moving from low to xhigh on the
same model multiplies the bill by about 11×. Switching models only doubles it. If you are
trying to cut spend, effort is the bigger lever — and it is the one teams forget they have.
Run the numbers with your own token counts on the calculator before you treat any of the figures above as your budget.
A routing rule you can actually implement
- Default everything to Sonnet 5.5 at
high. Measure quality on a real sample. - Push the easy half down. Every request whose output is a label, a date or
fixed-schema JSON goes to
low. This is usually where the biggest saving hides, because easy work is the majority of volume. - Escalate by exception. Route to Opus 5.5 when a cheaper attempt fails validation, or when the task is on a known-hard list (architecture, incident triage, multi-file migration).
- Cap the escalation. Set a per-day ceiling on Opus calls. Escalation without a ceiling is how a good policy turns into an unbounded bill.
When the answer is genuinely Opus
- The output is a decision someone will act on, not a draft someone will edit.
- The failure is silent — a wrong answer that looks right is far worse than an error.
- The task is long-horizon: many files, many steps, a plan that has to stay coherent.
- You are generating training data, evaluation sets or published material.
When Sonnet 5.5 is simply the right answer
- High-volume, user-facing traffic where latency is part of the experience.
- Extraction, classification, translation, formatting — anything deterministic.
- Anything you can verify automatically. Verification beats model strength.