Claude Sonnet 5.5 vs Opus 5.5: Is “Half the Price” Real Once You Count Cache and Output? (2026)
Short answer: Claude Sonnet 5.5 lists at half the price of Opus 5.5 on the Claude API: $2 in and $10 out per million tokens, versus $4 and $20. Cached input costs the same $0.20 on both, though. So on agent work with a big cached context, Sonnet’s real saving shrinks to about 40%. It stays cheaper unless it needs more than about 2.3x Opus’s output tokens for the same job. Anthropic says that gap closes at high effort settings.
“Half the price” is the headline every launch recap ran with on September 28. It’s true on the rate card. It stops being true the moment you look at what a real request is made of — how much is cached, how much the model writes, and which effort setting you left it on.
This is the per-task math for three common workloads, using Anthropic’s published prices. Plus the one line in the pricing docs that changes the comparison more than any benchmark does.
In This Comparison
- What Do Claude Sonnet 5.5 and Opus 5.5 Cost Per Million Tokens?
- Why Does Prompt Caching Shrink the Price Gap?
- How Much Does One Real Task Cost on Each Model?
- When Does Sonnet 5.5 Stop Being Cheaper?
- Where Does Opus 5.5 Still Earn Its Price?
- A Simple Routing Rule for Mixed Workloads
- Tips
- Warnings
- FAQ
- Sources
What Do Claude Sonnet 5.5 and Opus 5.5 Cost Per Million Tokens?
Claude Sonnet 5.5, released September 28, 2026, costs $2 per million input tokens and $10 per million output tokens on the Claude API, while Claude Opus 5.5 costs $4 and $20, according to Anthropic’s pricing documentation as of September 30, 2026. Both get a 50% discount through the Batch API.
Sonnet 5.5 is also priced the same as Sonnet 5. If you’re still on Sonnet 5, moving up costs nothing extra. Anthropic’s launch page says the new model runs 30%+ faster and costs up to 30% less for most work, because it needs fewer tokens per task.
| Price per million tokens (Claude API) | Sonnet 5.5 | Opus 5.5 | Opus premium |
|---|---|---|---|
| Base input | $2 | $4 | 2x |
| Output | $10 | $20 | 2x |
| 5-minute cache write | $2.50 | $5 | 2x |
| Cache hit (read) | $0.20 | $0.20 | 1x — same price |
| Batch input / output | $1 / $5 | $2 / $10 | 2x |
Table: Claude Sonnet 5.5 vs Opus 5.5 list prices from platform.claude.com pricing docs, checked September 30, 2026.
Why Does Prompt Caching Shrink the Price Gap?
Prompt cache hits on Claude Opus 5.5 are billed at 0.05x the base input price — $0.20 per million tokens — while Sonnet 5.5 uses the standard 0.1x multiplier, which also comes to $0.20, so cached context costs exactly the same on both models. That’s in a footnote on Anthropic’s pricing page. It’s easy to miss.
When I checked the table line by line, I found that this is the only row where Opus isn’t twice the price. It matters a lot for coding agents and document chat. Those workloads resend the same codebase or file on every turn, and caching turns most of the input into cache reads.
In practice, the more of your input is cached, the closer Opus gets to Sonnet on input cost. Output is still 2x.
How Much Does One Real Task Cost on Each Model?
A single coding-agent turn with 100,000 cached tokens of context, 5,000 new input tokens, and 3,000 output tokens costs about $0.06 on Claude Sonnet 5.5 and about $0.10 on Claude Opus 5.5 at September 2026 API prices. That’s a 40% saving, not 50%.
Here are three workloads side by side. I kept token counts identical across models so the table isolates price. Real output lengths will differ — more on that in the next section.
| Workload (per request) | Sonnet 5.5 | Opus 5.5 | Real saving |
|---|---|---|---|
| Document Q&A: 20K uncached in, 2K out | $0.060 | $0.120 | 50% |
| Agent turn: 100K cached + 5K new in, 3K out | $0.060 | $0.100 | 40% |
| Batch summary: 50K in, 1K out (Batch API) | $0.055 | $0.110 | 50% |
Table: calculated per-request cost at Anthropic list prices (September 30, 2026), assuming identical token counts on both models. Scale: 1,000 agent turns ≈ $60 on Sonnet 5.5 vs $100 on Opus 5.5.
When Does Sonnet 5.5 Stop Being Cheaper?
Claude Sonnet 5.5 stops being cheaper than Opus 5.5 on a cache-heavy agent turn once it produces more than about 2.3 times as many output tokens as Opus for the same job; on uncached document Q&A the break-even is about 4 times. Those are my own calculations from the list prices, not Anthropic figures.
Why would Sonnet write more? Effort. Anthropic’s launch page says Sonnet 5.5 “complements Opus 5.5 best when running at lower effort settings, where it costs less per task. At higher settings, it can perform comparably at a similar cost.”
Read that twice. Anthropic is telling you the discount mostly disappears at high effort.
The effort default also differs by surface. According to the launch page, Medium is the default in the Claude apps and High is the default on the Claude Platform. So an API integration you never tuned is already running at the setting where the savings are thinnest.
| Workload | Sonnet stays cheaper until its output is… |
|---|---|
| Agent turn (cache-heavy) | ~2.3x Opus’s output |
| Document Q&A (uncached) | ~4x Opus’s output |
| Batch summary (input-heavy) | ~12x Opus’s output |
Where Does Opus 5.5 Still Earn Its Price?
Claude Opus 5.5 outscores Sonnet 5.5 on seven of the eight benchmarks in Anthropic’s September 28, 2026 comparison table. The biggest gap is FrontierCode 1.1 (54.4% vs 46.2%). Humanity’s Last Exam is next (67.7% vs 64.5%). Sonnet 5.5 wins only Terminal-Bench 4.0, at 70.6% vs 66.4%.
When I compared the rest of the table, most gaps were small. OSWorld 2.1 is 81.8% vs 80.1%. CursorBench 4.0 is 57.8% vs 55.5%. GDPval-AA v2.1 is practically a tie at 1846 vs 1844.
Anthropic’s own guidance matches the numbers. It points Opus at “complex work requiring careful judgment,” and Sonnet at “well-scoped everyday tasks, fixing bugs, and creating polished documents.”
Honestly, that’s the most useful sentence in the whole launch.
A Simple Routing Rule for Mixed Workloads
- ✅ Default to Sonnet 5.5 at low or medium effort for bug fixes, drafts, formatting, extraction, and summaries.
- ✅ Send to Opus 5.5 for architecture decisions, hard multi-file refactors, and anything where a wrong answer costs more than the tokens.
- ✅ Turn on prompt caching either way. It’s the one discount both models share at the same rate.
- ✅ Move non-urgent jobs to the Batch API for 50% off on both.
- ✅ Log output tokens per task for a week on both models before you commit, and compare against the break-even table above.
If you’ve ever swapped models to save money and watched the bill barely move, output length is almost always why.
Tips
- Set the effort level explicitly in API calls instead of relying on the platform default of High.
- Check your usage dashboard’s output-token column, not just requests. That’s where Sonnet’s savings leak away.
- If you use a model alias, confirm which version it resolves to after a new release.
Warnings
- Benchmark scores come from Anthropic’s own launch table. Independent results may differ, and your workload isn’t a benchmark.
- US-only inference (
inference_geo: "us") adds a 1.1x multiplier to all token prices on both models, per Anthropic’s docs. - Prices were checked September 30, 2026. Confirm on the live pricing page before budgeting.
FAQ
Is Claude Sonnet 5.5 available to free Claude.ai users?
Anthropic says anyone can chat with Claude using Sonnet 5.5 on Claude.ai across web, iOS, and Android. Usage limits still depend on your plan.
Does Sonnet 5.5 use a different tokenizer than Opus 5.5?
No. Anthropic’s pricing docs say Claude 4.7 and later models share a newer tokenizer, so the same text produces roughly the same input token count on both models.
Is Haiku 5.5 a cheaper option?
Not yet. Anthropic says Claude Haiku 5.5 will join the 5.5 family in the coming weeks, but it had no published price as of September 30, 2026.
Sources
- Anthropic — Introducing Claude Sonnet 5.5 (September 28, 2026)
- Claude Platform Docs — Pricing
- Claude Platform Docs — Claude Sonnet 5.5 overview
🔗 Related guides
Fact-checked based on Anthropic’s Claude Sonnet 5.5 launch page and Claude Platform pricing documentation as of September 30, 2026. Per-task costs are calculations at list price, not measured bills.
