Claude Sonnet 5.5 is Anthropic’s second model in the Claude 5.5 family, announced on September 28, 2026. Its standard global Claude API price is $2 per million input tokens and $10 per million output tokens, with flat pricing across its 1M context window. If you searched for sonnet 5.5 vs haiku 5.5 or sonnet 5.5 bedrock, the key distinctions are Haiku’s lower, tiered API prices and AWS’s independently set Bedrock pricing.
The dated evidence gives three different views: Anthropic’s specifications and evaluations, Arena’s human preference ratings, and OpenRouter’s token usage. They answer different questions. The figures below use pricing documentation dated October 11, 2026, Arena data published October 8, and OpenRouter’s October 4–10 weekly window.
Sonnet 5.5 Pricing: Standard, Cache, and Batch Rates
The Claude API pricing page lists these global rates. All prices are USD per million tokens; cache durations describe the cache-write option.
| Token or processing category | Sonnet 5.5 price |
|---|---|
| Standard input | $2 |
| Standard output | $10 |
| 5-minute cache write | $2.50 |
| 1-hour cache write | $4 |
| Cache read | $0.10 |
| Batch API input | $1 |
| Batch API output | $5 |
There is a source discrepancy worth resolving before estimating costs: the launch announcement listed cache reads at $0.20, while the current pricing page lists $0.10, or 0.05x the standard input rate. This article uses the current pricing page. US-only inference has a 1.1x multiplier; the table describes global pricing.
Sonnet 5.5’s input and output rates remain flat across the 1M context window. For a budget estimate, distinguish ordinary input, cache writes, cache reads, and output instead of applying the standard input price to every token. Batch pricing is a separate processing option, so use its rates when estimating Batch API work.
The model overview specifies a 128K maximum output, with 300K available on the Batch API using a beta header. That output allowance and the Batch discount are separate facts: a larger allowed response does not mean a smaller total bill.
Sonnet 5.5 Rankings on Arena and OpenRouter
Arena measures human preference within a category, with style control in the overall text results cited here. The listed entry is claude-sonnet-5.5-xhigh, an xhigh effort variant. It should not be read as a rating for every possible Sonnet 5.5 configuration, including its default high effort setting.
| Arena category, October 8, 2026 | Rank | Rating | Confidence interval | Votes |
|---|---|---|---|---|
| Overall text, style control | 32 | 1475.8 | 1467.0–1484.6, 95% CI | 4,838 |
| Coding | 18 | 1531.8 | 1514.6–1549.0 | 1,176 |
→ Track it on AI Rank: Claude Sonnet 5.5 — Arena rating, rank and 90-day history
Vote counts are still low, leaving wide intervals. In overall text, GPT-6 Astra Max is ranked 35 with a rating of 1475.1, close to Sonnet’s 1475.8; that small point-estimate difference is insufficient grounds for a confident preference claim. Coding places Sonnet 5.5 above Sonnet 5 high, which is ranked 36 at 1519.1, but these entries also use different effort settings. Haiku 5.5 is not on Arena yet, so there is no Haiku 5.5 Arena rating to compare.
OpenRouter weekly text usage for October 4–10 places Sonnet 5.5 at rank 15 with 1.69T tokens, up 216% week over week. Haiku 5.5 is rank 23 with 818B tokens and marked new; Sonnet 5 is rank 28 with 679B tokens, down 17%. These are token volumes on OpenRouter, not quality scores or measurements of the entire market.
→ Live usage: See where it ranks this week in the OpenRouter model usage ranking on AI Rank (tokens, Oct 4–10, 2026 window).
Sonnet 5.5 vs Haiku 5.5: Which One Should You Use?
Price is the clearest documented difference. Haiku 5.5 has two prompt-length tiers; Sonnet 5.5 has flat rates across its context window. The table shows standard Claude API rates in USD per million tokens.
| Model and prompt length | Input | Output |
|---|---|---|
| Sonnet 5.5, across 1M context | $2 | $10 |
| Haiku 5.5, up to 100K tokens | $0.10 | $0.50 |
| Haiku 5.5, above 100K tokens | $0.50 | $2.50 |
Anthropic labels Sonnet 5.5 latency “Fast” and Haiku 5.5 “Fastest.” Their default effort settings are high and medium, respectively. Both have 1M context and 128K maximum output in the comparison table. Haiku is positioned for high-volume, cost-sensitive applications; Sonnet is the candidate to evaluate when its reported capabilities justify the higher token rates for your workload.
For a cost-sensitive application, start the comparison with Haiku’s applicable prompt tier. For work resembling the tasks in Sonnet’s published evaluations, consider Sonnet, then assess whether it meets your own acceptance criteria. The evidence does not establish a universal quality winner between these two models: Haiku has no Arena entry, and higher OpenRouter usage cannot settle that question.
For more detail, see AI Rank’s Haiku 5.5 API pricing guide and Haiku 5.5 benchmark article.
Sonnet 5.5 Bedrock: Model IDs and Pricing Notes
The AWS model card documents the following identifiers:
| Access route | Identifier |
|---|---|
| Claude API | claude-sonnet-5-5 |
| Amazon Bedrock model | anthropic.claude-sonnet-5-5 |
| Bedrock global inference profile | global.anthropic.claude-sonnet-5-5 |
| Bedrock US geographic profile | us.anthropic.claude-sonnet-5-5 |
| Bedrock EU geographic profile | eu.anthropic.claude-sonnet-5-5 |
Sonnet 5.5 has been available on Bedrock and Claude Platform on AWS since September 28, 2026. AWS also documents a bedrock-mantle endpoint supporting the Anthropic Messages API. Check the model card’s current regional availability when choosing an access route.
Bedrock pricing is set independently by AWS. The Claude API table above is not a Bedrock quote. Bedrock prices were not verified in the supplied evidence; use the AWS Bedrock pricing page for the relevant AWS rate.
What Anthropic Reports on Benchmarks
In its launch announcement, Anthropic reports the following results:
| Evaluation | Sonnet 5.5 | Vendor-reported comparison |
|---|---|---|
| Terminal-Bench 4.0 | 70.6% | Sonnet 5: 10.3%; Opus 5.5 at Xhigh: 66.4% |
| CursorBench 4.0 | 55.5% | Opus 5.5: 57.8% |
| OSWorld 2.1 | 80.1% | — |
| GDPval-AA v2.1 | 1844 | Opus 5.5: 1846 |
Anthropic also reports that Sonnet 5.5 is more than 30% faster than Sonnet 5 and costs up to 30% less per task in Anthropic’s testing. Cost per task is a vendor evaluation result, distinct from the published per-token rates. These are vendor-run evaluations, effort settings differ, and AI Rank did not independently reproduce them.
Limitations and What to Check
Before choosing, verify the provider, model identifier, effort setting, pricing date, and workload requirements. Sonnet 5.5 supports adaptive thinking and defaults to high effort; the Arena entry uses xhigh. Its documented knowledge cutoff is June 2026, and retirement is not scheduled sooner than September 28, 2027.
Treat Arena intervals, missing Haiku data, and OpenRouter’s platform scope as limits on conclusions. Recheck cache-read pricing and AWS regional availability, and compare performance on your actual tasks before treating a published benchmark as a deployment decision.
Compare before you choose: pricing changes and rankings move daily — check the current model rankings on AI Rank before you commit.
Or 📱 get the AI Rank app to follow them on your phone.
Prepared with automated editorial assistance and checked against official sources and dated AI Rank data on October 11, 2026. We did not run our own benchmarks for this article.