Claude Sonnet 5.5
What it is
Anthropic's mid-tier Claude model, stronger than Haiku at multi-step tool use. It scored highest of our three models on the τ²-bench airline support benchmark.
Best for
#1 of 3Model for Ship a SaaS MVP in a weekend
Total 7.2 / 10
- Task quality (30%)8 / 10WebDev Arena 1768 (#3; rank spread overlaps #2-#4).
Evidence:0078 (source: LMArena)
- Cost per typical job run (20%)6 / 10$2/$10 per MTok, but verbose: $5.46 per AA index task.
Evidence:0081 (source: Artificial Analysis)0016 (source: Anthropic)
- Latency / speed (15%)8 / 10141.4 output tokens/s.
Evidence:0081 (source: Artificial Analysis)
- Tool-calling & structured-output reliability (15%)8 / 1078.0% on τ²-Airline.
Evidence:0014 (source: OpenRouter)
- Context & modality support (10%)8 / 101M context.
Evidence:0017 (source: Anthropic)
- Openness / deployability (10%)4 / 10Closed weights; API + Bedrock + Google Cloud.
Evidence:0016 (source: Anthropic)
#2 of 3Model for Build a customer-support agent
Total 7.1 / 10
- Task quality (15%)8 / 1078.0% τ²-Airline.
- Cost per typical job run (25%)6 / 10$0.16/task; $2/$10 per MTok.
- Latency / speed (15%)8 / 1082 s/task, fastest on the board.
Evidence:0014 (source: OpenRouter)
- Tool-calling & structured-output reliability (25%)8 / 10
Evidence:0014 (source: OpenRouter)
- Context & modality support (10%)8 / 10
Evidence:0017 (source: Anthropic)
- Openness / deployability (10%)4 / 10
Evidence:0016 (source: Anthropic)
#3 of 3Model for Build a deep-research agent
Total 7.2 / 10
- Task quality (40%)8 / 10Top of held-out BrowseComp Extended (73.0%); 86.6% on the original set.
Evidence:0117 (source: Mercor)
- Cost per typical job run (20%)6 / 10$2/$10 per MTok, but $5.46 per AA index task (verbose).
Evidence:0081 (source: Artificial Analysis)0016 (source: Anthropic)
- Latency / speed (5%)8 / 10141.4 output tokens/s.
Evidence:0081 (source: Artificial Analysis)
- Tool-calling & structured-output reliability (15%)8 / 1078.0% on τ²-Airline.
Evidence:0014 (source: OpenRouter)
- Context & modality support (10%)8 / 101M context.
Evidence:0017 (source: Anthropic)
- Openness / deployability (10%)4 / 10Closed weights; API + Bedrock + Google Cloud.
Evidence:0016 (source: Anthropic)
Strengths
- 78.0% on τ²-bench airline (support agent with policy, tools and simulated customer).0014 (source: OpenRouter)
- 1M-token context window and 128K max output.0017 (source: Anthropic)
- Also available via Amazon Bedrock and Google Vertex AI.0016 (source: Anthropic)
Limits
- 20x Haiku's price per token ($2/$10 vs $0.10/$0.50 per MTok).0016 (source: Anthropic)
- Proprietary, API-only model.0016 (source: Anthropic)
Pricing
Paid
$2 input / $10 output per MTok.0016 (source: Anthropic)
As of . Check the vendor's page before buying.
License
Proprietary model (hosted access only).0158 (source: Anthropic)0174 (source: Anthropic)0159 (source: Anthropic)
Checked .
Alternatives in the same layer
Disclosures
- None