Claude Sonnet 5.5
The workhorse: faster and cheaper than Opus, and it edges ahead on agentic coding.
Overall Intelligence
#9
of 56
3.5 pts behind #1
2
races led
8
weeks at #1
85.2
score
- Released
- Sep 28, 2026
- 3 wk ago
- Context
- 1M
- 128K max output
- Price
- $2 / $10
- per 1M in/out · $4.00 blended
- Speed
- 142 tok/s
- 2.6s to first token
- Modalities
- text · image
- outputs text
- Weights
- Closed
- API and apps
The 30-second version
The workhorse of the Claude family. Sonnet 5.5 is fast, affordable and, unusually for a mid-tier model, it edges ahead of its bigger siblings on agentic coding benchmarks.
- 1Leads our agentic coding and terminal races in the current snapshot, at half Opus pricing.
- 2Around 140 tokens per second: fast enough to feel interactive in an IDE.
- 3The sensible default for most production coding agents and chat products.
- 4Less strong than Opus on ambiguous, judgement-heavy problems and deep debugging.
What you need to know
Purpose
A high-throughput frontier model for the bulk of real work: coding agents, customer-facing assistants and pipelines that run thousands of times a day.
What Anthropic optimised for
Agentic coding performance per dollar and per second. Anthropic tuned Sonnet 5.5 heavily on tool use and terminal workflows, which is why it can beat Opus on those specific races.
Unique strengths
- Best-in-class agentic coding for its price.
- Excellent terminal and shell fluency.
- Strong tool-use accuracy with low latency.
Where it ranks
Current snapshot · illustrative scores · click any race to watch it
Sonnet 5.5Best on each axis
Overall
Coding
Hard Reasoning
Math
Science
Writing & Language
Multimodal
Tool Use & Agents
Speed & Cost
Overall Intelligence over time
Best use cases
01
IDE and CLI coding agents
Implementing features, fixing tests and running commands with quick feedback loops.
02
Production assistants
Support bots and internal tools where speed and cost matter at scale.
03
Sub-agent for Opus
The executor in planner-executor setups where Opus plans and reviews.
Strengths
Agentic coding
Leads the agentic coding race in our current snapshot.
Speed
Roughly 50% faster than Opus 5.5.
Price/performance
Among the best intelligence-per-dollar ratios in the frontier tier.
Weaknesses
Ambiguity
More likely than Opus to take an underspecified request literally.
Hardest reasoning
Several points behind the top tier on frontier reasoning races.
Prompts that play to its strengths
Test-driven feature
Add pagination to the `/api/orders` endpoint (cursor-based, page size up to 100). Write the tests first, run them to see them fail, implement, then run the full suite. Show me the final diff and test output.
Why it works: Tight feedback loops and concrete success criteria are where Sonnet shines.
Shell troubleshooting
Our Docker build fails in CI but not locally. Here is the CI log: <log>...</log>. Diagnose step by step, propose the fix, and give me the exact commands to verify it locally.
Why it works: Plays to its terminal and DevOps fluency.
When to choose it, and when not to
Tasks are well specified and speed or cost matters.
The job is a coding agent or terminal workflow rather than writing or conversation.
You want top agentic coding without self-hosting or open-weights requirements.
Pick something else when…
- The request is vague and high-stakes: Opus 5.5 handles ambiguity better.
- You need the cheapest possible tokens at scale: Haiku 5.5 or DeepSeek V4.1 Flash.
Update history
The distilled version first; open any release for the deeper breakdown.
The Claude Sonnet family