Claude Opus 5.5
Fable-class capability at Opus prices: the default pick for serious agentic and coding work.
Overall Intelligence
#1
of 56
leading the field
12
races led
50
weeks at #1
88.7
score
- Released
- Sep 22, 2026
- 4 wk ago
- Context
- 1M
- 128K max output
- Price
- $4 / $20
- per 1M in/out · $8.00 blended
- Speed
- 96 tok/s
- 4.4s to first token
- Modalities
- text · image
- outputs text
- Weights
- Closed
- API and apps
The 30-second version
Anthropic's flagship for everyday serious work. Opus 5.5 brings most of the capability of the pricier Fable tier down to Opus pricing, and it is the most dependable model in the race for long, tool-heavy coding and agent sessions.
- 1The default choice for agentic coding, refactors and multi-hour tasks where reliability matters more than raw speed.
- 2Leads the overall race, though only by a photo finish, at roughly half the price of the top tiers.
- 3Exceptionally good at following long, layered instructions and keeping a plan coherent across hundreds of tool calls.
- 41M-token context, with recall that stays usable deep into large codebases and document sets.
What you need to know
Purpose
A frontier generalist that can be trusted to work unsupervised: plan, use tools, check its own work and finish the job, at a price that makes it practical as the main model in production agents.
What Anthropic optimised for
Anthropic optimised Opus 5.5 for sustained agentic reliability: fewer derailments over long sessions, better self-verification, safer tool use and stronger honesty about uncertainty, rather than chasing a single headline benchmark.
Unique strengths
- Keeps multi-step plans on track across very long sessions without drifting from the brief.
- Debugging and root-cause analysis that reads like a senior engineer's notes.
- Writes careful, well-structured prose and documentation, not just code.
- Strong calibration: tends to say what it is unsure about instead of bluffing.
Where it ranks
Current snapshot · illustrative scores · click any race to watch it
Opus 5.5Best on each axis
Overall
Coding
Hard Reasoning
Science
Writing & Language
Multimodal
Tool Use & Agents
Speed & Cost
Trust & Factuality
Overall Intelligence over time
Best use cases
01
Autonomous coding agents
Repo-scale changes, migrations and bug hunts run through Claude Code or similar harnesses.
02
Complex analysis
Reading long contracts, filings or research and producing defensible, sourced conclusions.
03
Orchestrating sub-agents
A planner that delegates to cheaper models (Haiku 5.5, Sonnet 5.5) and reviews their output.
04
High-stakes writing
Technical specs, policy drafts and documentation that must be precise and consistent.
Strengths
Long-horizon reliability
Among the best at finishing multi-hour tasks without human rescue.
Debugging
Leads our debugging and repair race in the current snapshot, just ahead of Fable 5.1.
Instruction following
Handles long system prompts with many conditional rules unusually well.
Weaknesses
Not the fastest
Around 90-100 tokens per second; Sonnet 5.5 or Gemini 3.5 Flash feel snappier for chat.
Peak math
GPT-6 Astra and Grok 4.7 edge it on research-grade and competition math.
Prompts that play to its strengths
Repo-wide refactor with guardrails
You are working in this repository. Goal: replace our hand-rolled retry logic with the shared `withRetry` helper everywhere. Constraints: - Do not change public function signatures. - Keep behaviour identical, including backoff timing. - Run the test suite after each package you touch. First list every call site you find and your plan. Then make the changes package by package, and finish with a summary of anything you were unsure about.
Why it works: Plays to its planning, tool use and honesty about uncertainty, and gives it a concrete definition of done.
Root-cause analysis
Here is a production incident timeline, the relevant logs and the diff that shipped before it. <timeline>...</timeline> <logs>...</logs> <diff>...</diff> Work out the most likely root cause. Rank alternative explanations by likelihood, say what evidence would confirm or rule out each, and propose the smallest safe fix.
Why it works: Rewards careful reasoning over evidence and explicit uncertainty, which is where Opus is strongest.
When to choose it, and when not to
You need steady multi-hour agent runs or deep debugging at roughly 40% of the price.
The task is ambiguous, judgement-heavy, or the cost of a wrong answer is high.
The work is mostly text and code rather than video, audio or images.
Pick something else when…
- You need sub-second responses for high-volume chat: use Haiku 5.5 or Gemini 3.5 Flash.
- The job is mainly understanding video or audio: Gemini 4 Argon is far ahead.
- You must run on your own hardware: pick an open-weights model such as Muse Spark Open or DeepSeek V4 Pro.
Update history
The distilled version first; open any release for the deeper breakdown.
The Claude Opus family