What are the strengths of this model?

20% price cut to $4/$20 and about 40% lower running cost than Opus 5 1M-token context with no long-context surcharge; cache reads 60% cheaper Pre-launch external evaluation by METR and Frontier Design; breakout attempts down about 85%

What are the weaknesses of this model?

Nearly every benchmark is vendor-reported at xhigh/max effort, above the API default CodeRabbit measured about 50% more output tokens on identical inputs - a 20% price cut can still raise the bill Always-on reasoning is a breaking change: it cannot be disabled, and legacy prompt assets degrade

What are the best use cases?

Multi-day engineering work: large codebase audits and migrations Computer use and desktop automation where a wrong click is expensive Regulated industries where third-party safety evaluation records are a procurement requirement

AnthropicProprietary

Claude Opus 5.5

Name: Claude Opus 5.5
Author: Anthropic

Compare this model

Anthropic's first Claude 5.5 model, released September 22, 2026. The claim is unusually narrow: Fable 5.1-level work on most tasks at about 40% lower running cost than Opus 5. List price is $4/$20 per million tokens (20% under Opus 5) on a 1M-token context window with no long-context surcharge. Self-reported system-card numbers: 89.9% on SWE-bench Pro, 66.4% on Terminal-Bench 4.0 at xhigh, 67.7% on HLE with tools. Adaptive thinking is always on - effort=low is the floor.

Share:X Hatena LINE

Parameters

Undisclosed

Context Window

License

Proprietary

Release Date

2026-09-22

Benchmark Performance

AA Intelligence Index

57.6

LMArena Elo

—

HLE

67.7

ARC-AGI-2

—

SWE-bench Verified

—

GPQA Diamond

—

MMLU-Pro

—

LiveCodeBench

—

AIME 2025

—

MATH-500

—

API Pricing

Input Price (per 1M tokens)

Output Price (per 1M tokens)

$20

Billing Mode: standard

Strengths

•20% price cut to $4/$20 and about 40% lower running cost than Opus 5
•1M-token context with no long-context surcharge; cache reads 60% cheaper
•Pre-launch external evaluation by METR and Frontier Design; breakout attempts down about 85%

Weaknesses

•Nearly every benchmark is vendor-reported at xhigh/max effort, above the API default
•CodeRabbit measured about 50% more output tokens on identical inputs - a 20% price cut can still raise the bill
•Always-on reasoning is a breaking change: it cannot be disabled, and legacy prompt assets degrade

Use Cases

•Multi-day engineering work: large codebase audits and migrations
•Computer use and desktop automation where a wrong click is expensive
•Regulated industries where third-party safety evaluation records are a procurement requirement

Deep Analysis

Artificial Analysis Intelligence Index

51.2 (medium) / 57.6 (max)

AA Index v4.3.2, ten evaluations at every effort level; GPT-6 Sol maxes at 47.5

Measured Cost per Task

$1.34 (medium) / $5.98 (max)

Same AA harness, list prices including cache reads and writes

Input / Output Price

$4 / $20 per 1M tokens

20% below Opus 5 ($5/$25); cached read $0.20, 5-min cache write $5.00

Context Window

1M tokens

128K max output, 300K via Batch beta; no long-context surcharge

Isolation-Boundary Escapes

about 85% fewer than Opus 5 / Mythos 5.1

Anthropic-reported; model was externally evaluated by METR and Frontier Design

Release

Sep 22, 2026

First model of the Claude 5.5 family; Sonnet 5.5 and Haiku 5.5 to follow

Strengths

・Fable 5.1-level output at 40% lower running cost than Opus 5, plus a 20% list-price cut and 30%+ faster output
・Reasoning effort scales from low to max on one endpoint, covering cheap and deep workloads without switching models
・No long-context surcharge across the full 1M window, unlike the 272K threshold that doubles GPT-6 Sol's input rate
・Shipped on Bedrock, Google Cloud, Microsoft Foundry and the Claude API simultaneously, and externally evaluated before launch

Weaknesses

・Still twice GPT-6 Sol's input and output rate at equal token volume; a 49-task route check put it at roughly 3.4x Sol's cost per completed task
・Reasoning cannot be disabled, and CodeRabbit measured about 50% more output tokens on identical inputs, so the 20% cut can raise real bills
・No published system card at launch
・SWE-bench Verified and GPQA Diamond figures for Opus 5.5 itself have not been published yet, so third-party coding numbers are still pending

Competitor Comparison

Model	Arena	SWE	GPQA	Price
GPT-6 Sol	AA 47.5 (max)	N/A	N/A	$2/$10
Claude Fable 5.1	AA 60	N/A	N/A	$10/$50
Claude Opus 5	AA 63	96.0%	N/A	$5/$25

Overview

Claude Opus 5.5 launched on September 22, 2026 as the first model in Anthropic's 5.5 family, and it is a cost play more than a capability leap. Anthropic says it matches Claude Fable 5.1 on most work while running 40% cheaper than Opus 5, with output 30% faster; the list price fell 20% to $4 per million input tokens and $20 per million output. On Artificial Analysis's Intelligence Index v4.3.2 it scores 51.2 at the default medium effort and 57.6 at max, against 47.5 for OpenAI's same-day GPT-6 Sol. The 1M-token window carries no long-context surcharge, but reasoning cannot be switched off. It is the first Anthropic model to ship after CEO Dario Amodei publicly called for slowing frontier development, and it passed external evaluation by METR and Frontier Design before release.

Benchmarks & Performance

Anthropic's own framing is comparative rather than absolute: Opus 5.5 reaches Fable 5.1 level on most tasks, and in CursorBench 4.0 it scored 11 points above GPT-5.6 Sol while costing about a third as much to run. Independent measurement through Artificial Analysis puts the model at 51.2 on the Intelligence Index at default medium effort (about $1.34 per task) and 57.6 at max effort (about $5.98 per task) — roughly ten index points clear of GPT-6 Sol at matched max effort. CodeRabbit's 24-to-48-hour evaluation found twice the bug detection of its baseline on complex tasks, but also about 50% more output tokens on identical input, which erodes the headline price cut. No SWE-bench Verified or GPQA Diamond score has been published for this specific model.

Detailed Comparison

**Claude Opus 5.5 vs GPT-6 Sol** - Same launch day, opposite bets: Opus 5.5 keeps the higher ceiling at $4/$20, Sol trades about ten AA index points for a $2/$10 rate card. - Anthropic's external evaluation and multi-cloud availability are real advantages; OpenAI's cost per task is roughly a fifth at matched medium effort. **Claude Opus 5.5 vs Claude Opus 5** - 20% cheaper on the rate card and 40% cheaper to run on typical workloads, with 30%+ faster output. - Anthropic reports about 85% fewer isolation-boundary escape attempts, though a system card is still missing. **Claude Opus 5.5 vs Claude Fable 5.1** - Fable 5.1 remains the flagship on capability and is still priced at $10/$50; Opus 5.5 aims to be the default workhorse instead.

Community Feedback

Reaction split along predictable lines. Teams already standardized on Bedrock or Google Cloud welcomed a cheaper Opus that needs no procurement change. Developers who ran their own evaluations pushed back on the headline discount after CodeRabbit reported roughly 50% more output tokens per request, arguing that thinking-token inflation can cancel a 20% rate cut. The missing system card drew criticism from safety researchers given Anthropic's own push for third-party evaluation. Investor attention focused on the timing, since Anthropic crossed $100 billion in annualized revenue and is targeting a November IPO.

Use Cases

**1. Long-horizon agentic coding** The model is tuned for long, multi-step software tasks and scales reasoning with the prompt. Use it where tasks run for hours and correctness matters more than throughput, and budget on output tokens rather than requests. **2. Regulated multi-cloud deployment** Simultaneous availability on Bedrock, Google Cloud and Microsoft Foundry means regulated buyers can run it inside existing contracts and controls without a new vendor review. **3. High-volume knowledge work** Financial research, document analysis and long-form writing are the advertised targets. At low and medium effort the cost per task drops sharply, which makes a single endpoint viable for both draft and review passes. Pin effort per task type; leaving everything at max is where the bill surprises people.

Latest News

**Sep 22, 2026**: Anthropic introduced Claude Opus 5.5 at $4/$20 per million tokens, 40% cheaper to run than Opus 5 and positioned at Fable 5.1 performance on most work. Artificial Analysis measured 51.2 index at medium and 57.6 at max effort. Sonnet 5.5 and Haiku 5.5 are announced for the coming weeks.

Sources

Analysis generated: 2026-09-23