Claude Opus 5.5: Specifications, Pricing, and Independent Benchmark Reality

Claude Opus 5.5 (model identifier claude-opus-5-5) is Anthropic's flagship frontier language model, released on September 22, 2026. Built for long-running autonomous agents, complex software engineering, and professional workflows, the model features a 1,000,000-token context window, a 128,000-token maximum output limit, and generally available cloud access alongside paid Claude subscription tiers.

Anthropic lowered API list pricing to $4.00 per million input tokens and $20.00 per million output tokens—a flat 20% reduction from the launch price of Claude Opus 5 ($5/$25). Prompt-cache reads dropped 60% to $0.20 per million tokens.

While Anthropic claims on its homepage that Opus 5.5 is "faster and costs 40% less to run," this is an attributed vendor claim regarding operational task cost rather than an across-the-board 40% reduction in API list prices (base input and output rates dropped 20%). Furthermore, Opus 5.5 always engages adaptive thinking before replying—with thinking tokens billed at standard output rates—meaning total task expenditure depends heavily on reasoning depth, cache hit rates, and workload shape.

On independent leaderboards, Artificial Analysis ranks Claude Opus 5.5 at #1 on its aggregate Intelligence Index (score 58 under adaptive reasoning with max effort). On the GDPval-AA v2.1 evaluation, Opus 5.5 also leads at max effort (Elo 1846, #1) and ranks second at xhigh effort (Elo 1820, #2). However, its scores change substantially across effort levels, with high (1692) and medium (1576) configurations ranking below top configurations of Claude Fable 5.1 and Claude Opus 5—demonstrating significant effort sensitivity rather than an inability to lead the benchmark.

Dimension	Claude Opus 5.5 (`claude-opus-5-5`)	Claude Opus 5 (`claude-opus-5`)	Material Change
Release Date	September 22, 2026	July 2026	New flagship release
Model Role	Frontier reasoning & agentic execution	Frontier reasoning & agentic execution	Iterative capability tier
Context Window	1,000,000 tokens	1,000,000 tokens	Identical context boundary
Maximum Output	128,000 tokens	128,000 tokens	Identical output ceiling
Input Modalities	Text, Image, PDF	Text, Image, PDF	Unchanged input support
Output Modality	Text	Text	Text generation only
Input Token Price	$4.00 / 1M tokens	$5.00 / 1M tokens	20% price reduction
Output Token Price	$20.00 / 1M tokens	$25.00 / 1M tokens	20% price reduction
Prompt-Cache Read	$0.20 / 1M tokens (1/20th input)	$0.50 / 1M tokens (1/10th input)	60% cheaper cache reads
Thinking Mode	Always thinks before replying (billed as output)	Adaptive thinking available	Mandatory thinking pass
Official Availability	Google Cloud (GA), Claude API, Claude Chat / Code	Google Cloud, Claude API, Claude Chat / Code	Generally available at launch

Technical Specifications and Platform Availability

Claude Opus 5.5 carries the API model identifier claude-opus-5-5. Google Cloud lists the model as generally available within its partner model catalog, positioning it as Anthropic's most capable Opus tier for long-horizon agent loops, enterprise coding, and professional knowledge work.

The 128,000-token maximum output limit accommodates substantial code files, extensive refactor diffs, and detailed reports in a single generation, provided client applications configure appropriate timeout governors.

API Pricing: Base Rates vs. Prompt Caching

The core economic change in Claude Opus 5.5 operates on two distinct pricing tiers: base token processing and prompt-cache retrieval.

Base Token List Rates

List prices for raw token processing dropped 20% across both directions:

Input: $4.00 per million tokens (down from $5.00/M on Opus 5)
Output: $20.00 per million tokens (down from $25.00/M on Opus 5)

Reuters corroborated these figures at launch, noting that the 20% price cut applies directly to API calls across major cloud platforms.

Prompt-Cache Economics

The more substantial structural shift occurs in prompt caching. On Opus 5, prompt-cache reads cost $0.50 per million tokens, representing one-tenth (10%) of the base input price.

On Opus 5.5, cache reads drop to $0.20 per million tokens—a 60% reduction compared to Opus 5. Cache reads now cost one-twentieth (5%) of the base input rate:

Opus 5 Prompt-Cache Read:
  $0.50 / 1M tokens  ──>  10% of base input rate ($5.00/M)

Opus 5.5 Prompt-Cache Read:
  $0.20 / 1M tokens  ──>   5% of base input rate ($4.00/M)

For software systems that maintain persistent system instructions, broad API documentation schemas, or repository AST indices across multi-turn sessions, the 60% reduction in cache read pricing provides a larger proportional rate discount than the 20% base input discount.

The "40% Lower Cost" Claim vs. Reality

On its launch homepage, Anthropic states that Claude Opus 5.5 is "faster and costs 40% less to run."

Technical builders should treat this as an attributed vendor claim about operating cost, not a 40% reduction in input/output list prices. Base input and output list prices fell by 20% ($4.00/M and $20.00/M vs. $5.00/M and $25.00/M on Opus 5), while prompt-cache reads dropped 60% ($0.20/M vs. $0.50/M).

In its analysis of task costs on Opus 5.5, Anthropic outlines why actual task-level savings vary:

Cache Read Proportions: Cache reads are 60% cheaper. Anthropic states that a session consisting mostly of prompt-cache reads can save up to 60% on input tokens, whereas a short question with no cache and a long answer saves up to 20%. Anthropic notes that most Claude Code tasks sit between those boundary cases.
Turn Count and Retries: Different tasks require different numbers of turns. Anthropic notes that if a model resolves an agent task in fewer turns or avoids false starts, the total volume of processed tokens decreases.
Output and Thinking Volume: Output and thinking token volume differ across queries and tasks.
Variable Work Performed: Opus 5.5 may perform a different amount of work than Opus 5 on a given task, altering total tokens generated.
Repository and Task Variance: Anthropic explicitly emphasizes that no single cost-per-task number applies to every codebase. Different repositories require different amounts of exploration, context ingestion, and code generation.

While heavy cache reuse can increase savings, it is not established as the only mechanism capable of producing a larger task-level reduction. Total task savings depend on workload shape, turn count, output/thinking token volume, cache behavior, retries, and the amount of work required.

Thinking Tokens Are Billed as Output

Anthropic documents that Opus 5.5 always engages in thinking before generating a final response. Crucially:

Intermediate thinking tokens are billed at standard output token rates ($20.00 per million tokens).
The model may use more thinking tokens depending on task complexity and instruction ambiguity.
If a harder query causes the model to generate 3,000 thinking tokens where an earlier model generated 1,000, the increased output token volume can partially or fully offset the 20% unit price discount.

Anthropic explicitly advises teams to measure effective cost per completed task on their own repositories and workflows rather than assuming a universal percentage drop.

Independent Benchmark Evidence: Artificial Analysis

Independent evaluations from Artificial Analysis provide the earliest empirical baseline for Claude Opus 5.5 across multiple reasoning effort tiers.

Aggregate Intelligence Index

On Artificial Analysis's aggregate Intelligence Index, Claude Opus 5.5 held the #1 ranking across 661 tested configurations at launch:

Configuration / Effort Level	Intelligence Index Score	Reported Benchmark Rank
Adaptive reasoning, max effort (default fallback)	58	#1 / 661 configurations
Extra-high effort (`xhigh`)	56	Top-tier
High effort (`high`)	54	Top-tier
Medium effort (`medium`)	51	Upper tier
Low effort (`low`)	42	Mid tier

These measurements demonstrate that reasoning effort significantly impacts model intelligence scores: dropping from maximum effort to low effort decreases the index score from 58 to 42.

GDPval-AA v2.1 Leaderboard and Effort Sensitivity

On the Artificial Analysis GDPval-AA v2.1 evaluation, Claude Opus 5.5 leads the leaderboard when configured at maximum reasoning effort, with its extra-high effort configuration ranking second:

Model & Configuration	GDPval-AA v2.1 Elo	Leaderboard Status
Claude Opus 5.5, Adaptive Reasoning, Max Effort, Default Fallback	1846	#1
Claude Opus 5.5, Adaptive Reasoning, Xhigh Effort, Default Fallback	1820	#2
Claude Fable 5.1, Adaptive Reasoning, Max Effort, Default Fallback	1735	Competing frontier configuration
Claude Fable 5.1, Adaptive Reasoning, Xhigh Effort, Default Fallback	1721	Competing frontier configuration
Claude Opus 5, Adaptive Reasoning, Max Effort	1708	Prior-generation configuration
Claude Opus 5.5, Adaptive Reasoning, High Effort, Default Fallback	1692	Reduced effort tier
Claude Opus 5, Adaptive Reasoning, Xhigh Effort	1676	Prior-generation configuration
Claude Opus 5.5, Adaptive Reasoning, Medium Effort, Default Fallback	1576	Reduced effort tier

These results establish that Opus 5.5 currently leads GDPval-AA v2.1 at max effort, with its xhigh configuration currently second. However, its benchmark score changes substantially with effort level: high effort (1692) and medium effort (1576) configurations rank below competing configurations such as Claude Fable 5.1 and Claude Opus 5 at max and xhigh effort.

This divergence demonstrates effort sensitivity, not that Opus 5.5 fails to lead the GDPval evaluation.

Builders should treat aggregate and component benchmark leads as indicators of peak capacity under specific configurations, rather than universal proof of superiority across every production workflow or throttled effort setting.

Claude Subscription Usage Impact

For developers using Claude via chat interfaces or Claude Code subscriptions, Anthropic notes that the reduced underlying inference cost is reflected in account usage.

According to Anthropic:

Account allowances on Pro, Max, and Team plans go approximately 25% further with Opus 5.5 compared to Opus 5 when factoring in cached context.

This 25% figure is an attributed Anthropic product statement, not an independently measured guarantee. Exact session allowances depend on context length, cache hit rates, file attachments, and prompt frequency.

(Note: AI Model Waddle does not track individual account limits, reset windows, or subscription tiers; refer to Claude Help documentation for account-specific usage rules.)

Where Claude Opus 5.5 Fits in Production Architectures

Understanding where Opus 5.5 fits requires analyzing how its capabilities align with system roles and operational constraints.

Review our 5-role decision framework for matching software tasks to AI model architectures

concept

Best-Fit Production Workloads

Long-Running Autonomous Coding Agents: Multi-file codebase refactors, debugging complex distributed systems, and generating end-to-end pull requests. The combination of a 1M-token context window, 128k output ceiling, and $0.20/M cache reads directly targets the requirements of repository-level agent workflows.
Multi-Turn Context-Heavy Pipelines: Workflows that repeatedly query large static references—such as compliance rulebooks, architectural blueprints, or extensive API documentation schemas—where prompt caching keeps marginal input costs at $0.20/M.
Complex Architectural Planning: Orchestrator agents that decompose ambiguous user specifications into multi-stage execution DAGs, where deep multi-step reasoning is required to coordinate downstream tool calls.

Study our 10-stage production coding agent lifecycle for managing context drift and verification gates

workflow

Poor-Fit Production Workloads

High-Throughput Inline Classification: Tasks such as webhook routing, support ticket triage, or strict policy guardrails where simple bounded classification is sufficient. In production architectures, routing these tasks to a premium frontier generative model with mandatory thinking introduces unnecessary unit cost overhead when specialized classifiers or smaller models can satisfy the task (see our model selection framework).
Simple Single-Pass Transformations: Straightforward data extractions, format conversions, or basic summarizations. Simpler transformations often do not require a frontier Opus-tier model and should be evaluated against lower-cost alternatives using representative tasks.
Output-Heavy Uncached Batches: High-volume batch generation tasks with unique prompts that cannot leverage prompt caching, where $20/M output fees accumulate quickly.

Explore how to decouple frontier reasoning engines from low-latency action execution across 8 agent subsystems

concept

What AI Model Waddle Has Not Tested

In accordance with our publication standards, we disclose what remains unverified by our own testing:

When first-party testing data becomes available, we will publish controlled runs following our standardized protocols.

Read our complete evidence standard for distinguishing public facts, vendor claims, and first-party measurements

concept

Technical Decision Summary

Recommendation

A significant economic upgrade for cache-heavy agent loops, not an automatic 40% discount for every workload.

If you currently run Claude Opus 5 for complex repository work, architectural planning, or deep agent orchestration, switching to Claude Opus 5.5 (claude-opus-5-5) provides an immediate 20% base token discount and a 60% reduction in prompt-cache costs ($0.20/M). On independent leaderboards, Artificial Analysis ranks it #1 on both its aggregate Intelligence Index and the GDPval-AA v2.1 evaluation at maximum reasoning effort (though scores drop under reduced effort configurations).

However, do not assume an automatic 40% reduction in monthly spend. Opus 5.5 always generates thinking tokens billed at $20/M output rates. If your workload involves short, uncached prompts with heavy reasoning, increased thinking output can offset token savings. Measure your total cost per completed task—including cached tokens and thinking output—on representative production runs before adjusting budget models.