Claude Opus 5.5

claude-opus-5-5

Claude Opus 5.5 is purpose-built for long-running agentic coding and knowledge work, delivering Fable 5.1-level performance on long-horizon coding and knowledge work.

Context window1.0M

ProviderClaude

Released2026/09/23

Playground

Pricing

The same model is available through multiple service channels — choose based on latency, reliability and cost.

Prices in $ / 1M tokens

To pick a channel, add a provider field to the request body, for example "provider": { "channel": "direct" }. Valid values are direct / stable / economical; omit it to use the default channel.

Direct

Direct upstream connection — best when you need native behavior and the full context window.

Input	Output	Cache read	Cache write
4.00/M	20.00/M	0.20/M	5.00/M

Overview

Input

Text Image

Output

Text

Claude Opus 5.5 API: Fable-level capability, long-horizon agentic model at an Opus price

Claude Opus 5.5 is the next-generation model in Anthropic's Opus series, released on September 22, 2026, less than two months after Opus 5. Its positioning is clear: achieve Fable 5.1-level performance on long-horizon coding and knowledge work while pushing unit prices down further—input and output are 20% cheaper than Opus 5, cache reads are 60% cheaper, and Anthropic's total cost estimate for a typical workload is about 40% lower than Opus 5.

The most notable capability is agentic coding: Terminal-Bench 4.0 scores 66.4% (xhigh level), higher than Fable 5.1's 55.8% and GPT-6 Astra's 57.9%; GDPval-AA v2.1 real-world occupational task evaluation scores 1846, 111 points higher than Fable 5.1. On the Artificial Analysis Intelligence Index v4.3.2, it tops the list with 58 points, leading the tied second-place Fable 5.1 and GPT-6 Astra by 5 points each.

Another thing with practical impact on production environments: this generation has made targeted adjustments for the "Claude voice"—putting the most important information first, using fewer terms and boilerplate. Anthropic says output is 40% more concise than Opus 5 with no change in accuracy. Output speed is also more than 30% faster than Opus 5.

SeaWhale AI provides Claude Opus 5.5 through an OpenAI-compatible interface and the Anthropic native Messages API, supporting reasoning Effort control, tool calling, streaming output, and image and PDF input.

Get API Key · Model ID: claude-opus-5-5

Why Choose Claude Opus 5.5

First in agentic coding — Terminal-Bench 4.0 scores 66.4%, Fable 5.1 is 55.8%, Opus 5 is 52.3%, GPT-6 Astra is 57.9%
First in knowledge work evaluation — GDPval-AA v2.1 scores 1846, Fable 5.1 is 1735, Opus 5 is 1708
Top of the AA Intelligence Index — v4.3.2 index 58, Fable 5.1 and GPT-6 Astra tied at 53, Opus 5 at 51
Prices lowered in reverse — input and output are 20% lower than Opus 5, cache reads are 60% lower, at roughly a 0.4x price tier compared with Fable 5.1
Writing that feels more human — output is 40% more concise than Opus 5, conclusions first, fewer terms, making agentic reports easier to read
More accurate image reading — Chartography chart recognition 89.0%; image-reading accuracy at the low level already exceeds Opus 5's highest level, without needing cropping and scaling tools
1 million token context + 128K output — flat unit price across the full window; can be extended to 300K output under the Batch API

Core Capabilities

01 Long-Horizon Autonomous Coding

Opus 5.5 is designed for coding sessions that run for hours. Terminal-Bench 4.0 66.4%, FrontierCode v1.1 54.4%, CursorBench 4.0 57.8%, all three higher than Fable 5.1. In a test case from Anthropic, a tester used it to complete a migration of 680,000 lines of code in one day; on a web page load optimization task it succeeded 39 out of 40 times, while Opus 5's changes often came with behavior changes.

Multi-file refactoring and large-scale code migration
Cross-session bug localization and code review
Unattended long-duration autonomous execution

02 Knowledge Work and Deep Research

GDPval-AA v2.1 scores 1846 and AA-Briefcase scores 1822; both knowledge work evaluations are currently first. Humanity's Last Exam with tools is 67.7% (Fable 5.1 is 65.6%), SciCode is 66.9% (the previous highest was Fable 5.1's 63.1%). AutomationBench automation tasks are 40.0%, 13 percentage points higher than Opus 5's 26.9%.

Document-intensive analysis in finance, policy, and legal domains
Multi-step scientific research and technical investigation
Produce complete reports and tables directly from a single question

03 Visual Understanding and Computer Use

This generation has a substantial change in image-reading ability: values on dense charts and position-dependent semantics (which two boxes an arrow connects, what the differences are between two versions of a chart, meeting start and end times in a calendar screenshot) can be read accurately without cropping and scaling tools. In Anthropic's tests, image-reading accuracy at the low level has surpassed Opus 5's highest level, while using only about one-tenth the output tokens. OSWorld 2.0 computer use is 81.8% (partial score); the default level already reaches the success rate of Opus 5's high level, with fewer steps and about 40% fewer tokens.

Parsing dense charts, technical drawings, and tables embedded in PDFs
Multi-step autonomous operation of browsers and desktop applications
Visual preprocessing scaffolding built for older models can be reassessed to see whether it is still needed

04 Clearer Communication and Lower Operating Costs

Writing style is a direction deliberately optimized in this generation: conclusions first, fewer terms, less boilerplate; during agentic tasks, progress reports and final summaries directly state what was done, what was found, and what is needed. Combined with cache reads dropping to 40% of Opus 5's cost, long sessions with dense tool calls are the scenario with the greatest savings.

Progress reporting and result summaries for agentic tasks
Higher adherence to writing instructions
The longer the cached prefix and the more it is reused, the greater the cost advantage

Best Use Cases

Scenario	Description
Overnight autonomous coding	Terminal-Bench 4.0 leads Fable 5.1 by more than 10 percentage points; large migrations can be completed in one day
Enterprise knowledge work	First in both GDPval-AA v2.1 and AA-Briefcase; end-to-end output of documents, spreadsheets, and reports
Chart and screenshot understanding	Accurately reads dense charts without tools; the `low` level is sufficient
Computer use agents	OSWorld 2.0 81.8%; default level achieves the effect of Opus 5's high level
Cache-intensive long sessions	Cache reads are billed at 40% of Opus 5, reducing total agent loop cost by about 40%
Cost reduction from Fable 5.1	Performance is comparable on most tasks, and the price tier is only 0.4x that of Fable 5.1

Differences Between Claude Opus 5.5, Opus 5, and Fable 5.1

Capability	Claude Opus 5.5	Claude Opus 5	Claude Fable 5.1
Model ID	`claude-opus-5-5`	`claude-opus-5`	`claude-fable-5-1`
Release date	September 22, 2026	July 24, 2026	September 1, 2026
AA Intelligence Index v4.3.2	58	51	53
Terminal-Bench 4.0	66.4%	52.3%	55.8%
GDPval-AA v2.1	1846	1708	1735
Terminal-Bench-Science 0.1	58.7%	29.0%	52.6%
OSWorld 2.0 (partial score)	81.8%	74.0%	80.7%
Context window	1 million tokens	1 million tokens	1 million tokens
Maximum output	128K tokens	128K tokens	128K tokens
Thinking mode	Always on, cannot be disabled	On by default, can be disabled below `high`	Always on, cannot be disabled
Default Effort	`medium`	`high`	`high`
Forced tool calling `tool_choice: any/tool`	Not supported, returns 400	Supported	Not supported, returns 400
Relative unit price for cache reads	0.05x base input price	0.1x base input price	0.025x base input price
Relative price tier	0.8x Opus 5	1	2x Opus 5
Positioning	The value-for-money first choice for long-horizon coding and knowledge work	Previous-generation Opus	Strongest publicly released model

Actual billing is subject to the real-time pricing card at the top of the page.

Frequently Asked Questions

When was Claude Opus 5.5 released? September 22, 2026. It launched on Claude API, Amazon Bedrock, Google Cloud, and Microsoft Foundry on the same day, while Opus 5 remains available. Anthropic says Sonnet 5.5 and Haiku 5.5 will follow within weeks.

What are the upgrades compared with Opus 5? Four areas. First, capability: Terminal-Bench 4.0 rose from 52.3% to 66.4%, GDPval-AA rose from 1708 to 1846, Terminal-Bench-Science doubled, and image reading no longer depends on cropping tools. Second, cost: input and output are 20% lower, cache reads are 60% lower, and total cost for a typical workload is about 40% lower. Third, output speed is more than 30% faster and writing is 40% more concise. Fourth, safety: behavior audit metrics outperform all recent Claude models, boundary-crossing attempts are 85% fewer than Opus 5, and it ties with Fable 5.1 for the lowest risk in the Gray Swan prompt injection evaluation.

Do I need to change code when migrating from Opus 5? There are four breaking changes. Thinking mode can no longer be disabled; thinking: disabled and budget_tokens both return 400; use effort to adjust depth. tool_choice values any and tool return 400; use auto with prompts, or strict: true and structured outputs. Thinking blocks are bound to the model that produced them; only Fable 5.1 and Mythos 5.1 can read Opus 5.5 thinking blocks, and when falling back to Opus 5 or 4.8 they are silently discarded; accounts created after August 31, 2026 also verify whether history has been edited, so conversations must remain append-only. Computer use only accepts computer_toolset_20260801; the older computer_20251124 returns 400. In addition, progress text between tool calls is now returned as a thinking block, empty by default; display: "updates" is needed to retrieve it, otherwise the frontend will be "silent" during long tasks.

What are the context and output limits? Context is 1 million tokens, the default is the maximum, with a flat unit price across the full window. Synchronous API maximum output is 128K tokens; with the output-300k-2026-03-24 beta header on the Batch API it can reach 300K. The tokenizer is the same as Opus 5, so token counts remain unchanged when migrating from Opus 5. Knowledge cutoff is June 2026.

Are there any areas where it lags? Yes. Terminal-Bench-Science 0.1 scores 58.7%, below GPT-6 Astra's 64.6%; AutomationBench 40.0% is also slightly below Astra's 41.4%. Artificial Analysis points out that it lags on CritPt, AA-LCR long-context reasoning, and GDP.pdf, and under the max level it consumes about 119,000 output tokens per task, 1.6x that of Opus 5 and more than 4x that of GPT-6 Astra—unit prices are lower, but total spending at high levels may not be. AA's own test of Terminal-Bench 4.0 is 59.6%, on par with GPT-6 Astra and below Anthropic's published 66.4%. On safety filtering, a new biology classifier has been added; dual-use research requests in virology, toxicology, and molecular design will be refused; most cybersecurity requests will be routed to Opus 4.8; requests that attempt to make the model reproduce internal reasoning in the body will be refused under the reasoning_extraction category and will not fall back to other models.

How do I choose the Effort level? Five levels: low, medium, high, xhigh, max; default is medium—note that Opus 5 defaults to high, so specify explicitly when migrating. At the same level, it thinks more than Opus 5, especially at xhigh and max; leave headroom in max_tokens for thinking. For routes that previously disabled thinking, start testing from low; image reading and computer use only need low or medium; for long-horizon coding, xhigh provides the largest gain. It supports switching Effort per message mid-conversation (beta), upgrading for difficult steps and downgrading for routine steps without invalidating the cache.

How do I choose between it and Fable 5.1? For most coding and knowledge work tasks, the two perform comparably; Opus 5.5 is even higher on Terminal-Bench 4.0 and GDPval-AA, while costing only 0.4x Fable 5.1. Fable 5.1's advantages are cheaper cache reads (0.025x vs. 0.05x), research tasks other than Terminal-Bench-Science, and being able to read Opus 5.5 thinking blocks whereas the reverse is not possible—if your workflow needs to upgrade from Opus 5.5 to a stronger model and continue, Fable 5.1 is the only option that does not lose reasoning.

Why Choose SeaWhale AI for the Claude Opus 5.5 API

Full passthrough of native Messages API — Effort control, per-message Effort switching, thinking blocks, prompt caching, task budgets, and beta features can all be used directly
Direct connection from mainland China — no overseas account or self-built proxy required; stability and latency are guaranteed by the platform
Pay-as-you-go — pay only for actual usage, with no minimum spending threshold
One key for multiple models — use Opus 5.5 daily, switch to Fable 5.1 when you need top-tier performance, and switch to Sonnet 5 for high traffic, all with unified billing under the same account

API

API integration

Model IDUse this value as the model in inference requests

claude-opus-5-5

API KeyBearer token used to authenticate inference requests

Manage API keys

Base URLOpenAI compatible · /chat/completions

https://api.haijingai.com/v2/

provider OptionalSelects a service channel; omit it and the system picks the default

"provider": { "channel": "direct" }

claude-opus-5-5 usage examples

SeaWhale AI is compatible with the OpenAI API protocol, so you can call it with the OpenAI SDK or plain HTTP requests. Streaming is enabled by default.

cURLPythonNode.js

curl https://api.haijingai.com/v2/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer <API_KEY>" \
  -d '{
    "model": "claude-opus-5-5",
    "messages": [
      {"role": "system", "content": "You are a helpful assistant."},
      {"role": "user", "content": "Hello!"}
    ],
    "provider": { "channel": "direct" },
    "stream": true
  }'
# provider is optional — remove this line to use the default channel

from openai import OpenAI

client = OpenAI(
    base_url="https://api.haijingai.com/v2",
    api_key="<API_KEY>",
)

stream = client.chat.completions.create(
    model="claude-opus-5-5",
    messages=[
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "Hello!"},
    ],
    stream=True,
    # Optional: pick a service channel; omit to use the default
    extra_body={"provider": {"channel": "direct"}},
)

for chunk in stream:
    if chunk.choices[0].delta.content:
        print(chunk.choices[0].delta.content, end="", flush=True)

import OpenAI from 'openai'

const client = new OpenAI({
  baseURL: 'https://api.haijingai.com/v2',
  apiKey: '<API_KEY>',
})

const stream = await client.chat.completions.create({
  model: 'claude-opus-5-5',
  messages: [
    { role: 'system', content: 'You are a helpful assistant.' },
    { role: 'user', content: 'Hello!' },
  ],
  stream: true,
  // Optional: pick a service channel; omit to use the default
  // @ts-expect-error provider is a SeaWhale AI extension, not in the OpenAI SDK types
  provider: { channel: 'direct' },
})

for await (const chunk of stream) {
  process.stdout.write(chunk.choices[0]?.delta?.content ?? '')
}