NEWEST OPUS1M CONTEXTCHEAPER THAN OPUS 5CACHE AT 5% OF INPUTTHINKING ALWAYS ON
Claude Opus 5.5 API Pricing
Claude Opus 5.5 replaced Claude Opus 5 as Anthropic's recommended default on September 22, 2026, and it costs less than the model it replaces: $4.00/M input and $20.00/M output against Opus 5's $5.00 / $25.00. The cache read is the bigger break with the house rule — every other Opus reads cache at 10% of input, this one at 5%, or $0.20/M. Thinking is always on and cannot be switched off, so reasoning tokens bill as output on every request.
Input - per 1M tokens
$4.00/M
Cheaper than the model it replaces -20% vs Opus 5
Output - per 1M tokens
$20.00/M
Includes thinking tokens -20% vs Opus 5
Cached input - 95% off
$0.20/M
Cache write $5.00 / 5 min -95%
Effective - agentic blend
$2.58/M
92/8 split - 82% cache
§ 01 / TERMINAL
Claude Opus 5.5 API cost calculator.
Live calculator pre-loaded with Claude Opus 5.5 rates. Thinking cannot be disabled on this model, so put your expected reasoning tokens in the reasoning field — they bill at the output rate whether you plan for them or not, and the effort parameter (default medium) is the only lever on how many there are.
$ /mo
Workload split
Prompt cache hit rate
Tokens you can process
—
Words equivalent (English)
—
Effective rate
—
§ 02 / SCENARIOS
Claude Opus 5.5 cost on real workloads.
CODING AGENT
Codebase-scale migration
$0.328/task
LONG DOC
Reading 100-page contracts
$0.352/doc
RAG SUPPORT
Ticket triage
$0.029/ticket
ASSISTANT
Research planning turn
$0.059/turn
Estimate · anthropic-bpe-estimate · ≈2.6 chars/token Auto-counts as you type
This is a chars-per-token approximation, not a real tokenizer. Actual tokens vary by language, code density, and tool-call overhead — counts are typically ±10–20% off for English prose, more for code or non-Latin scripts. For exact billing, use the vendor's official tokenizer.
Characters 740
Words 124
Tokens (estimated) 285 tokens
Cost as input · uncached $0.00114 USD
Cost as output · uncached $0.0057 USD
Cost as cached input $0.00006 USD
| Model | Input /M | Output /M | Effective blended | Context | Best for |
|---|---|---|---|---|---|
| Claude Opus 5.5 Current | $4.00 cache $0.20 | $20.00 | $2.58 agentic 92/8 | 1M | Long-running agentic coding and knowledge work |
| Claude Opus 5 | $5.00 cache $0.50 | $25.00 | $3.41 +32% | 1M | The model Opus 5.5 replaces |
| Claude Fable 5.1 | $10.00 cache $0.25 | $50.00 | $6.26 +2.4× | 1M | Anthropic's top capability tier |
| Claude Sonnet 5.5 | $2.00 cache $0.10 | $10.00 | $1.29 −2.0× | 1M | Speed-and-intelligence default |
| Claude Haiku 4.5 | $1.00 cache $0.10 | $5.00 | $0.682 −3.8× | 200K | Support and classification volume |
| Grok 4.7 | $2.00 cache $0.50 | $6.00 | $1.19 −2.2× | 500K | Cheaper frontier coding model |
| Gemini 3.8 Flash | $0.75 cache $0.075 | $3.75 | $0.481 −5.4× · promo | 1M | Mainstream multimodal work |
Claude Opus 5.5 has published scores on 10 benchmarks we track, and leads 3 of them and shares first place on ARC-AGI-1. 3 of these scores come from an independent leaderboard; the rest are the lab's own reports.
| Benchmark | Score | Rank | Leader | Source |
|---|---|---|---|---|
| ARC-AGI-1 Reasoning | 98.5% ARC Prize verified, best result | #1 of 99 | Tied for first with 5 more | Official leaderboard ARC Prize Foundation |
| ARC-AGI-2 Reasoning | 93.3% ARC Prize verified, best result | #3 of 105 | GPT-6 Astra · 95% | Official leaderboard ARC Prize Foundation |
| FrontierCode Coding · v1.1 Main (100 hardest tasks) | 54.4% adaptive thinking, max effort | #1 of 13 | Claude Opus 5.5 leads | Vendor-reported Anthropic |
| Terminal-Bench 4.0 Agents and computer use | 64.8% max effort, Claude Code harness, 330 trials, ±3.1 pts (95% CI) | #3 of 34 | Kimi K3.1 Pro · 65.7% | Official leaderboard Terminal-Bench (Stanford, Laude Institute) |
| CursorBench 4.0 Coding | 57.8% adaptive thinking, max effort | #1 of 7 | Claude Opus 5.5 leads | Vendor-reported Anthropic |
| Humanity's Last Exam Expert knowledge · With tools | 67.7% with tools, max effort | #3 of 20 | Claude Fable 5.2 · 74.2% | Vendor-reported Anthropic |
| OSWorld 2 Agents and computer use · 2.1, partial score | 81.8% partial reward, max effort | #1 of 6 | Claude Opus 5.5 leads | Vendor-reported Anthropic |
| Chartography Vision and documents · With tools | 89% with tools | #2 of 5 | Kimi K3.1 Pro · 90.2% | Vendor-reported Anthropic |
| Terminal-Bench-Science AI research | 58.7% adaptive thinking, max effort | #5 of 12 | Claude Mythos 5.2 · 75.3% | Vendor-reported Anthropic |
| AutomationBench Agents and computer use · Main set (as reported) | 40% run by Zapier, no fallback models | #11 of 22 | MiniMax M3.1 Flash · 62.8% | Vendor-reported Anthropic |
Claude Opus 5.5 API pricing FAQ.
Practical Claude Opus 5.5 pricing questions - what the rate card says, and where this model's billing differs from the Opus before it.
Q · 01 What is Claude Opus 5.5 priced at? +
Anthropic lists
claude-opus-5-5 at $4.00/M input, $0.20/M cache-hit input and $20.00/M output. Cache writes are $5.00/M for the 5-minute window and $8/M for the 1-hour window, and batch is $2.00/M and $10.00/M.Q · 02 Is Opus 5.5 cheaper than Opus 5? +
Yes, on every line. Input and output are 20% lower —
$4.00 / $20.00 against $5.00 / $25.00 — and the cache read is 60% lower, because Anthropic prices it at 5% of input here and 10% on Opus 5. Under our standard blend (92% input, 82% cache hits) that is $2.58/M against $3.41/M. Opus 5 remains on the price list.Q · 03 Why is the cache read 5% of input and not 10%? +
Because Anthropic says so on the pricing page: a cache hit costs 10% of the base input price on every model except the Fable and Mythos 5.1 and 5.2 models, where it is 2.5%, and Claude Opus 5.5, where it is 5%. That is
$0.20/M rather than the $0.40/M the house rule would give. It makes a warm prompt worth more here than on any other Opus.Q · 04 Can I turn thinking off to cut the bill? +
No. On Opus 5.5 thinking is always on:
thinking: {"type": "disabled"} and manual thinking budgets both return a 400 error, which is a breaking change from Opus 5. Depth is controlled by the effort parameter, whose default on this model is medium. Reasoning tokens bill at the output rate, so model them explicitly before migrating.Q · 05 Does the full 1M context cost extra? +
No. Anthropic includes the full 1M-token window at standard pricing on Claude 4.6 and later — a 900k-token request bills at the same per-token rate as a 9k-token one — and prompt caching and batch discounts apply across the whole window. Max output is 128k tokens.
Q · 06 What do batch and fast mode cost? +
Batch is half price:
$2.00/M input and $10.00/M output. Fast mode is $8/M input and $40/M output across the full context window. It left research preview on October 15, 2026, so it no longer needs a beta header, and it runs on the Claude API and on Claude in Microsoft Foundry for Azure-hosted deployments - not on Claude Platform on AWS or other partner-operated clouds, and not with the Batch API.Q · 07 How does it compare with Claude Fable 5.2 on price? +
Fable 5.2 lists at
$8.00/M input and $40.00/M output, twice this model's base card, but its cache hit is the same $0.20/M, so on our blend it costs $5.01/M against $2.58/M here. Anthropic's guidance is to start with Opus 5.5 and move up to Fable for the most ambitious, open-ended technical work. Fable 5.1 is still on sale at $10.00/M and $50.00/M.Q · 08 Does regional pricing differ? +
Yes. The Claude API is global by default; regional and multi-region endpoints carry a 10% premium over global ones, and that applies to this model as it does to every model from Claude Opus 4.5 onward. Opus 5.5 is also sold through Amazon Bedrock, Claude Platform on AWS, Google Cloud and Microsoft Foundry, each at that platform's own rates.
Q · 09 What else breaks when migrating from Opus 5? +
Three things beyond mandatory thinking, and none of them changes the price. Forced tool use is rejected —
tool_choice types any and tool return a 400; thinking blocks are tied to the model that produced them; and on the Claude API and Google Cloud the older computer_20251124 computer-use tool is not accepted. Text between tool calls now arrives inside thinking blocks, which can silence a progress indicator that streams it.