Claude Opus 5.5 Pricing: API Rates, Plans & Real Costs

Claude Opus 5.5 pricing is a genuine cut: every rate on the card is lower than Opus 5's. Anthropic launched Opus 5.5 on September 22, 2026, at $4 per million input tokens and $20 per million output tokens.

The rate card only tells half the story, though. What you actually pay depends on how many tokens a task uses, and that depends heavily on the effort setting. This guide covers the full rate card, Claude plans, and cloud billing, then shows where the advertised savings hold and where they don't.

Claude Opus 5.5 pricing at a glance: $4 input, $20 output

Opus 5.5 costs $4 per million input tokens and $20 per million output tokens on Anthropic's API. Thinking tokens count as output. The full 1M-token context window is billed at standard rates, so a 900,000-token request costs the same per token as a short one.

Table 1: Claude Opus 5.5 API rate card per 1M tokens, as of September 2026. Source: Claude Platform pricing documentation.

A few details on the card matter more than they look:

• Cache reads are unusually cheap: a cache hit costs 5% of the input price on Opus 5.5, against the standard 10% on most Claude models, according to Anthropic's pricing docs.
• Batch halves both directions: asynchronous batch jobs get 50% off input and output, and the discount stacks with caching.
• Fast mode costs double: it delivers up to 2.5 times faster output, runs only on Anthropic's own API, and can't be combined with batch.
• US-only inference adds 10%: on Anthropic's API and Claude Platform on AWS, pinning inference to the US multiplies every token category by 1.1. Bedrock and Google Cloud price regional endpoints separately.

Opus 5.5 costs 20% less per token than Opus 5, and 60% less on cache reads

Every line on the Opus 5.5 rate card is lower than Opus 5's. It's the first Opus price cut since Opus 4.5, which lowered the tier from Opus 4.1's $15/$75 to $5/$25, where it stayed through Opus 5.

Table 2: Claude Opus 5.5 vs Claude Opus 5 API pricing per 1M tokens, as of September 2026. Source: Claude Platform pricing documentation.

The cache-read line is the one to watch for agents. An agent re-reads its instructions, tool definitions, and conversation on nearly every step, and when that context is cached, those re-reads bill at the cache-read rate. By Anthropic's account, cache reads make up the majority of agentic and coding costs, so a 60% cut there moves the total far more than the 20% headline.

Where the 40% saving comes from, and where it disappears

In Anthropic's tests, Opus 5.5 cost about 40% less than Opus 5 on typical workloads at default settings. According to its launch post, that comes from two things together: lower prices per token and fewer tokens per task. That figure was measured at default settings, and at max effort the second half reverses.

At max effort, Opus 5.5 works much harder than Opus 5. Artificial Analysis found it generates about 1.6 times as many output tokens per task on its Intelligence Index. In a post on X, it then broke down what that does to the bill:

1. Extra tokens alone: with no price change, Opus 5.5's heavier token use would have raised the cost per task by about 80%, to $10.51.
2. The 20% base price cut: brings that down to $8.41.
3. The cheaper cache reads: take it the rest of the way to $5.98, roughly level with Opus 5's $5.86.

So at max effort, the price cut pays for the extra thinking, and that's all it does. On long coding-agent runs, it doesn't even do that. On Artificial Analysis's Coding Agent Index, which runs Opus 5.5 inside Claude Code at max effort, it scored 66, the top score on the index. Its cost per task rose 21%, to $13.04 against Opus 5's $10.79, because it used substantially more tokens per task, according to a follow-up post.

The practical rule is simple. The 40% saving is real at default and medium settings, and it shrinks or reverses as you push effort toward max.

Effort level is the biggest price lever on Opus 5.5

The effort setting controls how long Opus 5.5 thinks before answering, and it moves the cost per task by about 11 times from bottom to top. Opus 5.5 defaults to medium, down from high on Opus 5.

Table 3: Claude Opus 5.5 vs Claude Opus 5 weighted average cost per Artificial Analysis Intelligence Index task at each effort level, independently measured, as of September 2026. Source: Artificial Analysis.

At each model's default setting, Opus 5.5 costs 37% as much per task as Opus 5, while scoring higher on Artificial Analysis's index (51 against 48). Artificial Analysis calculates cost per task as a weighted average across its benchmark tasks, so treat these figures as a like-for-like comparison, not a forecast of your own bill.

For quality at every setting, see our Opus 5.5 benchmarks breakdown. For budgeting, the takeaway is to start at medium and raise effort only for tasks that clearly need it.

What Opus 5.5 did to our own bill on Emergent

We tested the new pricing on real app builds on Emergent before writing this guide. At medium effort, the agent doing the building used 18% fewer Emergent credits on Opus 5.5 than on Opus 5. Our figures come from separate fintech app builds, one per model and effort level, each from a different brief, with no quality scoring. The main agent is the model under test; helper agents for design and testing run on other models.

First, the rates held up. Dividing each cost line on our bill by the tokens used gave back exactly Anthropic's list prices for both models, including the $0.20 cache read on Opus 5.5.

Table 4: Where the main building agent's spend went at medium effort, Emergent internal builds, September 2026.

The saving came almost entirely from cheaper cache reads. Opus 5.5 still spent more on output and cache writes, which is the thinking-more pattern from the section above showing up on a real bill. The whole build, including the helper agents, used 9% fewer credits, and a second pair of builds at low effort showed a larger drop for the main agent.

One cost pattern applied to both models. About a quarter of the main agent's spend came from rebuilding context whose cache had expired, usually after a pause of five minutes or more. If your own workflow regularly pauses that long, the 1-hour cache write ($8 per million tokens) can cost less overall than rewriting a 5-minute cache several times.

What a real Opus 5.5 session costs: three worked examples

Real costs depend on how your tokens split between fresh input, cached context, and output. These three examples use Anthropic's list prices and exclude tool, search, and other feature charges.

Table 5: Worked costs at Claude Opus 5.5 and Claude Opus 5 list prices, as of September 2026.

The agent session shows the cache effect most clearly. It comes in about 27% cheaper on Opus 5.5 at identical token counts, against a flat 20% for the simple request. The batch month would cost $80 at standard rates, so batching halves it. None of these examples includes the extra savings Anthropic attributes to Opus 5.5 using fewer tokens per task.

Claude plans that include Opus 5.5

Most Claude plans charge a flat monthly fee instead of per token, and Opus 5.5 requires a paid plan. The Free plan includes Sonnet and Haiku but not Opus models, according to Claude's plans page.

Table 6: Claude plans and Opus 5.5 access, as of September 2026. Prices exclude tax. Source: Claude plans and pricing page.

Usage limits on every plan reset on a rolling five-hour window, and paid plans add weekly caps. Claude on the web, desktop, mobile, and Claude Code all draw from the same pool, and there's no fixed message count. When you hit a limit on a paid plan, you can turn on usage credits and keep working at standard API rates.

At launch, Anthropic said it raised five-hour usage limits on Pro, Max, Team, and seat-based Enterprise plans. It also gave subscribers a one-time rate-limit reset they can save and use when they choose.

Opus 5.5 pricing on AWS, Google Cloud, and Microsoft Foundry

Opus 5.5 is available on Amazon Web Services, Google Cloud, and Microsoft's cloud from launch, but billing works differently on each:

• Amazon Bedrock and Google Cloud: the cloud provider sets the price and invoices you, so check Bedrock or Google Cloud pricing directly. Regional and multi-region endpoints cost 10% more than global ones.
• Claude Platform on AWS and Microsoft Foundry: usage is rated at Anthropic's standard prices, then billed through the cloud marketplace in Claude Consumption Units at $0.01 each.
• Fast mode: available only on Anthropic's own API, not on Claude Platform on AWS, Bedrock, or Google Cloud.

How Opus 5.5 pricing compares with the rest of the Claude lineup

Opus 5.5 now sits in the middle of Anthropic's price range, well below Fable 5.1 and above Sonnet 5.

Table 7: Current Claude model pricing per 1M tokens, as of September 2026. Source: Claude Platform pricing documentation.

Opus 5.5 costs 60% less than Fable 5.1 on input and output tokens, and 20% less on cache reads. Anthropic's model guidance now recommends it as the starting point for most workloads. Fable 5.1 is for demanding reasoning, long-horizon work, or tasks where Opus at higher effort still falls short. For the full Fable rate card, see our Fable 5.1 pricing guide. Sonnet 5 matches Opus 5.5's cache-read price at half the token rate.

When Opus 5.5 is worth the price, and when a cheaper model wins

Opus 5.5 is worth its price for long, multi-step work where quality matters: agentic coding, research, analysis, and document-heavy tasks. Its low cache-read price makes it especially cost-effective for agents that carry a large context through many steps.

A cheaper model is the better buy for short, high-volume jobs like classification, routing, simple extraction, and bulk drafting. Sonnet 5 at $2 and $10 handles much of that work, and Haiku 4.5 at $1 and $5 goes lower still. Sonnet 5.5 and Haiku 5.5 are also due soon, according to Anthropic, which may shift that math again.

Whichever model you choose, effort is your main cost control. Medium is the default for a reason, and max effort can erase Opus 5.5's per-token savings entirely.

Build with Opus 5.5 on Emergent without managing token bills

Claude Opus 5.5 pricing is the best value the Opus line has offered. It's 20% cheaper per token than Opus 5, 60% cheaper on cache reads, and about 40% cheaper on typical work at default settings in Anthropic's tests. The catch is effort. Run it at medium or high and the savings are real, as our own 18% credit drop showed. Push it to max and the extra thinking eats most or all of the discount.

On Emergent, you don't manage per-token rates at all. Opus 5.5 is live on Emergent, and you pick it when you start a project. If the app you build needs its own AI features, the Universal LLM Key gives it one credential for Claude, GPT, and Gemini, with usage drawn from your Emergent credits.

Start Building on Emergent.