Claude Opus 5.5 is priced at $4 per million input tokens and $20 per million output tokens, with cache reads at $0.20 and a 50% Batch API discount. See what chat, code review, and agent workloads really cost, how it compares with Opus 5, Sonnet 5.5, and Haiku 4.5, which plans include it, and where free access ends.
Claude Opus 5.5 costs $4 per million input tokens and $20 per million output tokens on the Anthropic API. That is 20% below Claude Opus 5, which charges $5 and $25, and every other line on your bill is built from those two numbers. A short chat answer lands near two cents. A busy support desk lands near four hundred dollars a month. And the free route people keep searching for comes with a hard limit that most "free Opus" pages leave out. Here are real numbers for all three, so you can estimate your own bill before you write a single line of code.
Anthropic bills Claude by the token, and every rate below is per one million tokens (MTok). A token is roughly four characters of English, so 1,000 tokens is about 750 words, and a 10-page report is around 7,000 tokens.
| Item | Price per million tokens |
|---|---|
| Input, standard | $4.00 |
| Output, standard | $20.00 |
| Cache read | $0.20 |
| Cache write, 5-minute | $5.00 |
| Cache write, 1-hour | $8.00 |
| Batch input | $2.00 |
| Batch output | $10.00 |
| Fast mode, input and output | $8.00 and $40.00 |
Four more facts shape the bill:
claude-opus-5-5The formula is simple: (input tokens × 4 + output tokens × 20) ÷ 1,000,000 gives the cost in dollars. A request with 10,000 input tokens and 1,000 output tokens costs $0.04 plus $0.02, so six cents. Every API response reports the exact input and output token counts in its usage data, so log those numbers per request and multiply. Teams that skip this step tend to guess low on output, which is the expensive side.
💡 Output costs five times as much as input. A long answer drives the bill far more than a long prompt does, so set a sensible
max_tokenson every request.
Opus 5.5 always thinks before it answers. You cannot switch thinking off, and the reasoning tokens are billed at the output rate of $20 per million. The default effort level is medium. A hard coding question at high effort can spend thousands of thinking tokens before the first visible word, so a reply of 500 words may bill for far more output tokens than you can see.
Two more details change old estimates. Opus 5.5 uses the tokenizer introduced with Opus 4.7, which can produce up to about 35% more tokens than older Claude models for the same text. If your budget came from Claude 4.6 numbers, re-measure with the token counting endpoint before you trust it.
Rates mean little until you attach them to a task. Every figure in this section uses the standard $4 and $20 rates, with no caching and no batch discount.
| Task | Input tokens | Output tokens | Cost |
|---|---|---|---|
| One chat answer | 2,000 | 800 | $0.024 |
| 2,500-word article draft | 1,500 | 3,300 | $0.072 |
| Pull request review | 30,000 | 3,000 | $0.18 |
| Long contract review | 200,000 | 4,000 | $0.88 |
Even the contract review, with 200,000 tokens of input, stays under a dollar. The surprise is rarely a single call. It is the thousand calls nobody counted, and the thinking tokens that ride along with each one.
An agent that edits files and runs tests re-reads its growing context on every step. Picture a 40-step session that consumes 500,000 input tokens in total and writes 40,000 output tokens:
Now cache the repeated context so 90% of the input is read from cache. The 450,000 cached tokens cost $0.09, the 50,000 fresh tokens cost $0.20, and output stays at $0.80. The session drops to about $1.09, plus a one-time cache write on the first request.
A quick method keeps surprises small:
Try it on an app that makes 20,000 requests a month at 3,000 input and 700 output tokens each. Input is 60 million tokens, which costs $240. Output is 14 million tokens, which costs $280. The base bill is $520, and the 30% buffer brings the planning number to $676.
Opus 5.5 charges $4 and $20 where Opus 5 charges $5 and $25, a 20% cut on both sides. Cache reads fall from $0.50 to $0.20 per million tokens, a 60% cut. The context window, the output ceiling and the tokenizer stay the same, so for a team already on Opus 5 the price drop alone is a reason to test the switch. A team spending $5,000 a month on Opus 5 would pay about $4,000 for the same token volume on Opus 5.5, a saving of $1,000 a month or $12,000 a year, before any caching or batch discount.
Check two breaking changes first: thinking cannot be disabled on Opus 5.5, and forcing a specific tool with tool_choice returns an error. Code written for older Opus models may need small edits.
To compare the family fairly, take one workload: a support desk answering 1,000 tickets a day, with 1,500 input tokens and 400 output tokens each, over 30 days.
| Model | Input and output per MTok | Monthly bill |
|---|---|---|
| Claude Fable 5.1 | $10 and $50 | $1,050 |
| Claude Opus 5 | $5 and $25 | $525 |
| Claude Opus 5.5 | $4 and $20 | $420 |
| Claude Sonnet 5.5 | $2 and $10 | $210 |
| Claude Haiku 4.5 | $1 and $5 | $105 |
Opus 5.5 sits $105 below Opus 5 every month on this workload, and a Batch run would halve it to $210.
Price per call is not price per finished task. A cheaper model that needs two retries costs more than a stronger one that gets it right the first time. Still, the split is clear in practice:
Test your own prompts on all three, then compare cost per correct result instead of cost per request.
Say you ask ten questions about the same 100,000-token contract. Without caching, each request pays $0.40 for that input: $4.00 across ten. With a 5-minute cache, the first request writes the contract at $5 per million ($0.50) and the next nine read it at $0.20 per million ($0.02 each). The input total is $0.68.
Put the stable material first (instructions, documents, tool definitions) and the changing question last, because caching works on an exact prefix. One edited word near the top invalidates everything after it.
The Batch API takes requests in bulk and returns results asynchronously, usually within a day, at 50% of standard prices: $2 input and $10 output per million tokens. Anything nobody is waiting for qualifies: nightly summaries, catalog descriptions, dataset labelling, evaluations. The support-desk example drops from $420 to $210 a month if the answers can wait, and Batch can be combined with caching for a deeper cut.
Effort is the cheapest lever because it cuts thinking tokens without touching the prompt. The default on Opus 5.5 is medium. Chat, short rewrites and simple extraction rarely need more than low, while multi-file refactors and tricky debugging earn high. Measure on a sample of your real requests before changing a default, and pair the setting with a max_tokens ceiling.
💡 Fast mode doubles the price. It runs the same model at up to 2.5 times the output speed for $8 input and $40 output per million tokens. It is a research preview on the Anthropic API only, and it cannot be combined with the Batch API. Use it only where a person is waiting on the screen.
Let us be direct. As of early October 2026 there is no free tier for Opus 5.5 on the API, and the free claude.ai plan does not include it. Every working path has a price tag or a catch worth knowing about.
The free Claude plan gives you chat with Sonnet and Haiku class models, inside daily usage caps. Opus 5.5 sits behind a paid plan. That matches how the API works: pay-as-you-go from the first request, with prepaid credits and no permanent free allowance. Plans and limits change often, so read the plan page on Claude's site before paying for anything.
Claude Pro costs $20 per month and Claude Max starts at $100 per month. Both open Opus models inside the chat app, with usage limits instead of per-token billing. A rough comparison: at API rates, $20 buys about 830 chat answers of the size priced earlier (2,000 tokens in, 800 out). A casual user usually comes out ahead on a plan, while an app that makes thousands of calls needs the API.
Cloud credits are the other legitimate discount. Amazon Bedrock, Google Cloud Vertex AI and Microsoft Foundry all serve Claude, and a company that already holds credits with one of them may be able to spend those credits there. Bedrock and Vertex AI publish their own rates, so check them instead of assuming the figures in this article apply.
Search for free Opus 5.5 and you will find pages that promise unlimited access with no sign-up. Treat them with suspicion. Many are wrappers that resell a small quota, some route you to a cheaper model without saying so, and some collect whatever you paste. Never send contracts, source code or customer data through a site you cannot vet. A few cents per request on the official API is cheaper than a leak.
If you only need Opus-level answers a few times a week, the cheapest honest path is a small prepaid API balance. At about 2.4 cents per chat answer, every $10 buys roughly 400 answers of the size priced earlier.
The short version:
Opus 5.5 itself is not in the PicassoIA catalog at the time of writing. The closest Opus model is Claude Opus 4.7, and the same collection lists Claude Fable 5, Claude Sonnet 5, Claude Sonnet 4.6 and Claude 4.5 Haiku. You open a page, type, and read the answer. There is no developer account to configure and no billing dashboard to watch.
| Setting | Default | What it does |
|---|---|---|
| Prompt | Required | Your question or task |
| System Prompt | Empty | Sets role, tone and format rules |
| Max Tokens | 8,192 | Caps the length of the answer |
| Image | None | Adds a screenshot or photo |
| Max Image Resolution | 0.5 megapixels | Scales the image down before sending, saving time and money |
Three habits make the answers better. Give a role in the system prompt, such as "You are a careful reviewer of pricing pages." Ask for a table when you want a comparison. Lower Max Tokens when you only need a short reply.
💡 Want zero cost? The catalog also lists open models such as GPT OSS 120B and Llama 4 Maverick Instruct, both listed as free to try. Quality and limits differ from Opus, so test them on your own prompts.
A language model earns its place in a visual workflow too. Ask Claude for twenty detailed image prompts of about 120 words each. That is roughly 3,200 output tokens, or about 6 cents at the Opus 5.5 output rate, and the prompts carry the lens, lighting and texture detail that make generated photos look real.
A request that works well looks like this: "Write a 70-word photographic prompt for a morning kitchen scene. Include the subject, the environment, the direction of the light, the camera lens, and two surface textures. No text in the image." Claude returns a dense paragraph you can paste straight into an image model, and a second pass costs only a few more cents.
Then move to the image and video models on PicassoIA:
Using one tool for words and another for pictures keeps each cost visible. You pay cents for the prompt and let the image or video model do the heavy lifting.
You now have the Opus 5.5 price, the point where the free routes end, and three ways to stretch every token. The fastest way to put that to work is hands-on. Open PicassoIA, ask a Claude model to draft a photographic prompt for a scene you care about, and run it through Seedream 4.5. Then take the best frame into a video model and watch it move. Browse everything else on all PicassoIA models, and keep experimenting until the prompts feel like yours.
