Claude Opus 5 Pricing: What It Costs and What Actually Drives Your Bill

Anthropic released Claude Opus 5 on July 24, 2026 at exactly the price of the model it replaced. That sounds like a non-event for anyone budgeting a migration. It is not, and the reasons have almost nothing to do with the rate card.

Claude Opus 5 pricing is $5 per million input tokens and $25 per million output tokens. Those two numbers are the easy part. What decides your actual invoice is the effort setting, how much of your context repeats, whether thinking tokens are running by default, and a short list of charges that never appear in a per-token comparison table. This guide covers the published rates first, then the arithmetic that turns them into a monthly number.

Claude Opus 5 pricing: $5 and $25 per million tokens

Every published rate for Opus 5 on the first-party Claude API sits in one table.

Claude Opus 5 rate card, pricing as of July 2026. MTok = million tokens. Source: Anthropic.

The model ID is claude-opus-5. It has a 1M token context window as both the default and the maximum, with no smaller variant, plus up to 128k output tokens. Anthropic bills the entire window at standard rates. A 900k-token request costs the same per token as a 9k-token request, so there is no long-context surcharge to plan around.

Four things get billed, and thinking tokens are the one people miss

A single Opus 5 API call can generate four separate billable quantities. Most cost estimates account for two.

You pay for the prompt you send, including the tokens consumed by any tool definitions in the request. You pay for the visible answer. You also pay for thinking tokens, which bill at the output rate of $25 per million. And you pay for tool results returned into the conversation, which enter as input on the following turn.

That third item is where migrations go wrong. On Opus 4.8, a request that omitted the thinking field ran without extended thinking. On Opus 5, adaptive thinking runs by default. If you port a request body straight across and change only the model string, you will start paying for reasoning tokens you were not paying for last week.

One related setting needs a second look. max_tokens is a hard cap on total output, thinking plus response text together, so a value tuned for a no-thinking workload on Opus 4.8 will now truncate. Opus 5 also rejects a disabled-thinking request at xhigh or max effort with a 400 error.

The effort setting moves your bill more than the model choice does

Opus 5 exposes a five-level effort ladder: low, medium, high, xhigh, and max. It defaults to high on the Claude API and in Claude Code, and it is the cost control most teams never touch.

Effort trades intelligence against latency and token consumption inside one model. Anthropic's own launch material frames the performance gains as a curve across effort levels rather than a single score, and reports that at max effort on CursorBench 3.2, Opus 5 lands within 0.5% of Fable 5's peak score at half the cost per task.

Because the default is high rather than max, the saving is in stepping down, not in avoiding the top tier. Anthropic's guidance is to start at high and move in either direction based on your own evals, and it describes low and medium effort as producing strong quality at a fraction of the tokens and latency of higher settings. One legal AI team reported Opus 5 holding quality at lower reasoning levels while generating 26% fewer tokens on average than Opus 4.8 at max reasoning. Treat that as a vendor-published customer report rather than an independently verified benchmark.

One catch if you plan to tune effort per request. Changing the effort value invalidates cached message blocks, so switching it request by request can cost you the cache savings covered below. Setting effort explicitly to the model's default is the exception and invalidates nothing.

Caching and batch cut the rate, fast mode doubles it

Three features change what you pay per token. Two reduce it, one increases it.

Prompt caching charges a write price when a prefix is stored and a much lower read price when it is reused. Writes cost 1.25x the base input rate for a 5-minute cache and 2x for a 1-hour cache. Reads cost 0.1x, which is $0.50 per million on Opus 5. The break-even is fast: a 5-minute cache pays for itself after one read, a 1-hour cache after two. For anything with a stable system prompt, a fixed set of tool schemas, or a repository map that does not change between turns, caching is close to free money. Opus 5 also dropped the minimum cacheable prompt to 512 tokens from 1,024 on Opus 4.8, so prompts that were too short to cache last month can create entries with no code changes.

The Batch API halves both input and output for asynchronous work, putting Opus 5 at $2.50 and $12.50 per million. The discount stacks with prompt caching multipliers, so a cached batch workload compounds both savings.

Fast mode is the one modifier that raises your bill. It runs Opus 5 at roughly 2.5 times default output speed for twice the price, $10 input and $50 output, and it applies across the full context window including requests over 200k input tokens. Two constraints matter for planning. Fast mode is available on the first-party Claude API only, not on Amazon Bedrock, Google Cloud, or Microsoft Foundry. And it cannot be combined with the Batch API, so a workload cannot be both fast and half price.

Three worked examples show where the money actually goes

Rate cards are abstract. Here is the arithmetic on three shapes of work.

1. A single request

A 20,000-token prompt returning a 3,000-token answer:

• Input: 20,000 × $5 / 1,000,000 = $0.10
• Output: 3,000 × $25 / 1,000,000 = $0.075
• Total: $0.175

Run the same request in fast mode and it costs $0.35, exactly double.

2. An agent session with caching

An agent with a 25,000-token stable prefix, meaning system prompt, tool schemas, and a repo map, reused across 30 turns.

Without caching, that prefix costs 30 × 25,000 = 750,000 input tokens, or $3.75.

With a 5-minute cache, you pay one write and 29 reads:

• Cache write: 25,000 × $6.25 / 1,000,000 = $0.156
• Cache reads: 29 × 25,000 × $0.50 / 1,000,000 = $0.363
• Total: $0.52

That is an 86% reduction on the repeated portion alone, before counting the per-turn variable input and output. Each read refreshes the cache lifetime at no additional cost, so an actively running session keeps the prefix warm without repaying the write.

3. An overnight batch job

Ten million input tokens and 2 million output tokens processed asynchronously:

• Standard rates: $50 + $50 = $100
• Batch rates: $25 + $25 = $50

The 50% discount applies cleanly to both sides.

Opus 5 costs half of Fable 5 and the same as Opus 4.8

Opus 5 sits second in Anthropic's public lineup, above Sonnet and below Fable.

Claude model pricing comparison, pricing as of July 2026. Source: Anthropic.

Two footnotes on that table. Sonnet 5's $2 and $10 rates are introductory and run through August 31, 2026, after which standard pricing of $3 and $15 takes effect. And Claude Mythos 5 shares Fable 5's specs and pricing but is not generally available, sitting in limited access under Project Glasswing.

There is a trap in reading that table against any pre-4.7 model. Anthropic documents that Claude 4.7 and later use a newer tokenizer producing approximately 30% more tokens for the same text, while Sonnet 4.6 and earlier use the previous one. So if you are migrating from Sonnet 4.6 at $3 per million input, the real multiple on your bill is not 1.67x. The same document becomes roughly 30% more tokens before a single rate is applied. Opus 5 against Opus 4.8 stays clean, since both sit on the newer tokenizer. What still moves there is volume, because thinking runs by default and responses tend to run longer.

Subscription plans cap context at 200k, so the 1M window is an API purchase

Subscription plans and API access are billed separately. Once you reach a plan limit, paid plans can enable usage credits that continue the work at standard API rates.

Claude subscription plans, pricing as of July 2026. Source: Anthropic.

The context window on subscription plans does not match the API. Free, Pro, Max 5x, Max 20x, and Team all cap at 200k tokens. Enterprise gets 500k on the default model. The 1M window that makes Opus 5 attractive for large-codebase and long-document work is an API capability, not something a $200 Max subscription unlocks. If context length is why you want Opus 5, the API is the purchase.

One smaller correction worth making. Team plans have no five-seat minimum. Anthropic's own pricing page describes Team as built for teams of 2 to 150, though several third-party pricing roundups still repeat an outdated figure.

Six charges that never show up in a per-token comparison

• Web search costs $10 per 1,000 searches on top of the tokens the results consume.
• Claude Managed Agents adds $0.08 per session-hour of active runtime, metered only while a session is running rather than idle.
• US-only inference applies a 1.1x multiplier across input, output, cache writes, and cache reads when you set inference_geo to us.
• Tool use adds a system prompt of 286 tokens on Opus 5 for a tool choice of auto or none, and 406 tokens for any or tool. That is notably leaner than Opus 4.7, which needed 675 and 804.
• The bash tool adds a further 325 input tokens per request on top of the tool use system prompt.
• Code execution is free when used alongside web search or web fetch. On its own it bills by container hour after a monthly free allowance, with a 5-minute minimum per execution.

Web fetch adds no charge beyond the tokens of whatever it pulls into context, which is the one item on this list that works in your favour.

Sonnet 5 for most work, Opus 5 when the task earns it

Anthropic's own guidance is to start with Opus 5 for complex agentic coding and enterprise work, and move up to Fable 5 only where you need the highest available capability. The pricing supports that, since Opus 5 gives you second-in-lineup capability at half of Fable 5's rate on a model built for daily use rather than reserved for hard problems.

For most production workloads, though, Sonnet 5 is the honest default until a task demonstrably fails on it. High-volume classification, extraction, routing, and summarization belong on Haiku 4.5 at $1 and $5, and paying Opus rates for that work is waste. The test is not which model scores highest on a public benchmark. It is cost per finished task on your own workload, because a model that finishes in four attempts at $5 beats one that finishes in twenty at $2.

Budget for the effort setting, not the rate card

Claude Opus 5 pricing is $5 and $25 per million tokens, unchanged from Opus 4.8 and half of Fable 5. The rate card is also the smallest variable in your bill. Thinking runs by default and bills as output, the effort parameter swings token volume across a wide range inside a single model, caching cuts repeated input by 90%, and the newer tokenizer inflates counts by roughly 30% against pre-4.7 models. Set effort deliberately instead of accepting the default, cache every stable prefix, route asynchronous work through the Batch API, then measure cost per completed task on one real workload for a week. That number is the only one worth budgeting against.

If the goal is shipping an application rather than tuning a token bill, there is a shorter path. Claude Opus 5 is now available on Emergent, alongside OpenAI GPT and Google Gemini, all through a single Universal LLM Key with unified billing. Emergent builds and deploys the software around whichever model you pick: real backend, real integrations, real code you own. Skip the model comparisons and API setup. Describe your app and let Emergent handle the rest.

Start Building on Emergent.