Claude Opus API pricing: Anthropic, September 2026
Opus 5.5 is the cheapest of the eight Opus versions Anthropic prices, at $4 input and $20 output per million tokens (checked 28 September 2026).
Claude Opus 5.5 costs $4 per million input tokens and $20 per million output tokens through Anthropic’s API, with cache reads at $0.20 per million.1
Those figures come from Anthropic’s published pricing, checked 28 September 2026.
Opus 5.5 is the cheapest of the eight Opus versions Anthropic prices. The oldest two, Opus 4.1 and Opus 4, top all 18 Claude models on output price at $75 per million tokens.1
That’s almost four times the current Opus rate, for the same model family.
TL;DR
- Opus 5.5 is the cheapest Opus. It costs $4 input, $20 output, $5 for a 5-minute cache write, $8 for a 1-hour cache write and $0.20 for a cache read, per million tokens.1
- Five older versions share one card; the two oldest cost three times that. Opus 5, Opus 4.8, Opus 4.7, Opus 4.6 and Opus 4.5 each cost $5 input and $25 output. Opus 4.1 and Opus 4 cost $15 and $75.1
- Against Sonnet and Haiku, Opus 5.5 costs twice Sonnet 5 and four times Haiku 4.5 per input and output token. On cache reads it matches Sonnet 5 at $0.20.1
- Batch processing takes 50% off input and output, which puts Opus 5.5 at $2 and $10 per million tokens.1
- Caching decides an agent’s bill more than the model tier does. In Glamdring Research’s worked 100-call task, Opus 5.5 costs $5.60 with caching and $35.60 without it, against $9.00 on Opus 5 and $3.60 on Sonnet 5.
- Same price doesn’t mean same bill. Opus 4.7 and later count roughly 30% more tokens for the same text than Opus 4.6 and earlier, at identical rates.
How this ranking was built
This ranking belongs to our model infrastructure coverage and follows the publication’s research standard for captured and computed figures.
- Source: Anthropic’s pricing documentation at docs.claude.com, captured 28 September 2026 at 15:05 AEST, cross-checked against Anthropic’s summary pricing page captured the same minute. Neither page shows a release date, so the checked date stands in for one.
- Field: the documentation’s Output price under “Base tokens”, in US dollars per million tokens (the page writes “MTok”). Every model’s output price is five times its input price, so ranking by input gives the same order.1
- Currency: the documentation states that all prices are in USD.1
- Tax: the documentation makes no tax statement for direct API billing.
- Inclusion: every row in the documentation’s model pricing table, 4 current models plus 14 under “Additional models”. 18 captured, 18 ranked, no merges.
- Cross-check: the summary page shows 12 of the 18, split into “Latest models” and “Legacy models”, and every figure it shows matches the documentation. It omits Mythos 5.1, Mythos 5, Opus 4.1, Opus 4, Sonnet 4 and Haiku 3.5.
- Exclusions: subscription plans; charges billed per search, per container-hour or per session-hour; the fast mode and US-only multipliers, which the caching and version sections cover.
- Ties: tied models share a rank and appear in the documentation’s order.
These are Anthropic’s list prices: what Anthropic says it charges, before any negotiated discount.
Claude API models ranked by output price
| Rank | Model | Summary page status | Output price (US$ per million tokens) |
|---|---|---|---|
| 1= | Claude Opus 4.1 | Not listed | 75 |
| 1= | Claude Opus 4 | Not listed | 75 |
| 3= | Claude Fable 5.1 | Latest | 50 |
| 3= | Claude Mythos 5.1 | Not listed | 50 |
| 3= | Claude Fable 5 | Legacy | 50 |
| 3= | Claude Mythos 5 | Not listed | 50 |
| 7= | Claude Opus 5 | Legacy | 25 |
| 7= | Claude Opus 4.8 | Legacy | 25 |
| 7= | Claude Opus 4.7 | Legacy | 25 |
| 7= | Claude Opus 4.6 | Legacy | 25 |
| 7= | Claude Opus 4.5 | Legacy | 25 |
| 12 | Claude Opus 5.5 | Latest | 20 |
| 13= | Claude Sonnet 4.6 | Legacy | 15 |
| 13= | Claude Sonnet 4.5 | Legacy | 15 |
| 13= | Claude Sonnet 4 | Not listed | 15 |
| 16 | Claude Sonnet 5 | Latest | 10 |
| 17 | Claude Haiku 4.5 | Latest | 5 |
| 18 | Claude Haiku 3.5 | Not listed | 4 |
Output prices exactly as Anthropic’s model pricing table lists them.1
Among the four current models, output price falls at each step down the line: $50 for Fable 5.1, $20 for Opus 5.5, $10 for Sonnet 5 and $5 for Haiku 4.5.1
The older rows run the other way. Every older Opus costs more than Opus 5.5, and every older Sonnet costs more than Sonnet 5. Opus 4.1 and Opus 4 cost more than any Fable or Mythos model. Only Haiku 3.5, at $4, undercuts its successor.1
What does each Opus version cost?
Anthropic prices eight Opus versions on its API. Opus 5.5 costs $4 per million input tokens and $20 per million output tokens. Opus 5, Opus 4.8, Opus 4.7, Opus 4.6 and Opus 4.5 each cost $5 and $25. Opus 4.1 and Opus 4 cost $15 and $75.1
| Opus version | Input | Output | 5-min cache write | 1-hour cache write | Cache read | Batch input | Batch output |
|---|---|---|---|---|---|---|---|
| Opus 5.5 | $4 | $20 | $5 | $8 | $0.20 | $2 | $10 |
| Opus 5 | $5 | $25 | $6.25 | $10 | $0.50 | $2.50 | $12.50 |
| Opus 4.8 | $5 | $25 | $6.25 | $10 | $0.50 | $2.50 | $12.50 |
| Opus 4.7 | $5 | $25 | $6.25 | $10 | $0.50 | $2.50 | $12.50 |
| Opus 4.6 | $5 | $25 | $6.25 | $10 | $0.50 | $2.50 | $12.50 |
| Opus 4.5 | $5 | $25 | $6.25 | $10 | $0.50 | $2.50 | $12.50 |
| Opus 4.1 | $15 | $75 | $18.75 | $30 | $1.50 | $7.50 | $37.50 |
| Opus 4 | $15 | $75 | $18.75 | $30 | $1.50 | $7.50 | $37.50 |
All figures are US dollars per million tokens.1
At the 28 September 2026 check, Opus 5.5 undercuts every other Opus on every line. Against the five-version card it saves 20% on input, output and both cache writes, and 60% on cache reads.
The cache-read gap comes from a different multiplier as well as a lower base price. Most Claude models charge 10% of the input price for a cache hit, and Opus 5.5 charges 5%.1
Identical rates don’t produce identical bills. Claude 4.7 and later models use a newer tokenizer that produces approximately 30% more tokens for the same text, and the documentation says the exact increase depends on the content and workload. Claude Sonnet 4.6 and earlier models use the previous tokenizer.1
So Opus 4.6 and Opus 4.7 charge the same $5 per million input tokens, but the same document sent to Opus 4.7 can count about 30% more tokens. The same arithmetic trims Opus 5.5’s lead over Opus 4.6: 30% more tokens at a 20% lower rate comes to about 4% more for identical text. Against Opus 5 and Opus 4.8, which share its tokenizer, the 20% cut holds in full.
Versions also differ on fast mode. Opus 5.5 runs in fast mode at $8 input and $40 output, and Opus 5 and Opus 4.8 at $10 and $50. Opus 4.7 returns an error on fast mode requests, and Opus 4.6 runs them at standard speed and standard rates.1
Anthropic’s summary pricing page lists Opus 5 under “Legacy models” at $5 input and $25 output.2
That page also lists Opus 4.8, Opus 4.7, Opus 4.6 and Opus 4.5 as legacy models. Opus 4.1 and Opus 4 appear only in the documentation. Neither page says whether they remain open to new API accounts, so confirm availability before planning on either.
How does Opus compare with Sonnet and Haiku?
Opus 5.5 costs twice as much as Sonnet 5 and four times as much as Haiku 4.5 per input and output token. Sonnet 5 costs $2 and $10 per million tokens and Haiku 4.5 costs $1 and $5, against $4 and $20 for Opus 5.5. The gap narrows on cached context: Opus 5.5 and Sonnet 5 both charge $0.20 per million cache-read tokens, and Haiku 4.5 charges $0.10.1
| Model | Input | Output | 5-min cache write | Cache read | Input and output vs Opus 5.5 |
|---|---|---|---|---|---|
| Fable 5.1 | $10 | $50 | $12.50 | $0.25 | 2.5× |
| Opus 5.5 | $4 | $20 | $5 | $0.20 | 1× |
| Sonnet 5 | $2 | $10 | $2.50 | $0.20 | 0.5× |
| Sonnet 4.6, 4.5, 4 | $3 | $15 | $3.75 | $0.30 | 0.75× |
| Haiku 4.5 | $1 | $5 | $1.25 | $0.10 | 0.25× |
| Haiku 3.5 | $0.80 | $4 | $1 | $0.08 | 0.2× |
US dollars per million tokens. The ratio column is computed from Anthropic’s figures.1
The premium tiers are priced to reward caching. Fable 5.1 costs 2.5 times Opus 5.5 per fresh token but 1.25 times per cached read. Opus 5.5 costs twice Sonnet 5 per fresh token and the same per cached read.
The decision changes with the share of a workload that re-reads cached context. A chat that sends fresh text and returns long answers pays the full 2× premium for Opus 5.5 over Sonnet 5. An agent that re-reads a large codebase on every turn pays far less than 2×, because its biggest line is priced the same on both.
Anthropic’s own guidance is to choose Haiku for simple tasks, Sonnet for most production workloads and Opus for the most complex reasoning.1
These ratios hold only when two models spend the same tokens on the same job. A cheaper model that needs more calls, longer outputs or retries closes the gap, and the rate card publishes no token-efficiency data. For what drives inference cost beyond the list price, see our explainer on what controls AI inference cost.
What do prompt caching and batch processing cost?
On Opus 5.5, a 5-minute cache write costs $5 per million tokens, a 1-hour cache write $8 and a cache read $0.20, against $4 for ordinary input.1
The Batch API takes 50% off both input and output tokens. That puts Opus 5.5 at $2 input and $10 output per million tokens, and every Opus from Opus 5 to Opus 4.5 at $2.50 and $12.50.1
| Cache operation | Multiplier on input price | Opus 5.5 | Opus 5 to Opus 4.5 | Opus 4.1 and Opus 4 |
|---|---|---|---|---|
| 5-minute write | 1.25× | $5 | $6.25 | $18.75 |
| 1-hour write | 2× | $8 | $10 | $30 |
| Read (hit) | 0.05× on Opus 5.5, 0.1× on the others | $0.20 | $0.50 | $1.50 |
Multipliers from Anthropic’s prompt caching terms; dollar figures from its model pricing table, in US dollars per million tokens.1
A 5-minute cache on Opus 5.5 pays for itself on the first re-read. The write costs $1 per million tokens more than plain input, and each later read saves $3.80. A 1-hour write costs $4 more, so it needs two reads to break even.
Anthropic’s documentation gives the same break-even points for models on the standard 10% read rate: one read for the 5-minute cache and two reads for the 1-hour cache.1
The constraint is the cache lifetime. A prefix that isn’t read again inside five minutes has to be written again at the premium rate, which is the case for paying for the 1-hour write.
Anthropic states that the caching multipliers stack with other pricing modifiers, including the Batch API discount and data residency.1
Anthropic’s pricing page describes the Batch tier as for asynchronous workloads that can be processed together for better efficiency.2
An agent that needs each answer before its next step can’t sit in a batch queue, so batch pricing suits backlogs and overnight runs rather than interactive agents.
Two limits are explicit. Fast mode is not available with the Batch API, and the Batch API discount doesn’t apply to Claude Managed Agents sessions, because sessions are stateful and interactive.1
For Claude 4.6 and later models, US-only inference through the inference_geo parameter adds a 1.1x multiplier to all token categories, including input tokens, output tokens, cache writes and cache reads.1
On Opus 5.5 that makes $4.40 input, $22 output and $0.22 per cache read.
Claude 4.6 and later models include the full 1M-token context window at standard pricing, so a 900k-token request is billed at the same per-token rate as a 9k-token request.1
What does one long agentic task cost on Opus?
Glamdring Research’s worked estimate puts one long agentic task at $5.60 on Opus 5.5, against $9.00 on Opus 5 or any Opus from 4.5 to 4.8, $27.00 on Opus 4.1, $3.60 on Sonnet 5 and $1.80 on Haiku 4.5, at Anthropic’s published pricing, checked 28 September 2026. The task is illustrative: 100 model calls, each re-reading an 80,000-token cached context, adding 4,000 new tokens and generating 1,000. Without prompt caching, the same Opus 5.5 run costs $35.60.
Anthropic’s own worked example is smaller. It prices a one-hour Claude Managed Agents coding session on Opus 5 that uses 50,000 input tokens and 15,000 output tokens at $0.705, including $0.08 of session runtime. With 40,000 of the input tokens served as cache reads, the total falls to $0.525.1
| Model | No caching | 40,000 input tokens cached |
|---|---|---|
| Opus 5.5 | $0.580 | $0.428 |
| Opus 5 (Anthropic’s figures) | $0.705 | $0.525 |
| Opus 4.1 | $1.955 | $1.415 |
| Sonnet 5 | $0.330 | $0.258 |
| Haiku 4.5 | $0.205 | $0.169 |
US dollars. The Opus 5 row is Anthropic’s; the other rows apply the same token counts and the same $0.08 runtime to each model’s rates. Anthropic’s example lists no cache-write line, so read it as a session whose cache was already warm.
A session that small hides the effect of scale. The 100-call estimate assumes:
- 100 model calls in one task, the shape of a coding or research agent that reads files, runs tools and edits across many turns.
- 8 million cache-read tokens: each call re-reads an average 80,000-token cached context of instructions, files and history.
- 400,000 cache-write tokens at the 5-minute rate: each call adds 4,000 new tokens of tool results and file content to the cache.
- 100,000 output tokens: each call generates 1,000 tokens, including any reasoning tokens.
- Standard pricing, no batch, every read landing inside the 5-minute window, and the same token counts on every model.
| Model | Cache reads (8M tokens) | Cache writes (0.4M tokens) | Output (0.1M tokens) | Task total |
|---|---|---|---|---|
| Fable 5.1 | $2.00 | $5.00 | $5.00 | $12.00 |
| Opus 5.5 | $1.60 | $2.00 | $2.00 | $5.60 |
| Opus 5 to Opus 4.5 | $4.00 | $2.50 | $2.50 | $9.00 |
| Opus 4.1, Opus 4 | $12.00 | $7.50 | $7.50 | $27.00 |
| Sonnet 5 | $1.60 | $1.00 | $1.00 | $3.60 |
| Haiku 4.5 | $0.80 | $0.50 | $0.50 | $1.80 |
US dollars, computed from Anthropic’s per-million-token rates.
Three results follow from the arithmetic. On identical token counts, Opus 5.5 finishes the task for 38% less than Opus 5, and the biggest single saving is the cache-read line, which falls from $4.00 to $1.60.
Opus 5.5 costs 1.56 times Sonnet 5 on this task, well short of the 2× its token rates imply. Both charge $1.60 for the same 8 million cached reads, so the gap sits only in writes and output.
Caching is the largest lever of all. Sent as plain input, the 8.4 million context tokens cost $33.60 on Opus 5.5, and output adds $2.00. Caching cuts that $35.60 run by 84%.
Tools add input on every call. On Opus 5.5, any tool definition brings a tool-use system prompt of 286 tokens with tool choice set to auto or none.1
Web search costs $10 per 1,000 searches, plus standard token costs for the content it returns.1
The estimate holds token counts fixed, and that is its main limit. ModemGuides reports Anthropic’s estimate that Opus 5.5 at default settings costs 40% less than Opus 5 on typical workloads, because it needs fewer tokens and steps. It also reports Artificial Analysis measuring Opus 5.5 at about 1.6 times Opus 5’s output tokens per task at maximum effort, which left cost per task level with the older model.3
In this model, 1.6 times the output alone lifts Opus 5.5 to $6.80, still below Opus 5’s $9.00. A level result means those runs changed more than output volume. Effort settings move the answer as much as the rate card does.
Opus 5.5 can’t run with thinking disabled, so every response carries reasoning tokens that API users pay for.3
Use one real workflow to test the claim. Log one agent run’s cache reads, cache writes, uncached input and output as separate lines, then price each line at the rates in this article. That breakdown shows which model is cheaper for the job; the headline rate doesn’t.
Why do other published Opus prices disagree?
Opus prices that differ from Anthropic’s usually come from resellers and cloud listings that apply their own discount or multiplier to Anthropic’s list rates, or from pages written before Opus 5.5 launched. Three such pages captured on 28 September 2026 show the pattern: one reseller advertises up to 30% off, another 5% off, and a Bedrock calculator lists Opus 4.6 at 1.1 times Anthropic’s rate.
LLM.API’s Claude Opus 4.8 page lists Anthropic’s own price at $5 in and $25 out, and offers the same model at up to 30% below list price.4
OpenModel lists claude-opus-4-6 at $4.75 input, $23.75 output, $5.9375 cache write and $0.475 cache read, labelled 5% off, and a lower price available only via Claude Code.5
Each OpenModel API figure is 0.95 times Anthropic’s Opus 4.6 card. Those are resellers’ statements about their own channels. Confirm them in the current quote.
Holori’s calculator lists the Bedrock model us.anthropic.claude-opus-4-6-v1 at $5.50 input, $27.50 output and $0.550 cached input per million tokens.6
The same listing prices cache creation at $6.8750 per million tokens, and $11.0000 above one hour.6
Every one of those Bedrock figures is exactly 1.1 times Anthropic’s own Opus 4.6 rate, the same multiple Anthropic applies to US-only inference on its API.
Anthropic’s documentation says partner-operated platforms, Bedrock and Google Cloud, have independent regional pricing.1
Dates cause the rest. ModemGuides dates Opus 5.5’s release to 22 September 2026.3
A page written before that date can only quote an older Opus as the current model, at $5 and $25.
Subscription prices add more confusion. Claude Pro costs $17 a month with an annual subscription or $20 billed monthly, and Max starts from $100 a month.2
Those plans carry usage limits, not per-token rates, so they don’t compare directly with API pricing.
What changed with Opus 5.5?
Opus 5.5 cut every Opus token rate against Opus 5. Input and output fell 20%, from $5 and $25 to $4 and $20 per million tokens. Cache reads fell 60%, from $0.50 to $0.20, and cache writes fell from $6.25 to $5.3
Anthropic’s pricing documentation, checked 28 September 2026, carries the same rates for both models, and the 1-hour cache write fell from $10 to $8.
The fast mode rate fell with it: $8 input and $40 output on Opus 5.5, against $10 and $50 on Opus 5 and Opus 4.8.1
Opus 5 stayed on the summary pricing page as a legacy model at its old rates.2
The cache-read cut is the one that reaches agent bills. ModemGuides reports Anthropic stating that cache reads make up the majority of agentic and coding costs, because a coding agent re-reads the same repository context and instructions on every turn.3
That is Anthropic’s claim about its customers’ workloads, not a measured market figure. The worked 100-call task shows the mechanism either way: cache reads are the largest line on Opus 5 and fall by $2.40 on Opus 5.5.
Opus 5.5 also hands some requests to older models. ModemGuides reports Anthropic’s announcement saying that most cybersecurity tasks will be re-routed to Opus 4.8, and that biology and frontier-LLM-development requests fall back to Opus 5.3
The same report says the launch page doesn’t state whether a rerouted request is billed at the Opus 5.5 rate.3
On the rate card that is the gap between $4 and $5 per million input tokens. Teams in those fields should confirm billing for rerouted requests on their own invoices before modelling cost.
The Sonnet and Haiku comparison is due to move. ModemGuides reports that Sonnet 5.5 and Haiku 5.5 follow in the coming weeks with many of the same efficiency changes.3
Until Anthropic publishes their rates, Sonnet 5 at $2 and $10 and Haiku 4.5 at $1 and $5 remain the comparison points.
What the ranking cannot tell you
- What a buyer pays. These are list prices. Anthropic says volume discounts may be available for high-volume users, negotiated case by case.1
- What a task costs. Cost per task depends on token counts, effort settings, the tokenizer and cache hit rates. The rate card shows none of them.
- Partner-cloud prices. Bedrock and Google Cloud set their own rates. This ranking covers Anthropic’s own API.
- Tax. The documentation makes no tax statement for direct API billing.
- Which model answers. Rerouted requests, and how they’re billed, sit outside the rate card.
- Quality. Price ranks cost, not capability. A cheaper model that fails a task costs the attempt and the rerun.
This ranking changes when Anthropic publishes a new rate card, with Sonnet 5.5 and Haiku 5.5 the next expected. New research arrives through our email briefing.
Frequently asked questions
Does the Claude API cost money?
Yes. Billing is based on actual monthly usage, and new users receive a small amount of free credits to test the API.1
Input rates run from $0.80 per million tokens on Haiku 3.5 to $15 on Opus 4.1 and Opus 4, with Opus 5.5 at $4.1
Some tools carry their own charges. Web search costs $10 per 1,000 searches on top of standard token costs.1
How much does it cost to access the Opus 5 API?
Opus 5 costs $5 per million input tokens, $25 per million output tokens, $6.25 per million tokens for a 5-minute cache write and $0.50 per million for a cache read. Batch processing halves input and output to $2.50 and $12.50.1
Opus 5.5 charges $4 and $20 for the same input and output tokens, and Anthropic’s summary pricing page lists Opus 5 as a legacy model.1,2
Is Opus cheaper through a Claude subscription than through the API?
The two don’t compare directly. Pro includes Opus and costs $17 a month with an annual subscription or $20 billed monthly, and Max starts from $100 a month.2
Anthropic publishes usage limits for those plans rather than a token allowance, so no per-token comparison is possible.
For heavy coding sessions, Anthropic’s pricing page says Claude Code users can switch to pay-as-you-go API credits through a Console account.2
Sources checked
Company-owned pages establish what a company says. They do not prove a market conclusion. Each source is dated so readers can judge each claim.
