Claude Opus 5 API: How to Call It, What It Costs, and When It's Overkill

The Claude Opus 5 API is live on the Messages endpoint and on every OpenAI-compatible gateway that carries it. To call it directly you POST to https://api.anthropic.com/v1/messages with the model id claude-opus-5 and the anthropic-version: 2023-06-01 header. To call it with an OpenAI-style client — the route that needs no new SDK — you point your base URL at https://api.orcarouter.ai/v1 and pass the model id anthropic/claude-opus-5, which is how OrcaRouter exposes it. The price is identical on both: $5 per million input tokens and $25 per million output, exactly the list rate with zero markup. It launched on July 24, 2026, and it is the current Opus tier — Claude Opus 4.8 costs the same $5 / $25 and scores lower on the independent benchmarks. The interesting decision is not which dialect to use; it is whether Opus-tier is what your workload actually needs.

Two endpoints, one model

Most tutorials for the Claude Opus 5 API show only the native Anthropic call, and they stop there. There are really two dialects, and both reach the same weights.

• Anthropic native — /v1/messages. POST to https://api.anthropic.com/v1/messages with two required headers: x-api-key with your key and anthropic-version: 2023-06-01. The body carries model: "claude-opus-5", a max_tokens, and a messages array. This is what the Anthropic Python and TypeScript SDKs send under the hood.

• OpenAI-compatible — /v1/chat/completions. POST to https://api.orcarouter.ai/v1/chat/completions (or /v1/messages if you want the Anthropic dialect through the same host) with the model id anthropic/claude-opus-5. If you already run an OpenAI client, the switch is two lines: change base_url to https://api.orcarouter.ai/v1 and swap the API key. Nothing else in your request shape changes.

• Cloud models. On Amazon Bedrock the id is anthropic.claude-opus-5; on Google Vertex AI and Microsoft Foundry it keeps the first-party id claude-opus-5. Those paths are for teams already committed to a cloud provider; the OpenAI-compatible route is the one that keeps your code portable.

The model does not care which dialect you use. What differs is everything around the call — billing, keys, rate limits, failover, and how many other models you can reach without changing code.

What you get — and the two things that break old code

Claude Opus 5 is Anthropic's reasoning-heavy flagship: a 1M-token context window (the maximum is also the default), a 128K output cap, a May 2026 knowledge cutoff, and text, image, and file input with text output. On the independent Artificial Analysis leaderboard it measured AA Coding 78.0, ranked #1 of 132, and AA Intelligence 63.1, ranked #2 of 134, with a GPQA Diamond of 93.2 and a Humanity's Last Exam score of 54.9, as of July 24, 2026. Anthropic's own launch claims — roughly double Opus 4.8 on Frontier-Bench — are vendor-run and not independently reproduced.

Thinking is on by default. Adaptive thinking decides how much to reason on each call, controlled by the output_config.effort parameter, which accepts low through max and defaults to high. For coding and agentic loops, xhigh is the recommended starting point; low and medium are genuinely stronger on Opus 5 than they were on prior models, which makes them viable for real workloads rather than toy calls. Reasoning tokens count toward your output bill.

Two breaking changes from Claude Opus 4.8 bite anyone who copies an old request body:

• Tight max_tokens now truncates. Because thinking runs by default, a max_tokens sized around expected answer length gets consumed by reasoning first. If you move a 4.8 workload over unchanged, expect cut-off responses until you raise the cap or drop the effort.

• Sampling parameters are rejected. temperature, top_p, and top_k at any non-default value return a 400 error, as does an assistant prefill. Control output with effort, not temperature.

• Disabling thinking is capped. thinking: {"type": "disabled"} is accepted only at effort high or below; pairing it with xhigh or max returns a 400. The sane way to cut cost is lowering effort, not switching thinking off.

What it actually costs

Claude Opus 5 lists at $5 per million input tokens and $25 per million output — the same sticker as Claude Opus 4.8, which is the quietly good news of the release. Cache reads bill at $0.50 per million (prompt caching needs a minimum 512-token prefix and supports 5-minute and 1-hour TTLs). A research-preview fast mode runs the same model at up to 2.5x the output tokens per second for $10 / $50, and the Batch API returns results asynchronously at half price.

A concrete invoice: a 100K-token request that produces 20K output tokens bills $0.50 for input and $0.50 for output — $1.00 uncached. If that 100K prompt is served from cache on the next call, the input leg drops to about $0.05, for a $0.55 round trip. That cache arithmetic is the difference between Opus being affordable on long, repeated context and it not being affordable at all.

The tier above and below, for scale: Claude Fable 5 is Anthropic's most capable model at $10 / $50; Claude Sonnet 5 is $3 / $15 (with $2 / $10 intro pricing through August 31, 2026); Claude Haiku 4.5 is $1 / $5. On OrcaRouter every one of these is passed through at the vendor's list price with no per-token markup, so the comparison above is also what you actually pay.

The one-endpoint case

If you are reading a Claude Opus 5 API guide, you probably already have an integration and a key, and the real question is whether you want a second integration and a second key for the next model you try. OrcaRouter answers that with one OpenAI-compatible endpoint in front of 200+ models: the same client, the same base URL, and a different model id per model — anthropic/claude-opus-5 today, a cheaper model when the workload stops earning Opus-tier, a new frontier model the week it ships, all without a code change.

The economics are the honest part. OrcaRouter adds $0 per token — you pay each provider's published rate, so Claude Opus 5 stays at Anthropic's $5 / $25, and a vendor price cut is live on our side the same day it is announced. Bring your own key is a first-class option: hand over your existing Anthropic API key, keep your own rate limits and credits, and the provider bills you directly. Guardrails — a PII shield and a content policy — are enforced before the request is billed, not after. The model page for anthropic/claude-opus-5 shows the live price, the 1M context, and the current throughput numbers, and the routing DSL composes a failover so that a cheaper model answers when Opus 5 errors or when your traffic shape lets it.

Where Opus 5 is the wrong answer

Being the best is not the same as being right for your workload, and there are three places the $5 / $25 flagship loses.

• High-volume, low-stakes calls. A summarization or classification pipeline doing 10M input tokens a month costs $50 on Claude Opus 5 against $10 on Claude Haiku 4.5 — same endpoint, same key, one id change. Haiku is not Opus, but for routing, extraction, and formatting it is close enough that the 5x is hard to justify.

• Long agent loops. Opus 5 verifies its own work unprompted, so a "double-check your answer" instruction that was tuned for an older model now triggers over-verification and burns reasoning tokens. At xhigh and max effort the thinking budget grows by design. If your agent does not need frontier reasoning, Claude Sonnet 5 at $3 / $15 is the default that most production teams should start from, and the effort dial is where you tune it down.

• Where the model is not ours. OrcaRouter routes and bills; Anthropic runs the weights. The endpoint, the zero markup, the failover, and the guardrails are ours, but inference quality and safety behavior are Anthropic's, and with BYOK the money goes straight to them. That is the deal stated plainly: you pay the provider rate either way, and the question is whether one endpoint for 200+ models is worth it.

Bottom line

Call the Claude Opus 5 API with the native Messages endpoint if you have no other integration, and with the OpenAI-compatible endpoint if you want one client for everything. It is the right model when the task is hard, the context is long, and the output is worth $25 per million tokens — and it is the wrong model when the task is routine. The cleanest setup is the one that lets you change your mind: one endpoint, your own key, $0 markup, and a model id you can swap the day the workload does.