Claude Opus 5 API Pricing: $5 / $25 per 1M Tokens, Cache at $0.50, and What It Really Costs

Claude Opus 5 costs $5.00 per million input tokens and $25.00 per million output tokens on the Claude Opus 5 API — the same rate the model carries on the OrcaRouter directory, checked 2026-08-18, at zero markup. That makes it half the price of Claude Fable 5 and exactly what Claude Opus 4.8 cost, and the rate card on the live Claude Opus 5 model page is the one to budget against. The levers matter more than the headline: reused input bills at $0.50 per million via prompt caching, batch drops to $2.50 / $12.50, and its fast-mode beta sits at $10 / $50. This page is the pricing reference for the model — the exact rate card, what a request actually costs, how the price compares to the field, when it is the wrong buy, and how to call it through one OpenAI-compatible endpoint with your own key.

Prices are per million tokens (input / output), as of 2026-08-18. The dollar figures are read from Anthropic's published API rates and cross-checked against the OrcaRouter model page for Claude Opus 5; cache, batch and fast-mode rates are Anthropic-announced. Prices move — verify before you build.

The exact price, verified today

Here is the full rate card for Claude Opus 5, verified 2026-08-18:

• Input — $5.00 per 1M tokens on a cache miss.

• Output — $25.00 per 1M tokens.

• Prompt cache write — $6.25 per 1M at a 5-minute TTL, $10.00 per 1M at a 1-hour TTL.

• Prompt cache read — $0.50 per 1M tokens, a 90% discount on any input you resend.

• Batch API — $2.50 input / $12.50 output, a flat 50% off, asynchronous.

• Fast mode — $10 input / $50 output, a 2x premium for up to roughly 2.5x output speed, Anthropic beta.

• Context window — 1M tokens, up to 128K output.

OrcaRouter passes the provider rate through with no markup, so the figure on the directory is the figure you pay. On the live model page the card reads $5.00 / $25.00 per 1M, cache read $0.500, cache write $10.00 (the 1-hour TTL rate), with Vision, Tools, JSON and Reasoning badges.

The levers that change the effective price

The base rate is only the starting point. Three modifiers change what Opus 5 actually costs you:

• Prompt caching — reused context bills at $0.50 per million instead of $5 fresh. The minimum cacheable prefix dropped to roughly 512 tokens on Opus 5, so even a modest stable system prompt qualifies. For any agent or chat that re-sends a large prefix on every turn, caching is the single biggest lever on the bill — and because Anthropic keys the cache per API key, staying on one key (one channel) is what keeps the hits landing.

• Batch — Anthropic's batch API runs jobs asynchronously, with results typically available within 24 hours, at exactly half price: $2.50 / $12.50. It stacks with prompt caching and suits evals, bulk classification and offline generation where latency does not matter.

• Fast mode — $10 / $50 for up to ~2.5x output throughput, useful for latency-critical high-output work. It does not combine with the batch API and is an Anthropic first-party beta; the standard pass-through rate from a router is the base $5 / $25.

What a request actually costs

Per-token prices are abstract, so here are worked examples at today's rates:

• Single chat call, 50K input + 20K output — $0.25 + $0.50 = $0.75.

• 1M input + 1M output — $5 + $25 = $30.

• Agentic turn, 200K cached context + 20K fresh input + 30K output — $0.10 + $0.10 + $0.75 = $0.95. The same turn with a cold cache — 220K fresh input at $1.10 — totals $1.85. Caching nearly halves it.

• A 10M-token month at a 70% input / 30% output split, with 80% of input served from cache: 1.4M fresh input ($7.00) + 5.6M cached ($2.80) + 3M output ($75) ≈ $85. Uncached, the same month is $110.

Two caveats raise the effective rate. Opus 5 is a thinking model — adaptive thinking is on by default, and thinking tokens bill as output at $25 per million. A hard prompt at a high effort setting can spend a large share of its output budget on reasoning, so the effective output cost on difficult work runs well above a naive estimate. And Anthropic's current tokenizer is more verbose than the previous generation's, so the same English text costs more tokens (and more money) than it did on older Claude models. Lower the effort setting for easy calls and cap output where you can.

How Opus 5 pricing compares to the field

Read against published list prices on 2026-08-18:

• Claude Opus 4.8 — $5 / $25. Identical price for the previous generation. Opus 5 is the same cost for the newer model, which is the easy case for upgrading.

• Claude Fable 5 — $10 / $50. Opus 5 is half the price of Anthropic's top tier, with the same 1M-token context.

• Claude Sonnet 5 — $3 / $15 standard ($2 / $10 intro through August 31, 2026). The Sonnet tier is 40% cheaper on input and output — the natural default when you don't need Opus-level reasoning.

• GPT-5.6 Sol — $5 / $30. Same input price, cheaper output on Opus 5.

• GPT-5.6 Terra — $2 / $12. The volume middle tier for high-volume non-frontier work.

• Gemini 3.6 Flash — $1.50 / $7.50. A Flash-tier model for cheap, fast, multimodal traffic.

• Kimi K3 — $3 / $15. An open-weights flagship at roughly half Opus's rates.

• DeepSeek V4 Pro — $0.44 / $0.88 off-peak (peak windows double it). The cheap open-weights flagship, a different price universe entirely.

The honest framing: Opus 5 is not a cheap model — it is a flagship priced at half the top tier (Claude Fable 5) and exactly what the previous generation cost. It wins on price-per-long-context and price-per-hard-reasoning, not on the absolute rate. If your traffic is short prompts and simple tasks, a $3 Sonnet or a $0.44 DeepSeek model is the cheaper call.

When Opus 5 is the wrong choice

Four cases where this price is not the one you want:

Short, simple prompts. Below roughly 20K tokens with no deep reasoning, the Opus premium is wasted. Claude Sonnet 5 at $3 / $15 (or $2 / $10 through the intro window) handles the volume tier for 40% less, and a Gemini Flash or DeepSeek-class model costs far less again.

Reasoning-heavy volume. Because thinking bills as output, a hard prompt at high effort can cost well above the nominal output line. If your workload is many difficult calls, measure actual output tokens before committing — the effective rate can surprise you.

Price-per-token as the only goal. If a frontier benchmark score or the 1M context isn't the hard requirement, cheaper flagships (GPT-5.6 Terra, DeepSeek V4 Pro) or Flash-tier models (Gemini 3.6 Flash) deliver far lower per-token rates for the easy majority of traffic.

Strict latency budgets. Opus 5 is a large thinking model — p50 time-to-first-token around 7 seconds in OrcaRouter's 7-day measurement. For real-time chat where sub-second TTFT matters, it is the wrong model regardless of price.

How to call it on OrcaRouter

Claude Opus 5 is hosted on OrcaRouter as model ID anthropic/claude-opus-5, served through the OpenAI-compatible endpoint at api.orcarouter.ai/v1 at the provider rate with zero markup. If you already have an OpenAI client, migration is a base-URL and model-ID change — nothing else. The same key and endpoint carry 200+ models, so Opus 5 sits next to Claude Sonnet 5, Claude Haiku 4.5, the GPT-5.6 tiers, the Gemini family and the open-weight models you'd want to compare it against, and you can route by difficulty instead of re-integrating each one.

Two OrcaRouter specifics worth naming on a pricing page. BYOK: bring your own Anthropic key and Anthropic bills you directly at its own rates — OrcaRouter adds $0 per token and keeps your provider rate limits and credits intact. Guardrails: the PII Shield and content policy run before billing, so a blocked request returns a clean 400 and is never charged, and the Agent Firewall grades tool and MCP calls before they execute.

The honesty note that matters for a pricing page: we route the model, we don't set its price. The $5 / $25 figure is Anthropic's rate passed through, and when Anthropic changes it the listing follows — which is exactly why this page is date-stamped. If you call Anthropic directly you get the same rate; OrcaRouter exists for the single-key, zero-markup route if you want one endpoint across the whole model line.

Sources and date

All prices above were verified on 2026-08-18. The Claude figures are read from Anthropic's published API pricing and cross-checked against the OrcaRouter model page for Claude Opus 5 (checked the same day). The GPT-5.6 post-cut prices are OpenAI's published rates; the Gemini, Kimi and DeepSeek figures are the current OrcaRouter directory listings, with DeepSeek V4 Pro shown at its off-peak rate (peak windows double it). Anthropic prices have been stable since Opus 5's launch on July 24, 2026, but they move — treat any earlier "Opus 5 pricing" page you find with the date checked. The numbers here are current as of today.

Every price above is the live rate on the Claude Opus 5 model page — verified 2026-08-18 at zero markup, ready for your own key.