![Claude Opus 4.5 API Pricing: Platform Cost Comparison [2026]](/_next/image?url=https%3A%2F%2Fstatic.linkmodel.ai%2Fblog-cover%2Fcover-claude-opus-pricing.png&w=3840&q=75)
The Version Question First
If you searched "Claude Opus 4.5 pricing," here's the update: Opus 4.5 has been succeeded by Opus 4.6, 4.7, and the current flagship, Claude Opus 4.8 (released May 28, 2026). The good news for budgeting is that the entire Opus 4.x line has held the same headline rate — so whether your code pins 4.5 or the latest 4.8, the per-token math below applies.
All figures here come from Anthropic's official pricing page. Rates change, so verify there before finalizing a budget.
Current Claude API Rates
Per million tokens (MTok), input / output:
| Model | Input | Output | Notes |
|---|---|---|---|
| Claude Opus 4.8 (flagship) | $5.00 | $25.00 | 1M context, no long-context surcharge |
| Claude Sonnet 4.6 | $3.00 | $15.00 | Best price/quality balance |
| Claude Haiku 4.5 | $1.00 | $5.00 | Cheapest current model |
| Claude Fable 5 / Mythos 5 | $10.00 | $50.00 | Top tier; reserve for hardest work |
Opus 4.8's rate is unchanged from 4.5/4.6/4.7 — a level Anthropic dropped to from the Opus 4.1 era's $15/$75, roughly a 67% cut. Opus 4.8 adds a cheaper Fast mode ($10/$50) and effort controls, but note its default effort moved to "high," which can quietly raise costs unless you set it.
The Discounts That Actually Move Your Bill
Two levers stack and can cut effective spend dramatically:
- Batch API — 50% off input and output, for asynchronous jobs (within 24 hours). Opus drops to $2.50/$12.50.
- Prompt caching — up to ~90% off cached input reads. For Opus, a cache hit is about $0.50/MTok instead of $5.00. Chatbots and agents that resend the same system prompt or context benefit most.
Watch two multipliers that push the other way: US-only inference (inference_geo: "us") adds ~10%, and cloud-marketplace regional endpoints (Bedrock, Vertex) add ~10% over global. A "$5/$25" model can land nearer $5.50/$27.50 for a US-resident workload before caching.
Platform Cost Comparison
The same Claude tokens cost the same base rate wherever you buy them — the differences are operational and in the discounts each route exposes:
| Route | Pricing behavior | Best for |
|---|---|---|
| Anthropic direct | List price; full Batch + caching | Direct control, newest models first |
| Amazon Bedrock | ~10% regional premium; AWS billing | AWS-native stacks |
| Google Vertex AI | ~10% regional premium; GCP billing | GCP-native stacks |
| Aggregator gateways | One key across providers; discounts vary | Multi-model apps, unified billing |
For teams already running image and video generation through a gateway like LinkModel — which aggregates models behind one API at up to 30% below official rates — consolidating billing and observability across providers is the operational win. Check the model catalog and pricing page for current availability and rates, since aggregator model coverage changes over time.
Which Claude Model Should You Pay For?
- Haiku 4.5 ($1/$5) for high-volume classification, extraction, routing.
- Sonnet 4.6 ($3/$15) as the general-purpose default — the price/quality sweet spot.
- Opus 4.8 ($5/$25) for the hardest reasoning, agentic coding, long-horizon work.
- Fable 5 / Mythos 5 ($10/$50) only when fewer failed agent loops justify 2x the token rate.
Routing simple tasks to Haiku and reserving Opus for genuinely hard prompts creates a 5–25x cost spread — the single biggest lever most teams leave unused. More tactics in how to reduce AI API costs.
In Context
Against peers, Claude sits at the premium end on raw tokens — Sonnet 4.6 ($3/$15) runs slightly above GPT-5.4 ($2.50/$15) and Gemini 3.1 Pro (~$2/$12), while budget options like DeepSeek V4 Flash ($0.14/$0.28) undercut everyone. Anthropic's counterweights are 1M-context at flat pricing, strong coding, and aggressive caching economics. The full cross-vendor rate card is in the AI API pricing comparison; for the cheapest generation routes specifically, see the cheapest AI API guide.