Anthropic's current Claude family covers three practical tiers. Fable 5 is the most capable widely released model, Opus 5 targets complex agentic coding and enterprise work, and Sonnet 5 balances speed and intelligence for production systems.
All three provide a 1 million token context window and 128,000 token synchronous output limit according to Anthropic's current model documentation. Availability spans the Claude API and major cloud platforms, with exact platform features and pricing worth checking before implementation.
The three-tier decision
Fable is appropriate when maximum available capability matters more than latency and unit cost: long-running agents, dense knowledge work, and the hardest ambiguous tasks. Opus is a more economical quality tier for complex coding and enterprise workflows. Sonnet should be the starting point for fast coding, assistants, and repeatable production tasks.
Sonnet 5 has introductory API pricing through August 31, 2026, after which the listed standard rate rises. Budgeting should use the post-introductory rate unless the workload is temporary.
Access through Anthropic and cloud platforms
Claude models are available through Anthropic's first-party API, Amazon Bedrock, Claude Platform on AWS, Google Cloud, and Microsoft Foundry, though a particular model or feature can arrive at different times on each platform. Use the first-party API for the clearest access to Anthropic-specific features; use a cloud marketplace when procurement, regional controls, or an existing enterprise agreement makes it preferable.
Model IDs from the newer generations are fixed snapshots rather than permanent moving aliases. Record the exact ID used in every evaluation and production trace so a behavior change is explainable.
Caching, tokenization, and cost
Prompt caching can materially reduce repeated context cost, but cache writes and reads have distinct prices. Newer Claude models also use a newer tokenizer that can produce more tokens for the same text than earlier generations, so a price comparison based only on per-million rates can be misleading.
Measure the complete bill on your actual documents. Track uncached input, cache writes, cache reads, output, and any regional or fast-mode multipliers alongside task success.
Start with Sonnet 5, compare Opus on the failures, and reserve Fable for cases where it demonstrably changes the outcome.
Primary sources