Choosing a Claude Model
Anthropic offers a family of models at different capability/cost/speed points. Choosing well is mostly about matching the model to the job — and not overpaying for capability you don't need.
- Read the Haiku → Sonnet → Opus ladder as a capability/cost/speed tradeoff
- Start from the right default instead of guessing, then move up or down deliberately
- Mix tiers in one system — the biggest cost lever most people never pull
- Look up the exact model ID the right way, so upgrades stay a one-line change
The current models
Last verified: 2026-08-07 · Official models & pricing
| Model | API model ID | Tier | Best for |
|---|---|---|---|
| Claude Opus 5 | claude-opus-5 | Opus — most capable | Complex agentic coding and enterprise work: 1M context, 128k output, with materially better numbers than Opus 4.8. Default on Claude Max, strongest available on Claude Pro. (Supersedes Claude Opus 4.8, now a legacy model.) |
| Claude Sonnet 5 | claude-sonnet-5 | Sonnet — balanced | The default workhorse and Claude Code's default model: the best combination of speed and intelligence, with a 1M-token context window. Start here for most work. (Supersedes Claude Sonnet 4.6, now a legacy model.) |
| Claude Haiku 4.5 | claude-haiku-4-5-20251001 | Haiku — fastest & cheapest | High-volume, latency-sensitive tasks: classification, extraction, routing, simple chat, and cheap sub-agents. |
| Claude Fable 5 | claude-fable-5 | Mythos-class — flagship (above Opus) | Anthropic's most capable widely released model. Multiday autonomous runs, first-shot on ambiguous specs, dense vision, sustained delegation to parallel subagents. GA June 9 2026. Refuses in-band (HTTP 200, stop_reason: refusal) — configure server-side or client-side fallback to another current Claude model (Opus 5 is the recommended default). Adaptive thinking only; raw thinking never returned. 30-day retention (NOT ZDR-eligible). |
| Claude Mythos 5 | claude-mythos-5 | Mythos-class — restricted (Project Glasswing) | Same capabilities as Fable 5 without the safety classifiers. Limited availability to approved Project Glasswing partners only; not self-serve. Customers without access should use Fable 5. |
Model IDs are exact; tiers and 'best for' are guidance. Pricing, context-window sizes and rate limits change frequently and are intentionally NOT listed here — always read them on the official page above. This file is the single source of truth: never hard-code model facts in prose, link to this table instead.
Try it: which model fits?
Answer three questions and get a starting recommendation:
The balanced default — strong reasoning and coding at a fraction of Opus cost. Start here, and only move up if you hit a real quality ceiling.
Look up the exact model ID on the models table. Rule of thumb only — run a quick eval on your own inputs to be sure.
The mental model: a capability ladder
- Start with Sonnet. It's the default workhorse — strong reasoning and coding at a sensible cost. Most tasks should begin here.
- Move up to Opus only when Sonnet struggles and quality matters more than cost (hard reasoning, tricky agents, gnarly code).
- Drop to Haiku for high-volume, latency-sensitive, or simple work (classification, extraction, routing, cheap sub-agents).
How to actually choose
- It is the balanced workhorse. Starting anywhere else means you are optimizing before you have evidence about your actual task.
- Do not upgrade the whole workload. Find the cases Sonnet fails and route only those to Opus — you buy the quality without paying for it everywhere.
- Classification, extraction, routing, and cheap subagents rarely need a bigger model. Test it rather than assuming.
- Use Haiku for cheap pre/post-processing and Sonnet/Opus for the hard core. This model tiering is one of the biggest cost levers — see Cost & Latency.
Model tiering is worth its own read: Cost & Latency.
:::tip Don't pick from benchmarks alone Public benchmarks are a starting hint, not a verdict for your task. Run a tiny eval on a handful of your real inputs across two models — it takes minutes and beats guessing. :::
Looking up the exact model ID
Always pass the current API model ID (e.g. in your messages.create call). Get it from the models table above or the official models page — and prefer reading it from config over hard-coding it in many places, so model upgrades are a one-line change.
Check yourself
0/4- Haiku → Sonnet → Opus is a capability/cost/speed ladder — pick a rung, don't guess a model.
- Default to Sonnet and ship; move up or down only with evidence from your own task.
- Upgrade the hard subset, not the whole workload — routing beats blanket upgrades.
- Mixing tiers in one system is one of the biggest cost levers available to you.
- Benchmarks are a hint; a tiny eval on your real inputs is the verdict.
- Read the model ID from config and look it up in the live models table — never hard-code model facts.