Skip to content
CW101
Comparison

Fable vs Opus vs Sonnet vs Haiku

Four Claude model tiers, ten times apart in price. Which one you want is usually decided by task shape rather than by benchmark scores.

Last verified August 21, 2026

The short answer

Start with Opus 5. Move down to Sonnet 5 if it holds up on your evaluation — it is 60% cheaper and usually does. Use Haiku 4.5 for high-volume, latency-sensitive work that does not need a 1M-token window. Reach for Fable 5 only when a long-running agent is failing on capability, not on prompting.

Can you use both? Yes, and routing by task is the main cost lever most teams never pull. Use Haiku for classification and extraction, Sonnet for the bulk of production work, Opus for the hard steps.

Side by side

Option A

Claude Fable 5

$10 in / $50 out. 1M context, 128K output, adaptive thinking always on, slowest. Framed around long-running agents.

Pick it when

A long-running agent measurably fails on Opus 5 and the failure is capability, not context or prompting.

Option B

Claude Opus 5

$5 in / $25 out. 1M context, 128K output, May 2026 knowledge cutoff — the most recent of any current model.

Pick it when

Complex agentic coding and enterprise work, and any time you are not sure. This is the documented default.

Option C

Claude Sonnet 5

$2 in / $10 out. 1M context, 128K output, fast, positioned close to Opus 4.8.

Pick it when

Most production workloads. Try it before assuming you need Opus — the price difference compounds fast at volume.

Option D

Claude Haiku 4.5

$1 in / $5 out. 200K context, 64K output, fastest. Uses explicit extended thinking, not adaptive.

Pick it when

High volume, tight latency, bounded tasks — classification, extraction, routing, simple edits.

The table

Fable 5 Opus 5 Sonnet 5 Haiku 4.5
Input / MTok $10 $5 $2 $1
Output / MTok $50 $25 $10 $5
Context 1M 1M 1M 200K
Max output 128K 128K 128K 64K
Adaptive thinking Always on Yes Yes No
Thinking parameter No No No Yes
Latency Slowest Moderate Fast Fastest
Knowledge cutoff Jan 2026 May 2026 Jan 2026 Feb 2025

The three things that decide it in practice

1. Output is five times input on every tier. Which means a verbose model on a cheap tier can cost more than a concise one on an expensive tier. Measure your actual input:output ratio before optimising by rate card.

2. Thinking tokens are output tokens. On adaptive-thinking models the effort parameter is therefore a direct cost control, and it defaults to high on Opus 5 and Sonnet 5 via the API and Claude Code. Leaving it at default is a decision, even if you did not make it deliberately.

3. Context is not free even when it fits. A 1M-token window does not mean filling it is wise. See context window management.

The tokenizer boundary

Claude Opus 4.7 and later — every Claude 5 model — use a newer tokenizer producing roughly 30% more tokens for the same text. Haiku 4.5 and Sonnet 4.6 use the old one.

Comparing $1/MTok on Haiku against $2/MTok on Sonnet 5 therefore understates Haiku’s advantage: the same document is about 30% more tokens on Sonnet 5. The effective ratio is closer to 1:2.6 than 1:2.

How to actually choose

  1. Build an evaluation from tasks you care about, not benchmarks.
  2. Run Sonnet 5 first. It is the value case.
  3. Where it fails, check whether the failure is capability or context. Context problems are not solved by a bigger model.
  4. Route the failures to Opus 5. Route the easy volume to Haiku 4.5.
  5. Only consider Fable 5 for long-horizon agent work that is still failing on Opus 5.

Most teams that “need Opus everywhere” have never run step 2.

Common misconceptions

  • "Haiku is the cheap one, so it is five times cheaper than Opus." On input yes, but Haiku uses the older tokenizer and produces about 30% fewer tokens for the same text — so the real gap is wider than the rate card suggests.
  • "All current models have a 1M context window." Haiku 4.5 has 200K. It is the only current model that does not.
  • "Extended thinking works everywhere." It is the reverse of what people expect: Haiku 4.5 supports the thinking parameter and no Claude 5 model does.
  • "Fable is just a better Opus." It is twice the price, slower, and cannot be run without thinking. It is a different trade, not a strict upgrade.

Sources

  1. 01Models overview — Claude Platform DocsOfficialplatform.claude.com
  2. 02Pricing — Claude Platform DocsOfficialplatform.claude.com

Start typing to search every entity in the reference.