Claude Sonnet 5 vs Opus 5 vs Opus 5.5 vs Fable 5.1: Which One Should You Use?

Anthropic released four major Claude models between June and September 2026. If you blinked, you missed one. I use Claude every day, in Claude Code for coding, in the Claude apps for writing and research, and through the API in projects I build, and I've used all four. This is my practical guide: what each model is for, what it actually costs, and which one I reach for.

The four models at a glance

Two things jump out. First, Opus 5 is already a legacy model, less than two months after launch. Second, the newest Opus is cheaper than the one it replaced: Opus 5.5 costs 20% less than Opus 5.

All four read text and images, use tools, and share the same 1M-token context window, which is roughly 555,000 words.

What each model is for

Claude Sonnet 5: fast and affordable

Anthropic describes Sonnet 5 as "the best combination of speed and intelligence." It's the fastest of the four and the cheapest by far: $2 / $10, down from $3 / $15 on Sonnet 4.6, which it replaces.

It's a good fit for high-volume work: features inside apps where every request costs you money, quick coding questions, summaries, and everyday writing. Its knowledge cutoff is the oldest of the four (January 2026), which matters if you ask it about recent libraries or events.

For developers: Sonnet 5 now rejects non-default temperature, top_p and top_k values with an error, and adaptive thinking is on by default. If your code sets temperature, remove it before switching.

Claude Opus 5: good, but already replaced

Opus 5 launched in July as the new top Opus model. It's still available (Anthropic commits to keeping it running until at least July 2027), but Anthropic now recommends moving to Opus 5.5. There's little reason to start a new project on Opus 5: its successor is cheaper and newer.

Claude Opus 5.5: the new default

Opus 5.5 is five days old at the time of writing, and Anthropic now recommends it as the starting point for most work: "for long-running agentic coding and knowledge work." What changed from Opus 5:

• Cheaper: $4 / $20 instead of $5 / $25. Cache reads dropped from $0.50 to $0.20 per million tokens, the same as Sonnet 5.
• Thinking is always on. You control how hard it thinks with the effort setting, and it can't be switched off.
• Default effort dropped to medium (Opus 5 defaulted to high). If a task feels less thorough than on Opus 5, set the effort explicitly.
• It thinks more per turn than Opus 5 at the same effort level, especially at the top settings. So the lower price per token doesn't automatically mean a lower price per task. Measure it on your own work.
• Better with images: Anthropic says it reads values off dense charts, diagrams and screenshots much more precisely.

Claude Fable 5.1: the most capable, and the most expensive

Fable 5.1 is Anthropic's most capable widely available model, "for demanding reasoning and long-horizon agentic work." It costs 2.5× Opus 5.5 and is the slowest of the four. Anthropic's own advice is to use it when Opus 5.5 at higher effort still isn't good enough, not as a default.

One nice change: it costs the same as Fable 5, but cache reads are now a quarter of the old price ($0.25 per million tokens).

What a real task costs

Price per million tokens is hard to picture, so here's a worked example: a typical coding-agent request that reads 200,000 tokens (your code, instructions and conversation) and writes 20,000 tokens (its thinking plus the answer).

Calculated from Anthropic's published prices; cache-write costs are left out for simplicity.

Three takeaways:

1. Caching changes everything. Coding agents like Claude Code resend mostly the same context on every step, so most input is served from cache at a small fraction of the normal price.
2. Once input is cached, output dominates the bill. Output includes the model's thinking, so the effort level often matters more than the list price. Lowering effort on routine tasks is the easiest saving there is.
3. The big jump is to Fable 5.1. With caching, a request costs about 28 cents on Sonnet 5 and 52 cents on Opus 5.5, but $1.25 on Fable 5.1, about 2.4× Opus 5.5. The step from Sonnet to Opus is often worth paying for hard work; the step to Fable needs a real reason.

If you don't need an answer immediately, the Batch API halves all of these prices.

If you're upgrading code from Opus 5 to Opus 5.5

Changing the model ID is most of the work, but four things can break:

• Remove any setting that disables thinking. On Opus 5.5, thinking: {"type": "disabled"} returns an error. Lower the effort instead.
• Forced tool use returns an error. tool_choice of any or a specific tool isn't supported. Use auto, turn on strict tool use, and tell the model in the prompt when to use the tool.
• The old computer-use tool (computer_20251124) isn't accepted on the Claude API and Google Cloud. Move to the new computer-use toolset.
• Progress messages between tool calls moved. They now arrive as thinking blocks that are empty by default. If your app shows users "working on it…" updates, it will go quiet until you set the thinking display option.

Which one I use

Opus 5.5 is my default. It's the model Anthropic itself recommends starting with, it's cheaper than the Opus it replaced, and its knowledge is the most recent. With caching, it costs about twice as much as Sonnet 5, and for coding and server work I'd rather pay that than redo work.

Sonnet 5 for speed and volume: quick questions, simple edits, and AI features inside client apps where cost per request really matters.

Fable 5.1 for the hardest problems only: when Opus 5.5 at high effort still can't get there. At 2.5× the price and noticeably slower, it's a specialist tool, not a daily driver.

Opus 5: move off it. It still works, but Opus 5.5 is better value.

For developers here in Ghana, where every API bill is paid in dollars, my advice is simple: start with Opus 5.5 at the default effort, use caching, drop to Sonnet 5 for high-volume features, and save Fable 5.1 for the problems that genuinely need it.

Sources: Anthropic's official model documentation (platform.claude.com). This post was drafted with help from Claude; the choices and opinions are my own.