Claude models compared: Fable 5, Opus, Sonnet, Haiku

When I show people Claude in a session, the first stumbling block is usually the model picker. Four Claude models, all with poetry names, no obvious clue which one to use. And as of this week there's a new one at the top: Fable 5.

This is my plain-English guide to the Claude models compared. What each one is for, what they cost, and which one you should actually pick. No benchmarks you'll never read, no token-counting maths.

One framing before we start. I call these tools Autocomplete Machines, because that's what they are: very good prediction engines, not magic. The models below are the same machine at different sizes. Bigger means smarter, slower and more expensive. That's the whole trade-off.

The short answer

Use Sonnet for almost everything. It's the default for a reason.

Use Haiku when you want speed and the task is simple. Use Opus when the work genuinely needs deep thinking, like pulling apart a 200-page document. Use Fable 5 when even Opus isn't cutting it and the task is long, complex and worth the cost.

That's it. The rest of this article is the detail behind that advice.

The Claude models compared

Model	Think of it as	Best for	Context window	API cost ($ / M tokens, in / out)	Where you get it
Haiku 4.5	The sprinter	Quick answers, summaries, pulling info out of text	200k tokens	$1 / $5	Free plan and up
Sonnet 4.6	The daily driver	Writing, analysis, spreadsheets, most everyday work	200k tokens (1M via the API)	$3 / $15	Free plan and up
Opus 4.8	The deep thinker	Research, long documents, complex multi-step problems	200k tokens (1M via the API)	$5 / $25	Pro plan and up
Fable 5	The new flagship	The hardest, longest tasks; senior-level analysis	200k tokens (1M via the API)	$10 / $50	Paid plans (included until 22 June 2026, then usage credits)

Don't worry about what a token is. The column to read is the ratio: Fable 5 costs ten times what Haiku does. If you're paying a monthly subscription rather than building on the API, the cost shows up differently: the bigger models burn through your usage allowance faster.

The context window is the model's working memory: everything it can hold in its head during one conversation. In the Claude apps you get the standard 200k window. The 1M version only applies if you're building on the API, so don't pick a model for the bigger window if you're just using the chat. In plain terms, 200k tokens is roughly 150,000 words, or a 500-page book, that the model can keep in view at once.

Haiku: the sprinter

Haiku 4.5 is the smallest and fastest model. It answers almost instantly.

Use it for the boring-but-frequent stuff. Summarise this email thread. Pull the invoice numbers out of this document. Categorise these 50 customer enquiries. If a task feels like admin rather than thinking, Haiku will do it well and won't dent your usage limits.

If you run a small business and you're automating anything high-volume, like tagging incoming enquiries, this is the model you want behind it. Paying Opus prices for that job is like booking a removal van to post a letter.

Honestly, I barely touch Haiku myself. The one time I reach for it in a session is to show people something important: its intelligence is jagged. Ask it whether to walk or drive to a car wash 50 metres away and it can earnestly talk itself into driving. Speed isn't the same as judgement. So use Haiku to pull, sort and summarise. Don't lean on it to reason.

Sonnet: the daily driver

Sonnet 4.6 is the one to default to. It handles writing, analysis, document work and multi-step problems well, and it's fast enough that you're not sitting around waiting.

In my sessions, most of what people actually want to do lands squarely in Sonnet territory. Drafting client emails. Reviewing a contract. Building a first version of a board pack.

In my own setup, Sonnet is the workhorse that actually runs things. Once a workflow is designed, Sonnet is what executes it day to day: my morning AI news briefing, my SEO analytics, the scheduled jobs that run without me watching. It's capable enough to carry out the plan reliably, and cheap and fast enough to do it on repeat.

If you've never thought about which model you're using, you've probably been using Sonnet. And honestly, that's fine. The mistake isn't using Sonnet too much. It's using the expensive models on tasks Sonnet handles perfectly well.

Opus: the deep thinker

Opus 4.8 is built for problems that need sustained reasoning. Long documents, research, analysis where being wrong is expensive.

The test I give people: did Sonnet struggle with this? If yes, move up to Opus. If you never tested Sonnet first, you're guessing.

A concrete example of where Opus earns its keep: feeding in a lengthy regulatory document and asking for the implications for your specific business, not just a summary. The cheaper models will give you a decent precis. Opus will notice the clause on page 140 that contradicts the one on page 12.

Opus is also what I build with. When I'm designing an automation, writing something substantial, or putting together an app, I reach for Opus. Setting up a system that then has to run unattended is exactly the kind of sustained reasoning it's good at. Then I hand the day-to-day running of that system to Sonnet. Design with Opus, execute with Sonnet.

Opus needs a paid plan (Pro or above), and it uses your allowance faster. Reserve it for work that deserves it.

Fable 5: the new flagship

Fable 5 launched on 9 June 2026, and it's a step above Opus. Anthropic calls it a "Mythos-class" model, a new tier above Opus, and says its lead over the other models grows as tasks get longer and more complex.

Two things stood out to me from the launch material. First, Stripe reported that Fable 5 did a code migration in a day that would have taken a team over two months. Second, on Hebbia's Finance Benchmark, which tests senior-level reasoning over documents, charts and tables, Fable 5 scored highest of any model. If your work involves dense documents and judgement calls, that second one is the result to care about.

I've only just started using Fable 5, but one thing already stands out: one-shotting. Hand it a large, complicated app to build and it can come back with a working first version in a single pass, where with other models I'd expect several rounds of fixing. That tracks with the Stripe result.

There's a catch worth knowing about. Fable 5 ships with safety guardrails: ask it about certain topics (mainly cybersecurity and biology) and your request gets quietly answered by Opus 4.8 instead, with a notice telling you so. Anthropic says this happens in under 5% of sessions. For normal business use you'll rarely see it.

And a pricing wrinkle: Fable 5 is included free on paid Claude plans until 22 June 2026. After that it moves to pay-as-you-go usage credits until Anthropic has the capacity to fold it back in. So if you're curious, this fortnight is the cheap time to try it.

At $10 in / $50 out per million tokens on the API, it's double Opus. Worth it for the genuinely hard stuff. Wasted on email.

How to pick without overthinking it

Start every task on Sonnet. If the answer comes back shallow or wrong, escalate to Opus. If Opus still isn't getting there, and the task is long and high-stakes, that's a Fable 5 job. If the task is trivial and repetitive, drop down to Haiku.

If you're building a system rather than just chatting, the pattern flips a little: design it with Opus, then let Sonnet run it on a schedule. That split — the expensive model designs, the cheaper model executes — is how I keep my own automations both good and affordable.

One more thing the official guidance says that most people miss: each new model release is a fresh training run, not an upgrade patch. A task that suited Opus last year might sit better with Sonnet now. When a new model lands, spend ten minutes re-running your usual tasks across the options. It's the cheapest performance review you'll ever do.

Garbage in, garbage out still applies to all four. A well-written prompt on Haiku beats a lazy one on Fable 5 more often than you'd think.

Where to go from here

If you want to see what these models can do with your actual work rather than someone else's examples, that's what my sessions are for. I've written up what happens in a 90-minute session if you want the detail, and there's a wider comparison of the main AI models if you're weighing Claude against ChatGPT and Gemini.

If you're not sure whether a session is right for you, message me. I'll tell you honestly.