Which AI model should a business use now that Claude Fable 5.2 is out?

Anthropic released Claude Fable 5.2 on 9 November 2026. In an independent test it finished 41.3 percent of real freelance jobs to a standard a client would accept, and it costs less per token than the version before it. So which AI model should a business use? For long, open-ended work there is now a much stronger option. For routine office processes, a cheaper model is usually still the better pick.

What did Anthropic announce with Claude Fable 5.2?

Fable is the most capable model line Anthropic sells to everyone. It shipped alongside Mythos 5.2, a version with looser cybersecurity safeguards that only vetted organisations can use. For most companies, Fable is the one that matters. According to the announcement:

• It costs 8 US dollars per million input tokens and 40 dollars per million output tokens. heise and The Decoder put the previous price at 10 and 50 dollars.
• A fast mode generates text up to 2.5 times faster at double the price.
• Anthropic says responses arrive up to 30 percent faster than with Fable 5.1.
• You can use it through Anthropic directly and on Amazon Web Services, Google Cloud and Microsoft Azure.

A token is a chunk of text, usually part of a word. You pay for what the model reads and for what it writes.

What do 41 percent of real freelance jobs actually mean?

The figure comes from the Remote Labor Index, run by Scale Labs and the Center for AI Safety. It uses 240 finished projects that freelancers on Upwork were really paid for. The average project took about 29 hours of human work and paid around 630 dollars. Fable 5.2 delivered 41.3 percent of them at acceptable quality. Fable 5.1 managed 17.9 percent and Claude Opus 5.5 21.3 percent. According to The Decoder, the best score about a year ago was 2.5 percent.

There are two caveats. The testers say the biggest gains came in audio editing, architecture, interior design and 3D models, not in bookkeeping or order intake. A closer match for office work is AutomationBench, which tests business workflows. Anthropic's own table puts Fable 5.2 at 51.4 percent there, Opus 5.5 at 42.5 percent, and Google's Gemini 4 Carbon slightly ahead at 53.9 percent.

The takeaway is that a model can now carry a task through many steps on its own, where a year ago it would have stalled early. That still doesn't make the strongest model the right one for every process.

Which AI model fits which office task?

Anthropic itself recommends Opus 5.5 as the default for most work and Fable for the most demanding, open-ended projects. The same split works inside a company.

Routine processes with clear rules rarely need the top model. Think of order entry, meaning typing up purchase orders that arrive by email, checking invoices against orders, or answering "where is my delivery" questions. The agent reads similar documents over and over. An AI agent is a program that reads emails and documents, looks things up in your systems and prepares a draft that a person approves, or that goes straight through when the case is clear. In processes like these, reliability depends mostly on your rules, mapping tables and test cases.

Tasks without a fixed path are different. Take a made-up example: a wholesaler wants to pull the line items out of a 60-page tender, match them against its own range and draft questions for the customer. Every tender is laid out differently, appendices contradict the main document, and quantities sit in tables one time and in running text the next. This is where a model that can follow a long chain of steps without losing the thread earns its price. Fable 5.2 is a serious candidate for that kind of work.

How much does an agent on Fable 5.2 cost by comparison?

Here is a worked example with assumed numbers. For each case, an agent reads an order email with its attachment and master data, 5,000 input tokens in total, and writes a draft of 1,000 output tokens. At Anthropic's list prices, 1,000 cases come to:

• Fable 5.2 (8 and 40 dollars per million): about 8 US cents per case, roughly 80 dollars in total
• Opus 5.5 (4 and 20 dollars): about 4 cents, roughly 40 dollars
• Sonnet 5.5 (2 and 10 dollars): about 2 cents, roughly 20 dollars

The per-token price also hides something. A strong model thinking at a high effort level uses more tokens. The Decoder reports that Artificial Analysis measured 26.50 dollars per test task for Fable 5.2 at its highest setting, against 10 dollars for Opus 5.5. A lower price per token doesn't make Fable 5.2 the cheaper choice by default.

Compared with the staff time those 1,000 cases take by hand, all three amounts are small. For office processes, the model you pick decides reliability far more than it decides the bill. Our article on AI agent running costs shows how these amounts add up in practice, and the piece on LLM costs at scale covers what happens when volumes grow.

Should you switch to the new model now?

Only if a test on your own cases justifies it. Benchmarks measure freelance projects and coding tasks, not your delivery notes. The clean approach is a comparison: run cases you have already processed through the new model and put the results next to the old ones. Does it read your documents at least as reliably, and at what cost? Switch only when both answers are good.

That only works if the agent is built so the model can be swapped without rewriting rules and integrations. Our article on avoiding vendor lock-in explains how to stay independent of a single provider.

At Sophera Consulting this comparison is part of running an automation. Every automation comes with a maintenance agent that tests changes against past cases and handles the move to new AI models. We build at a fixed price with no subscription, and token costs are billed directly to you. In the free automation check we work out which process makes a good starting point and which model suits it.

Our recommendation

Fable 5.2 is a real step forward for long, open-ended work, and the price per token has dropped. For most office processes in wholesale, logistics and hospitals, the advice stays the same: start with one process, use the cheapest model that reliably passes your test cases, and keep the model replaceable. Where an agent still gets stuck on long, ambiguous cases, Fable 5.2 is worth another try. Fable is only available as a cloud service. If patient data must not leave the building, a model can run locally on your own hardware.