Anthropic announced new models Claude Fable 5 and Claude Mythos 5. Fable 5 is designed for general use, Mythos 5 uses the same underlying model with relaxed constraints and remains available only to selected partners. You can access Fable 5 via the Claude API as claude-fable-5 at $10 per million input tokens and $50 per million output tokens.
That's the part everyone reads. But honestly? The price per token is the least interesting piece of the whole story. Let's look at why.
What Anthropic announced
Fable 5 is said to be significantly stronger at coding, analytical work, image understanding, long context and scientific tasks. Anthropic mentions large-scale code migrations, document and spreadsheet analysis, precise reading of data from charts and long agentic tasks that require ongoing planning.
The key word is "agentic". Fable 5 is not primarily a model you ask one question and get one answer. It is built for tasks where the model works across multiple steps, maintains its own plan and progresses toward the goal relatively autonomously. And that is exactly where the way we think about cost starts to break.
A new trend: models are getting more verbose
We are seeing a shift that can't be read from a pricing table. The latest models are becoming more verbose. To solve the same task, they take more steps, "think through" the approach more and analyse things more deeply.
This is not a flaw — on the contrary. It is precisely what enables them to handle more complex tasks. But it has a consequence: the same problem that a model used to solve in a single pass may now go through several rounds of reasoning, intermediate steps and self-checking.
Cost per token vs. cost per outcome
For years we've been fixated on one metric: how much does a million tokens cost. It's comparable, looks objective and fits neatly in a table. But it measures input, not value.
The real question is not "what does a token cost" but "what does it cost to get a task to completion". That is: how many tokens, how many attempts and how much human oversight does one finished thing consume — one analysed contract, one completed migration, one report you don't have to manually correct.
Once you start calculating this way, the model ranking by price per token stops making sense. A cheaper model that wanders may cost more than a more expensive model that hits the result first time.
Building automations and AI on company data?
In Apexloop, data, documents, automations and permissions share one common model — so the output of an AI agent can be checked, audited and measured, not just run.
What this means for enterprise deployment
For teams building internal applications, automations or AI assistants on company data, this shift in thinking matters more than a few dollars in the price list.
The move toward long agentic tasks is appealing — a model that better maintains a plan and context is more practical for workflows like contract analysis, data migration, report preparation or database review. But that same shift is exactly what increases token consumption. Capability and cost rise hand in hand here.
So the practical advice is not "pick the cheapest model" but:
- Define what a "completed result" means for a given task and measure costs against it.
- Track first-attempt success rate, not just price per token.
Where a simpler task is sufficient, don't deploy the most verbose model. The goal is not maximum reasoning but the minimum reasoning needed for the result.
A detail that also affects the outcome
One more thing that feeds into "cost per result" calculations. Fable 5 is generally available, but Anthropic added conservative safety classifiers on top. If a query falls into sensitive areas — cybersecurity, biology or chemistry — the response may be taken over by Claude Opus 4.8.
For a company this means that for a portion of queries, model behaviour may change. Anthropic states that this affects fewer than five percent of sessions on average and that the safeguards sometimes catch an innocuous request. If you're building a product on this model, it's worth accounting for this variability — it too belongs in the question of how reliably you get the expected result.
What to take away
A stronger model is not automatically a complete solution. In real deployment, what still decides is the data model, permissions, audit, output testing and clearly designed boundaries of what the AI agent is allowed to do autonomously.
But even the thinking about cost deserves an upgrade. A new model is an opportunity to switch the mindset: stop asking what a token costs and start asking what a result costs. That is the metric that actually matters.