When Anthropic’s Fable model returned to the market on July 1st after a brief suspension, businesses rushed to get their hands on what was billed as a safer, guardrailed version of the company’s flagship Mythos model.
What many soon discovered had nothing to do with safety at all – it was cost. Within weeks, some businesses were staring at bills far larger than expected, and getting an early, expensive lesson in how using the most powerful frontier model can drive up business expense.
The Token Burn
Ad tech platform Scope3 is one company that quickly realised that despite its power, using Fable was hiking up unwarranted costs. The company’s co-founder and COO, Anne Coghlan, shared in a viral LinkedIn post that her fellow co-founder and company CEO Brian O’Kelley unexpectedly racked up over £20,000 of token spend due to Fable-use.
The primary driver of this cost spike was Fable’s tendency to overstep its role, Coghlan shared. While the team originally intended to use the high-end model strictly for planning, Fable frequently took matters into its own hands. “What we realised pretty quickly was that you’d ask Fable to do the plan and then hand it over to one of the other models for the build and if it wasn’t working quite fast enough, Fable would just be like, ‘I’ll just do it myself,'” Coghlan explains.
Despite the eye-watering bill, Coghlan clarifies that the situation wasn’t a financial crisis, but rather a valuable lesson in model management. Reflecting on her LinkedIn post about the incident, she notes: “It was more of a joke… I was like, ‘Dude, make sure that you’re using the right model, please,’ versus a, ‘Oh my gosh, we’ve broken the bank this month.'”
Costly Agentic Projects
The cost conversation isn’t only about specific models, but also what models are being used for agentic or non-agentic tasks – with agentic projects running continuously in the background, using up more tokens and being more costly.
In a report from McKinsey & Co from July 2026, it was found that 93 percent of organisations are exceeding their AI budgets. When looking into agentic AI economics, the paper revealed that roughly 60 percent of an agentic task’s total cost isn’t the model producing its first answer – it’s from the reasoning AI agents need to do to find the best final outcome. This includes checking that answer, correcting it, and re-verifying the result before it’s fit to use.
Therefore, enterprises comparing models on price-per-token are often measuring the wrong thing entirely – the more consequential cost driver is how much refinement, retry, and correction a given workflow generates downstream.
An agent that costs three times as much per call but requires no review cycle at all can be cheaper, in total, than one running on a bargain model that needs more supervision.
The Right Model for the Task
There are a few things companies can do to gauge how much their AI models are costing them. Businesses must understand how much an entire task – including retries and reviews – costs, and calculate cost per task over cost per token.
Prompting correctly with enough context is important too. This is so the model doesn’t go into internal reasoning loops, which can be costly. In this sense, building proper systems from the outset to make sure outputs are accurate and satisfactional first time becomes crucial.
There’s a temptation to assume that any model should, or can, be used for any type of task. But using a powerful and expensive model like Fable to check emails would be an inappropriate use of it, and of course drive up costs unnecessarily.
Additionally, the model with the lowest price per token isn’t always the cheapest model to run. In research from data intelligence platform Databricks, it was found that using the cheapest model doesn’t always mean ending up with the lowest cost overall. The company tested model outputs against engineer-written pull requests across eight different models, spanning a mix of task complexity.
Anthropic’s Sonnet 5 model is roughly 1.7 times cheaper than its Opus 4.8 model on a per-token basis. On Databricks’ own tasks, however, Sonnet ended up more expensive than Opus, not less because it used substantially more tokens to get to the final output, and finished with lower quality output on top of the higher bill.
The lesson cuts against the instinct to default to whichever model is cheapest. Revealing that a pricier model that gets a task right in one pass, with no rework, can easily beat a ‘cheap’ model that needs three attempts and a human review cycle before the output is usable.
To avoid getting burned, Coghlan advocates for a more strategic deployment of Fable – specifically utilising it to coordinate other systems rather than execute tasks directly.
“Use it for reasoning, planning and the specing before you use it to run all of your agents because agents can run for a really long time,” she advises.
To control these long-running costs, Scope3 treats Fable as a manager rather than a builder.
“Fable is the orchestrator… and then you have a swarm of agents underneath that are really cheap, and not necessarily that smart…” Coghlan says, suggesting that the smarter model should review and quality-control the cheaper models’ work. By establishing these architectural boundaries, businesses can harness Fable’s reasoning power without experiencing a runaway token burn.
Workflow automation AI platform Pega has taken an alternative radical approach, eliminating per-token pricing altogether in its platform. At PegaWorld 2026, CEO Alan Trefler unveiled Pega Infinity 26, under which clients building AI-driven workflows are no longer charged by token usage. The system moves the heavy AI reasoning to the design phase, so that agents running in production are fast and cheap to operate, rather than reasoning from scratch at every step. Trefler has been blunt about the alternative, calling the industry’s reliance on thousands of autonomous, prompt-driven agents a “philosophy of madness.”
The Takeaway
The key takeaway isn’t to not use extremely powerful models altogether, but instead understand what to use them for and where they most appropriately fit. Used well, Fable is a powerful tool. Used carelessly, it could be an expensive one.
