GPT-5.6 Sol vs Claude Fable 5: the pricing cliff switching test

On July 9, OpenAI shipped GPT-5.6 Sol and ChatGPT Work, its hours-long workplace agent that can read and edit local files straight from the desktop. Today, July 12, Anthropic's free promotional window for Claude Fable 5 closed, and its prepaid rate of $10 per million input tokens and $50 per million output tokens went live, the highest published price on any of its generally available models. We ran that collision through PredictAible: 100 simulated US professionals and founders who each pay for at least one AI assistant, all asked one question, will you switch, stay, or run both.

No exodus, but a real split

Nobody stampeded for the door. In this synthetic room (stated intentions, not measured market share) the split settled at roughly 48% staying, 31% moving, 21% keeping both. What made the number worth reading was the reason behind it. Almost nobody moved over quality. They moved over price. The two forces on the table today are the Fable 5 pricing cliff and ChatGPT Work's capped plan, and the cliff is what moved feet.

The cliff did the moving

The loudest argument was the $50 per million output rate landing on the exact day the free window closed. One synthetic switcher laid out the budget math without drama: at $10 input and $50 output, Fable 5 "would triple what I pay now." For freelancers and small teams on lumpy income, that beat any benchmark chart. The brief's sources put Sol at about one-third less than Fable 5 with a coding edge; treat both as input assumptions of the run, dated the week it ran. For coders and growth marketers the call read as arithmetic, not loyalty.

Predictability pulled its own weight. ChatGPT Work is priced up to $100 per month with credit-based usage limits, and users on unpredictable budgets said a known ceiling beats an open-ended token bill. The counter-doubt was just as sharp. Several pointed out that "ultra" mode and hours-long agent runs could burn through those credits, and a plan that quietly turns unpredictable is worse than one that is honestly expensive. A good share of them are waiting a full billing cycle before they commit.

Files, and a Sonnet 5 side door

Local-file access cut both ways. The desktop agent editing files for hours was the single strongest non-price reason to move, with one podcaster claiming forty minutes back per episode. The same power was the top fear: an agent that can change your repo unsupervised can also quietly break it. The common resolution was to adopt Work but sandbox it to non-sensitive folders.

Two quieter moves matter. Plenty of Anthropic loyalists aren't leaving at all; they are dropping to Sonnet 5, the newer agentic Sonnet, to sidestep the Fable 5 rate while keeping their prompt libraries and SOPs intact. Anthropic loses Fable 5 volume there, not the customer. And a symmetric freeze held on both sides: Sol's national-security review and Fable 5's foreign-national ban left compliance-minded users unwilling to put sensitive work on either model until an independent audit lands. As one said, "a lower price tag doesn't mean lower risk."

What this run is not

These are stated preferences of simulated agents, not real buyers. Our earlier piece on how well these agents mirror real people is the honest place to start. One caveat outweighs the rest here: this simulation ran on Claude models, so a finding about whether people leave Claude for OpenAI carries a plain conflict of interest, and you should read the split with extra skepticism. We plan a cross-model replication before we treat the split as solid. This product category is also only days old, so intentions expressed in week one may not survive the first invoice.

The market brief behind this run was built from sources dated within the past week, each fact date-stamped, so the agents reasoned from what was true this morning rather than last quarter.

Want to pressure-test your own pricing change or launch before you make it? Build a simulation at /builder. Then test the loudest objections on real customers.