Capability moved behind a paywall the same week independent tests showed the top models still fail most real professional work.
A developer who left Claude Fable 5 running overnight could now wake up to a four-figure bill. On July 8 Anthropic pulled its most capable model off the flat-rate plans and put it on usage credits, $10 per million input tokens and $50 per million output, double the rate of Opus 4.8. The company called the change temporary and blamed a capacity crunch. Temporary or not, the subscription that felt unlimited now has a meter on it, and a heavy day of agent runs lands near $37.
Hold that next to a second fact from the same week. Snorkel's GDPval+ test handed roughly two thousand real professional tasks to the frontier models, and the best of them, xAI's Grok 4.5, passed 29% of the expert criteria. GPT-5.5 managed 22%. Opus 4.8 managed 21%. So the model you are now paying for by the token fails most of the work a professional is actually paid to deliver. That gap between the meter and the proof is the real story of the week, and it points two ways at once. Buyers just picked up a dependency they cannot yet verify, and anyone who sells software or advice has a proof layer worth building.
The escape hatch closed the same week. For years Meta sold openness as the alternative to all of this, free Llama weights you could download and host. Muse Spark 1.1 arrived as the opposite, a closed model reachable only through Meta's first paid developer API. The company that made free weights its wedge now charges by the token like everyone else. If your fallback plan was to drop off the meter and self-host Meta's best model, that plan lost an option this week.
Now look at where the vendors point you for reassurance, because that is where the second crack opened. Cursor pulled its own coding benchmark after finding that Grok 4.5, the model it co-launched, had an earlier snapshot of Cursor's codebase in its training data, the exact thing CursorBench is meant to test blind. On the one independent index Grok could not train against, it ranks fourth. xAI, for its part, published four benchmarks for Grok 4.5 and the model won two of them. A company grading its own exam is not evidence. This week that stopped being a cynic's line and became a documented event.
So the honest position for a buyer is uncomfortable. The price of frontier capability went up and turned variable, the free alternative narrowed, and the numbers vendors use to prove capability got less trustworthy, all in the same seven days. The change is not that AI got more expensive. It is that cost and proof moved in opposite directions at once.
Meanwhile the agents are being sold straight into the middle of the business, not the edges. Josh Kushner's Thrive Holdings is raising about $2 billion to buy accounting firms outright and automate the work from inside, and one of its companies has already acquired 48 of them. Mercor is buying Deeptune to build what it calls training gyms, simulated Excel sheets and Salesforce screens where an agent rehearses the office job and gets scored on it. And Anthropic's own data from the Claude Cowork mobile launch shows software development was only 8.7% of agent sessions, while business-process grind, reconciling spreadsheets and stitching updates into reports, was the largest slice at 33.4%. The work being automated first is the routine professional deliverable, which is precisely the work GDPval+ says the models still fail two-thirds of the time.
That is the tension a decision maker has to hold. The market is pricing agents as if they can reconcile the spreadsheet and file the tax return, while the cleanest independent test available says they clear less than a third of expert-graded work. Both can be true. The models are good enough to sell and not yet good enough to trust unattended, and the meter now runs whether the output is right or wrong.
Who this puts on the clock. Buyers who pinned a roadmap to a specific model lost the freedom to wait. Google's Gemini 3.5 Pro is still stuck in preview a month past its promised date, and Anthropic's metering cutover ran on a fixed calendar. Vendors gained leverage on both price and proof at once. And the reassurance most teams lean on, the launch-day benchmark chart, is the thing that got least reliable this week.
There is real agency left, and it sits in verification. The leaderboard is now the least useful number you have. The useful number is how the model does on your own deliverables at the price you now pay per run. That is measurable this week, not next quarter.
For buyers and operators, stop trusting the vendor's chart and build your own. Take a hundred of the real deliverables your team is actually paid to produce, the client memo or the month-end reconciliation, and grade the model against your own acceptance criteria rather than a puzzle score. Then compute cost per accepted output at the new metered rate, and test one cheaper fallback against it. Terra matches last year's GPT-5.5 at half the token price, and Grok 4.5 finishes a coding task in roughly a quarter of the tokens Opus 4.8 burns. Put a human approval step wherever money, compliance, or client trust changes hands, because that is exactly where the 71% failure lives.
For sellers, consultants, agencies, and software teams, the demand this week is verification, not adoption. Sell a workflow-specific eval and a cost-per-successful-task scorecard: a hundred of the client's own cases, graded blind, priced at current metered rates, with fallback routing to a cheaper model where it holds up and a named failure mode where it does not. Vendor benchmarks got less trustworthy this week and the meter made token efficiency real money, so the proof a client cannot get from a launch chart is the thing worth charging for.
The assumption worth dropping is that capability improves on a fixed subscription, so adopting more of it is a one-way ratchet. This week priced that out. Capability is a metered variable now, and the vendor's proof of it is not proof. Treat both as things you measure yourself, on your own work, before you route another dollar through them.
The week in one line: The frontier is metered and its benchmarks are compromised, so the only number that counts is how the model does on your own work at the price you now pay per run.
Sources this week: Claude Fable 5 goes pay-per-use, GDPval+ professional-work test, Meta Muse Spark closed API, Cursor pulls contaminated benchmark, Grok 4.5 token efficiency, Thrive Holdings buys accounting firms, Mercor buys Deeptune, Claude Cowork usage data, GPT-5.6 Terra pricing, Gemini 3.5 Pro still in preview