Frontier model releases now arrive every few weeks, and each one arrives framed as an upgrade, usually on the strength of a lower list price or a better leaderboard score. At CData Labs, we run models as they're released on a set of tasks reading and writing enterprise data through CData Connect AI Gateway. We want to know how much it costs models to get the right answers or take the right actions when using an optimized Connect AI configuration.
On what we call R1 (ranking accounts by health), the two most recent releases share a habit their predecessors mostly lack: both double-check every critical account one-by-one before listing the bottom 50 accounts. For Gemini 4 Carbon, that raises the cost per correct answer to about 2.4 times Gemini 4 Argon's, while Claude Fable 5.2 stays about level with Fable 5.1, which already ran the same checks in about two of every three runs.
Google announced Gemini 4 Carbon on October 28, 2026, and opened it to developers and enterprises on November 10, while Anthropic shipped Claude Fable 5.2 on November 9, each just weeks after the model it follows. The mechanism differs by vendor: Carbon costs the same per token as Argon, so every extra check lands on the bill, while Fable 5.2's 20% lower per-token price roughly pays for its checks.
In this article, we walk through the reason (both new models check every record individually) and what your teams can do about it.
At a glance: Cost per correct answer is a model's median cost per run divided by its median accuracy. Anthropic released Claude Fable 5.2 on November 9, 2026, at $8 / $40 per million input and output tokens. Google made Gemini 4 Carbon available to developers on November 10, 2026, at an introductory $2 / $10 through November 30. On CData Labs' account-health task, both checked every critical account individually in nearly every run. Carbon's cost per correct answer was about 2.4 times Argon's. Fable 5.2's was about the same as Fable 5.1's, which ran the same checks in about two of every three runs.
Two releases, one shared behavior
The same behavior shows up in two unrelated model families from two different labs, released within two weeks of each other. One model doing this would read as a quirk, while two vendors arriving at it independently suggests a direction frontier models are heading. The task hands each model a list of 50 accounts that already includes each account's review eligibility. Gemini 4 Argon always answers from that list, and Fable 5.1 does so in about one run in three. Fable 5.2 and Gemini 4 Carbon almost always call a per-account lookup tool for each of the 26 critical accounts first, which takes about 30 tool calls instead of four. The checks take time as well as tokens: Carbon's median run took about 3.5 minutes, against about 40 seconds for Argon.
Those extra calls did not improve the answer. Argon, Carbon, and Fable 5.2 returned the exact answer in every run. Fable 5.1 missed in five of its 40 runs, whether or not it checked.
How the benchmark measures cost per correct answer
R1, the task behind these numbers, asks a model to produce an account health report for high-priority accounts using a curated model context protocol (MCP) toolkit over customer relationship management (CRM), warehouse, and IT service management (ITSM) data. Cost per correct answer builds on the idea of taskonomics, pricing AI by the work it gets right rather than the tokens it consumes. Costs come from recorded token usage at each provider's list price, and the open-source harness and reproduction guide let you rerun every figure against your own data.
What the per-record checks cost on each model
R1 (account health report), curated MCP toolkit, pooled runs from November 11 and 12, 2026: 40 for Fable 5.1, 20 for each other model. The CData Labs table uses 10 runs per model.
Model | List price, input / output per million tokens | Runs that checked every critical account | Runs with the exact answer | Median tool calls per run | Cost per correct answer |
Fable 5.1 | $10 / $50 | 27 of 40 | 35 of 40 | 28 (about 3 or 29) | $0.61 |
Fable 5.2 | $8 / $40 | 19 of 20 | 20 of 20 | 29 | $0.51 |
Gemini 4 Argon | $2 / $10 | 0 of 20 | 20 of 20 | 4 | $0.030 |
Gemini 4 Carbon | $2 / $10 | 19 of 20 | 20 of 20 | 31 | $0.073 |
The Fable comparison shows how the checks are priced. Fable 5.1's runs fall into two groups, about $0.31 when it answers from the list and about $0.63 when it checks every account, because every call adds its result to the context the model reads on the next turn, and that growing context is billed again with each step. Fable 5.2 checks every time at about $0.51 per run, so its 20% per-token price cut roughly pays for the extra calls. The two land at about the same cost per correct answer, and Fable 5.2 gets there with predictable behavior and an exact answer in every run.
The Gemini comparison isolates the effect. Argon and Carbon share identical introductory list prices, so Carbon's cost per correct answer, about 2.4 times Argon's, comes entirely from the extra lookups. Google has also said Carbon's price rises to $4 per million input tokens and $20 per million output tokens after November 30, which would widen the gap further if Argon's pricing holds.
That said, results on the CData Labs model table's other two tasks are more mixed, which fits the pattern, since the effect is strongest where the bulk result already holds the answer, as it does in a good share of enterprise reporting.
Why newer models check every record
Per-record verification is a reasonable default for a model trained to be careful. Before reporting data it already has, the model double-checks each item against its source, the same instinct a diligent analyst brings to an unfamiliar dataset. Google's own launch post acknowledges the trade-off, describing Carbon as using more tokens than Argon to "explore thoroughly, reason deeply, and run extended verification loops." For enterprise agents, the value of that instinct depends on the data. When the bulk result is already authoritative, the checks add latency and tokens without improving accuracy, and because each lookup is a live query, the extra calls also show up as load on your CRM, warehouse, and ITSM systems. None of this shows up on a leaderboard, so it is worth tracking tool calls per task alongside accuracy whenever your teams evaluate an upgrade.
What this means for teams choosing a model
Test each new release on your own tasks and compare cost per correct answer, tool calls, and accuracy before upgrading, since list price and leaderboard rank describe a model in isolation while your tasks describe it in production. It also helps to keep one definition of each business rule and expose it through every tool, so careful models that verify records get the same answer the bulk view gives.
That consistency is why business context matters more than model choice for enterprise agents. Models will keep changing every few weeks, and your business definitions should not have to change with them.
That's where CData Connect AI fits. At CData, we built Connect AI Gateway to give agents one governed MCP endpoint with real-time access to hundreds of enterprise sources, through curated toolkits and derived views. Because every tool draws on the same governed definitions over live data, the per-record checks a careful model runs come back consistent with the bulk view, and your agents get one consistent answer whichever model is asking. The same three tasks anchored our earlier 178x cost spread study, and our gateway benchmark reached correct answers in a tenth of the tool calls when the gateway ran the computation. The CData Labs model table covers all 38 models across R1, R2, and A1, and the open harness lets you reproduce every result.
Benchmark your own AI agents on live enterprise data with Connect AI
Run the CData Labs harness against your own sources and see what each model costs per correct answer before you upgrade. Connect AI gives every agent governed, real-time access to your enterprise data through curated MCP toolkits. Start a free trial and compare models on the tasks that matter to you.
Frequently asked questions
Is Claude Fable 5.2 better than Fable 5.1 for AI agents?
On this task, yes, if you value predictable cost. On CData Labs' account-health task, Fable 5.2 returned the exact answer in all 20 runs at about $0.51 per correct answer. Fable 5.1 returned it in 35 of 40 runs, and its cost per run swung between about $0.31 and $0.63 depending on whether it checked each account individually. Averaged over all runs, the two cost about the same.
How does Gemini 4 Carbon compare with Gemini 4 Argon on cost?
Carbon costs about 2.4 times as much per correct answer on the account-health task. The two models share the same introductory list price per token, and Carbon matched Argon's exact answer in every run, but it checked each critical account individually in 19 of 20 runs, and its runs took about five times as long.
Why do newer AI models make more tool calls?
Because they verify each record individually. In this test, Fable 5.2 and Gemini 4 Carbon called a per-record lookup tool in nearly every run, even when the bulk result already contained the answer. That caution can help when data is unreliable, but on authoritative data it adds cost and latency without changing the answer.
Will Gemini 4 Carbon's price change?
Yes. Google's introductory price of $2 per million input tokens and $10 per million output tokens ends November 30, 2026. After that, Carbon costs $4 and $20, which would widen its cost gap with Argon on per-record workloads.
What is cost per correct answer?
It is a model's median cost per run divided by its median accuracy. It shows what you pay for a correct result, which list prices and leaderboard scores do not.
