Gemini 4 Carbon Is Cheaper Than Fable. Its Launch Price Has an End Date.

Demis Hassabis, wearing glasses and a dark jacket, rests his chin on his hand in a portrait.
Google DeepMind cofounder Demis Hassabis in July 2025. Carbon’s broader November 2026 release brings another model into practical cost comparisons. Image credit: Christopher Michel, CC BY-SA 4.0. Resized, full frame retained.

Another frontier AI model has arrived, and the usual temptation is to ask which company is ahead this week. My first question about Gemini 4 Carbon is more practical: what will the same work cost after the introductory price ends? A useful trial in November should not quietly become a more expensive habit in December.

Google opened Carbon to paying API developers and enterprises on November 10, following an earlier restricted release. It is also rolling the model out to Google AI Ultra subscribers in products including the Gemini app and Antigravity. The availability announcement therefore matters as much as the performance claims: more people can now evaluate a model that was previously available to a narrower group. Google’s availability announcement and pricing footnote

That gives us a concrete comparison with Anthropic’s Claude Fable 5.2. I am reading the published terms and results here, rather than reporting a hands-on contest. The decision I care about is which model earns its place in a workflow once quality, cost and review time are counted together.

Read the pricing footnote before running a trial

Carbon’s introductory API price is $2 per million input tokens and $10 per million output tokens. Google’s footnote says that after the introductory period expires on November 30, 2026, the public-access model will cost $4 and $20 respectively. The ongoing listed rates are twice the introductory rates. Google’s availability announcement and pricing footnote

Anthropic lists standard Fable 5.2 API rates of $8 per million input tokens and $40 per million output tokens. At identical billable token counts, that makes Carbon’s introductory rate one quarter of Fable’s, and its announced subsequent rate one half. That comparison concerns ordinary token prices, not every feature, discount or way of using the products. Anthropic’s API prices

Published API token rates, checked November 11, 2026
Model and price period1M input tokens1M output tokens
Carbon: introductory period ending Nov. 30$2 ≈ €1.78$10 ≈ €8.92
Carbon: after that introductory period$4 ≈ €3.57$20 ≈ €17.83
Fable 5.2: standard rates$8 ≈ €7.13$40 ≈ €35.67

The approximate euro amounts use the November 10 ECB reference rate of $1.1214 per euro. They are currency conversions of published US-dollar API prices, not euro checkout quotes. Taxes, payment conversion charges, tools, caching and other billable features can change the final amount. ECB reference rate, November 10

This is exactly the kind of footnote I want recorded alongside a benchmark. Otherwise, a team could compare a promotional Carbon run with a normal Fable run, choose a winner and then keep quoting the trial’s savings after the pricing basis has changed.

A small job makes the difference easier to see

Take an illustrative job that bills 250,000 input tokens and 25,000 output tokens, with no additional features or discounts. Carbon’s introductory token bill would be $0.75, about €0.67. At its announced later rate, the same counts would cost $1.50, about €1.34. At Fable’s standard rate, they would cost $3.00, about €2.68.

Those numbers are arithmetic, not measured costs for a real task. They deliberately hold the token counts fixed so that the price difference is visible. In actual use, the models may take different routes, request different tools and produce different amounts of billable output.

Suppose a low-cost attempt fails a requirement and needs several retries. Its initial price is no longer the whole price of finishing the job. Conversely, suppose the cheaper model produces an acceptable answer on the first attempt. Paying a higher token rate for a more prestigious model would need some other justification.

I also count my own attention. A result that needs fifteen minutes of careful repair can be less attractive than a slightly more expensive result that I can verify quickly. That is especially relevant to research and blogging, where a polished sentence can still contain a missing condition, a wrong attribution or an unsupported conclusion.

The leaderboard needs a task attached to it

Google reports strong Carbon results on coding and workflow benchmarks, including CursorBench, Perplexity’s WANDR and Zapier’s AutomationBench. It also explains that Carbon can use more tokens than Argon for deeper exploration and verification. Both points belong in the comparison: strong results and inexpensive individual tokens do not establish a universally inexpensive completed task. Google’s availability announcement and pricing footnote

I find broad labels such as “best model” least helpful when they combine different jobs. Editing a website, analyzing a spreadsheet, tracing a long argument and using an external service impose different requirements. Even a clear lead on one evaluation cannot tell me which mistakes will be most expensive in my own work.

For a research example, I might ask two models to compare statements from several papers while preserving their hypotheses. I would check whether each identifies where the assumptions differ, whether its references support its claims and whether it admits when the supplied material does not settle the question. A fluent overview is not enough to pass.

For a blog example, I might ask for a source-backed product comparison with exact regional variants and dated prices. A model that quietly mixes a US specification sheet with a European store listing has missed the task even if the prose is pleasant. These are proposed evaluation cases, not results I have already obtained from Carbon or Fable.

Keep the comparison small enough to finish

My starting point would be a handful of representative tasks with acceptance criteria written before either model runs. Include one ordinary job, one longer job and one awkward case where the correct answer requires acknowledging missing information. A tiny trial cannot establish an industry ranking, but it can expose a poor fit.

Use the same source material and record the actual model version, settings and available tools. Save the outputs before revising the prompts. If one model gets an extra hint or a different tool, record that too. Otherwise, it becomes difficult to tell whether the model improved or the test became easier.

For each task, track the full billable usage across retries, the elapsed time and the amount of human repair. Mark the result as accepted only when it meets the original requirements. An abandoned attempt belongs in the cost record, because it still consumed resources even though it produced nothing useful.

Then examine the failures. If both models struggle with missing source material, changing models may not help much. If one consistently mishandles an important format or loses a required qualification, that could be a useful reason to route that category of work elsewhere. The point is to make a decision small enough to act on.

API prices and subscriptions answer different questions

The table above does not compare monthly subscriptions. An API lets a developer pay for metered usage under its billing rules; an app subscription packages access, tools and usage allowances differently. A model appearing in a subscriber’s menu does not make its API price the cost of every conversation. Our earlier Fable and Mythos guide explains the separate access routes on Anthropic’s side. Our guide to Fable and Mythos access

For someone already paying for an eligible plan, the sensible first step may be to test the newly available model within the access they already have. Before adding another subscription, check whether the desired model is included, what limits apply and whether the needed tools are available in that product.

For a small application, metered API use may be easier to compare, provided the developer records the complete bill. A price per million tokens looks precise, but the number of tokens used is a behavior of the application and the model together. Prompt length, repeated context and repeated attempts all matter.

I would also keep a spending limit for the experiment and an explicit stopping point. When new models arrive quickly, testing can become a project of its own. There is little benefit in spending the week arranging a perfect contest if the original task was to finish an article or repair a small tool.

Competition is useful without a weekly migration

Carbon’s broader release gives users another serious option and a price schedule they can examine. That is useful competition. It is also a reminder to separate a company’s launch timetable from my own need to change tools.

My existing workflow has value: saved prompts, familiar review habits and an understanding of where I need to double-check an answer. A new model earns a switch by improving enough of that work to justify the adjustment. It does not need to replace everything to be useful; a narrow role can be a sensible outcome.

For November, I would record both Carbon price periods, run a limited comparison and keep the accepted outputs. Then I would decide using the rate I expect to pay over the intended period. The interesting winner is the model that helps me finish reliable work at an acceptable total cost, even after the launch offer disappears.

Comments

Popular posts from this blog

Obsidian’s New Search: Less Friction for Notes in Two Languages

OPPO Find X9 Ultra vs Find X10 Pro Max: Is the Older Ultra Still Worth Buying?

OPPO Find X10 Technical Preview: Cameras, Design and Missing Specs