On 9 November, Anthropic released Claude Fable 5.2. The next day, Google made Gemini 4 Carbon available to paying developers and enterprises. Both are each company's strongest model, and both are pitched at hard, long coding work. Somewhere on your team, someone has already asked whether you should switch.
The most useful line in either announcement is easy to miss. Both vendors say their new top model is not what most of your work should run on.
What was released
Claude Fable 5.2. Anthropic says it communicates more clearly than Fable 5.1 and produces output up to 30% faster. On FrontierCode 2.0, a benchmark Anthropic runs, it reports 49.1% at its highest effort setting, up from 32.4% for Fable 5.1. Standard API pricing is $8 per million input tokens and $40 per million output tokens, with an optional fast mode at twice the price. Medium effort is the default in Claude apps and Claude Code. The announcement recommends starting with Claude Opus 5.5 for most coding and moving up to Fable for the most ambitious work.
Gemini 4 Carbon. Google calls it its strongest model and aims it at complex multi-step coding, debugging and research. It says Carbon ranks first on CursorBench 4.0, at less than half the cost per task of the next best external model, without giving the score. Pricing is $2 and $10 per million input and output tokens until 30 November, then $4 and $20. Google also says Carbon deliberately uses more tokens than Gemini 4 Argon, which it positions as the default for most software engineering work.
Why the benchmarks won't settle it
Each launch reports the numbers it chose. Anthropic's table puts Fable 5.2 ahead of Gemini 4 Carbon on the coding benchmarks it reports, and on FrontierCode 2.0 it compares Fable at its highest effort with Carbon at medium. Google cites a ranking on Cursor's benchmark without a score. Neither is wrong, but neither tells you how the model does on your codebase, with your conventions and your tests.
Cost is just as hard to read from a price list. A model with a lower price per token can still cost more per task if it thinks for longer, and both vendors say effort settings change the token count a lot. Anthropic, for example, says Fable 5.2 at high effort roughly matches Opus 5.5 on FrontierCode 2.0 at about half the cost per task. That is a claim about one benchmark, not a promise for your work.
Long agent runs show how fast this adds up. Decrypt reports that one reviewer's multi-day Fable 5.2 builds of full game and spreadsheet clones cost an estimated $11,700 to $14,710 each in API usage, while a small physics demo came in at $91. Those were deliberately huge tasks, but the spread is the point.
How to decide for your team
- Keep a default and a step up. Do what both vendors recommend: one everyday model for most work, and the top model for the tasks that are actually hard, like a tricky migration, a cross-cutting refactor or a bug nobody can find. Write down which is which so people don't pick by habit.
- Test on your own tasks. Pick five to ten recent, real pieces of work with a known good result: a merged pull request, a fixed bug, a small feature. Run each candidate model on the same tasks with the same instructions and the same effort setting, and compare the results.
- Measure what you care about. Did the change pass your tests? How much did a reviewer have to fix? How long did it take, and what did it cost per task, not per token? A model that is right first time on a hard bug can be cheaper than a cheap one that needs three tries.
- Watch the introductory prices. Carbon's price doubles after 30 November. If you test it this month, work out your costs at the later rate before you make it a default.
- Set effort on purpose. Effort settings change both quality and cost. Agree on a default effort for everyday work and when to raise it, the same way you agree on the default model.
- Keep your setup portable. New top models now arrive every few weeks. Rules, project context and tests that any model can use make switching a configuration change, not a project.
What to do this week
- Don't switch the whole team because of a launch post.
- Let one or two people try the new models on genuinely hard tasks and write down what they find.
- Run the small test above before changing your default.
- Check your usage limits and budgets, since top models use them up faster.
The takeaway
Two strong new models in two days is good news, and both may be worth having for hard problems. But the vendors' own advice is the right starting point: use the everyday model for everyday work, and keep the expensive one for tasks that need it. The way to know which is which on your team is a small, honest test on your own code.
Sources: Anthropic, “Claude Fable 5.2 and Claude Mythos 5.2” (9 November 2026), Google, Gemini 4 Carbon announcement (10 November 2026), Decrypt (9 November 2026).