Nadya Dudka 
Andrew Dyuzhov Google made Gemini 4 Carbon generally available on November 10, 2026, one day after Anthropic released Claude Fable 5.2. They are the strongest models from each lab, and they are priced far apart. This month Carbon costs a quarter of Fable 5.2 per token, and from December 1 it costs half.
Which should you use?
- You want answers fast. Use Fable 5.2 at medium. In our test it finished in 4.2 minutes, against 12.7 for Carbon at medium.
- You run long agent jobs and watch the bill. Try Carbon at high. In our test it cost $1.29 per task this month against $7.70 for Fable 5.2 at high.
- You need the highest score you can buy. Use Fable 5.2 at high or above. It has two effort levels above high, and at max it scores 65 on Artificial Analysis, five points above Carbon’s best.
- Your work is tax, legal research, or finance. Carbon ranks first on several of those Vals benchmarks, where Fable 5.2 ranks lower. Test on your own documents.
Both labs point to another model for everyday work. Anthropic recommends starting with Opus 5.5, at half Fable 5.2’s token price, and Google recommends Gemini 4 Argon, at the same token price as Carbon.
What we measured
We ran one coding task through Atomic Agent, our open-source agent, on OpenRouter. The provider was pinned to Anthropic for Fable 5.2 and to Google AI Studio for Carbon. The task is a warehouse stock library of about 1,400 lines with 13 planted bugs and one missing feature. The agent has to read a spec, fix the bugs, build the feature, and get the tests passing. We then score the result against 59 hidden tests the agent never sees.
Each setting ran three times on the same build of the agent, which is included in the 0.7.0 release. Carbon costs are at the November price.
| Model and setting | Hidden tests | Time | Steps | Cost |
|---|---|---|---|---|
| Carbon, medium (API default) | 59/59, 3 runs | 12.7 min | 56 | $0.95 |
| Carbon, high | 59/59, 3 runs | 20.4 min | 59 | $1.29 |
| Fable 5.2, medium | 59/59, 3 runs | 4.2 min | 6 | $1.87 |
| Fable 5.2, high (API default) | 59/59, 3 runs | 26.6 min | 19 | $7.70 |
The Carbon high cost is the average of two runs. The longest run recorded only 23 of its 69 steps, so we left its cost out, and the true average is probably higher.
If the token mix stays the same, Carbon’s December price makes these about $1.90 at medium and about $2.60 at high.
What the numbers show:
- At medium, Fable 5.2 was three times as fast. It cost about twice as much this month, and about the same from December.
- At high, Carbon cost far less. Fable 5.2 at high cost about six times as much this month and about three times as much from December. Carbon at high also finished sooner on average in our runs.
- The two models work very differently. Carbon took 56 steps at medium and Fable 5.2 took 6. Fable 5.2 thinks longer per step and spends more per step, while Carbon takes many short ones. Most of Carbon’s input came from the cache.
- The API defaults point in opposite directions. Fable 5.2 defaults to high, its most expensive setting in our test. Carbon defaults to medium, its cheapest.
Limits of this test. It is one task, three runs per setting, and every setting scored 59 out of 59, so this test says nothing about which model is smarter. It shows what each setting spent to reach the same result. Runs varied: Carbon at high took between 14.6 and 29.6 minutes. The Fable runs were on November 9 and the Carbon runs on November 10, and some runs on both days went in parallel, so treat time comparisons as rough. We did not test Fable 5.2 at xhigh or max, or Carbon at low.
Pricing
| Per 1M tokens | Carbon until Nov 30 | Carbon from Dec 1 | Fable 5.2 |
|---|---|---|---|
| Input | $2 | $4 | $8 |
| Output, including thinking | $10 | $20 | $40 |
| Cache read | $0.10 | $0.20 | $0.20 |
| Cache write | not charged | not charged | $10 for 5 min, $16 for 1 hour |
| Batch input / output | $1 / $5 | $2 / $10 | $4 / $20 |
| Faster or priority tier | $3.60 / $18 | $7.20 / $36 | $16 / $80 |
Prices from Google and Anthropic. The last row compares different products: Anthropic’s fast mode raises output speed by up to 2.5 times, and Google’s pricing page does not describe what its Priority tier changes.
Carbon’s caching is automatic, with no separate charge to write to the cache and no explicit cache control. Anthropic lets you choose what to cache and charges for writes. Carbon has no free tier on the API.
The December change matters most for planning. At the same token count, Carbon goes from 25% to 50% of Fable 5.2’s price. A budget you build in November on Carbon’s current price will be half of what you pay in December.
Independent results
Artificial Analysis prices Carbon at the December rate. These are its figures as listed on November 10:
| Model and setting | Index | Cost per task | Time per task |
|---|---|---|---|
| Fable 5.2, max | 65 | $26.54 | 44.4 min |
| Fable 5.2, xhigh | 62 | $14.83 | 26.1 min |
| Fable 5.2, high | 62 | $9.08 | 17.5 min |
| Carbon, high | 60 | $7.76 | 25.6 min |
| Carbon, medium | 58 | $6.59 | 23.6 min |
| Fable 5.2, medium | 56 | $3.80 | 8.1 min |
| Carbon, low | 54 | $7.19 | 22.5 min |
| Fable 5.2, low | 54 | $3.10 | 6.5 min |
At high, Carbon scores two points less than Fable 5.2 and costs 15% less, or about 57% less at the November price by our arithmetic. At medium, Carbon scores two points more and costs 73% more at the December price. Fable 5.2 at medium and low finished in about a third of Carbon’s time on this index.
At the December price the two sources roughly agree that Fable 5.2 at medium is not the more expensive one: on the index it costs 42% less than Carbon at medium, and in our task they were about even, $1.87 against about $1.90. Only at the November price was Carbon cheaper in our task.
Vals AI tested both at high effort:
| Benchmark | Carbon | Fable 5.2 |
|---|---|---|
| Vals Index | 70.45% (3rd) | 71.31% (1st) |
| Vibe Code Bench v1.1 | 89.06% (9th) | 93.77% (1st) |
| Terminal-Bench 4.0 | 63.13% (5th) | 65.15% (1st) |
| Finance Agent v2.1 | 67.16% (7th) | 69.47% (1st) |
| Vals Legal Index | 75.93% (3rd) | 69.17% (20th) |
| Tax Agent Bench | 51.78% (1st) | 32.09% (27th) |
| Cost per test | $10.51 | $17.94 |
Vals figures as of November 9, at the $4 and $20 price for Carbon. Vals first published Carbon on October 28.
The pattern is clear. Fable 5.2 leads on coding and agent benchmarks. Carbon leads on legal and tax work by wide margins and costs 41% less per test. The overall index is close to a tie, with both scores inside about one point of margin.
Benchmarks from the labs
The two labs publish different benchmark sets. On these three rows both labs report the same score for Opus 5.5, and Anthropic’s own table shows the same Carbon scores as Google’s:
| Benchmark | Carbon | Fable 5.2 | Opus 5.5 |
|---|---|---|---|
| FrontierSWE 2 | 61.2% | 69.9% | 62.3% |
| AutomationBench | 53.9% | 51.4% | 42.5% |
| OSWorld-2.1 | 81.5% | 89.3% | 87.2% |
Carbon figures from Google’s model evaluation, Fable 5.2 figures from Anthropic’s announcement. Google took FrontierSWE 2 from Proximal and AutomationBench from Zapier. Both labs report OSWorld-2.1 on the offline subset with partial scoring, and Google’s Carbon score is the best of three runs.
Fable 5.2 leads on software engineering and computer use. Carbon leads on business workflow automation, and Google claims first place there at less than half the cost per task of the next best external model.
Effort and thinking levels
| Carbon | Fable 5.2 | |
|---|---|---|
| Levels | low, medium, high | low, medium, high, xhigh, max |
| API default | medium | high |
| Default in the labs’ own apps | not stated | medium in Claude apps and Claude Code |
| Can thinking be turned off | no off setting offered | no |
This is the part most likely to surprise you. If your agent or code sends no effort value, Carbon runs at medium and Fable 5.2 runs at high. In our test that meant $0.95 for Carbon and $7.70 for Fable 5.2 on the same task, an eightfold gap, against about twofold when both run at medium. Set the level yourself on both.
On Carbon the levels change the bill much less than on Fable 5.2. On Artificial Analysis, Carbon at high costs 18% more than at medium, while Fable 5.2 at high costs 2.4 times as much as at medium.
Limits and other differences
| Carbon | Fable 5.2 | |
|---|---|---|
| Context window | 1M tokens | 1M tokens |
| Output per request | 65,536 by default, more with background requests | up to 128,000 |
| Knowledge cutoff | March 2026 | June 2026 |
| Sampling settings | temperature, top_p, top_k ignored | temperature, top_p, top_k must be omitted |
Both models can take several minutes on hard prompts. Google recommends a client timeout of at least 15 to 20 minutes for Carbon.
Both labs restrict high-risk security work. Fable 5.2 blocks it and returns a refusal, with an optional fallback to an older Claude model. Google offers fuller cybersecurity features only to verified defenders in its Fairwind Program. Fable 5.2 also requires 30-day data retention.
On consumer plans, Carbon is in the Gemini app, AI Mode, and Antigravity for Google AI Ultra subscribers. Fable 5.2 is included in Max and Ultra plans and premium Team seats, while Pro and standard Team seats had it only for a launch week ending November 16 and otherwise pay through usage credits. Details are in our Fable 5.2 review.
How to try both
To repeat our test on your own work, download Atomic Agent for Mac, Windows or Linux, or install the terminal version.
curl -fsSL https://atomicagent.io/install | shatomic-agentOn Windows, run irm https://atomicagent.io/install.ps1 | iex in PowerShell. The source is on GitHub.
Connect OpenRouter with your own key and choose google/gemini-4-carbon or anthropic/claude-fable-5.2 from the live model list; see our OpenRouter guide. To set the effort level, add one line to your openrouter entry under llm.providers in ~/.atomic-agent/config.json:
"extraBody": { "reasoning": { "effort": "medium" } }The setting applies to every model on that entry. Without it, each model runs at its own default, so you would compare Carbon at medium with Fable 5.2 at high. More in our models guide.
For each model on its own, see our Gemini 4 Carbon review and our Fable 5.2 review. Fable 5.2 against Opus 5.5 is in our earlier test. If you are comparing agents rather than models, see our list of open-source Claude Code alternatives.
The takeaway
Fable 5.2 is slightly stronger on the independent indexes and clearly stronger on coding benchmarks. Carbon is close, leads on legal and tax work, and costs much less per token. In our test both solved the task completely. Fable 5.2 at medium was the fastest setting, and Carbon at high was far cheaper than Fable 5.2 at high.
Set effort yourself: with each model on its API default our task cost eight times as much on Fable 5.2, against about twice as much when both ran at medium. If you are deciding this month, remember that Carbon’s price doubles on December 1, and plan your budget on the December price.
