Two flagship language models sit at the top of the same leaderboards, and the gap at the top is small while the gaps underneath are much larger. These are text and code models, not text-to-image generators, so this piece compares reasoning and work-task benchmarks rather than picture quality. Each section explains what a test actually measures, shows the numbers once in a table, then keeps the prose focused on what the gap means in practice.
As of November 10, 2026, Fable 5.2 sits 1st on the Vals Index and 1st on the Artificial Analysis Intelligence Index, with Carbon 3rd on Vals and the next distinct model family on Artificial Analysis.
For background on Carbon before this head-to-head, see our Gemini 4 Carbon overview, and for how Anthropic’s prior generation compared with OpenAI, see our GPT-6 vs Claude Fable 5.1 comparison.
Vals Index: almost tied overall, split underneath
The headline is a narrow lead for Fable. The components show two different strengths rather than one model winning everywhere.
| Vals component, v2.2 (task-completion accuracy) | Fable 5.2 | Carbon | What the test asks the model to do |
|---|---|---|---|
| Vals Index overall | 71.31% | 70.45% | Weighted mix of finance, coding, legal, and tax work |
| Finance Agent v2.1 | 69.5% | 67.2% | Multi-step financial reasoning |
| Excel Modeling Benchmark v1.1 | 77.4% | 76.5% | Build and edit financial models in spreadsheets |
| Terminal-Bench 4.0 | 65.2% | 63.1% | Solve problems through the command line |
| Vibe Code Bench v1.1 | 93.8% | 89.1% | Build complete small apps end to end |
| Code Migration v1.1 | 85.8% | 80.1% | Move projects to another language, including older code |
| Legal Research Bench | 46.2% | 55.8% | Answer legal questions with cited cases and statutes |
| HLAB-Vals-1.1 | 5.0% | 17.5% | Produce longer legal work product |
| Tax Agent Bench | 32.1% | 51.8% | Answer research-grade U.S. corporate tax questions with citations |
Fable’s clearest separation comes in building and moving code. Carbon’s clearest separation comes in cited research tasks, especially tax and legal retrieval.
Artificial Analysis: benchmark results at matched effort
At the high effort setting for both models, Fable leads six of the nine Index evaluations, while Carbon leads three.
| Artificial Analysis Index evaluation, v5.0.1 (independent task-completion results) | Fable 5.2, high | Carbon, high | What the test asks the model to do |
|---|---|---|---|
| FrontierCode 2.0 | 49.1% | 40.6% | Ship code changes accepted by open-source maintainers |
| HLE-Diamond | 73.8% | 69.1% | Answer hard expert questions across subjects |
| CritPt v2, hard subset | 73.0% | 69.0% | Solve hard physics reasoning problems |
| Terminal-Bench Science 0.1 | 70.0% | 64.8% | Run scientific computing tasks in the terminal |
| AA-Omniscience v1.1 | 64 | 54 | Answer knowledge questions with fewer fabrications |
| AA-Briefcase v1.1 | 1677 | 1611 | Complete office documents and spreadsheets (preference Elo) |
| GDP.pdf | 21.8% | 28.4% | Do real document knowledge work end to end |
| AutomationBench-AA | 71.4% | 77.8% | Automate business workflows |
| Terminal-Bench 4.0 | 59.6% | 65.2% | Solve problems through the command line |
Carbon takes terminal coding and business automation on this harness, plus the document task. Fable takes the rest, with its widest margins in agentic coding and factual knowledge.
What the top scores cost in time and money
Artificial Analysis lists the same model more than once because effort changes the result. Higher effort usually raises the score and also raises cost and time.
| Artificial Analysis row (composite score plus measured cost and time) | Intelligence Index | Cost per task | Time per task |
|---|---|---|---|
| Fable 5.2, max | 65 | $26.54 | 46.4 minutes |
| Fable 5.2, xhigh | 62 | $14.83 | 26.4 minutes |
| Fable 5.2, high | 62 | $9.08 | 17.6 minutes |
| Carbon, high | 60 | $7.76 | 26.5 minutes |
| Carbon, medium | 58 | $6.59 | 23.5 minutes |
| Fable 5.2, medium | 56 | $3.80 | 8.3 minutes |
| Carbon, low | 54 | $7.19 | 21.3 minutes |
| Fable 5.2, low | 54 | $3.10 | 6.5 minutes |
The top scores are also the slowest and most expensive runs.
What the labs claim about their own models
Provider tables are useful but one-sided. Each lab chooses the tasks, settings, and rivals, so read them as claims to check against independent runs rather than neutral verdicts.
| Provider-claimed test (vendor-run accuracy, not independent) | Fable 5.2, per Anthropic | Carbon, per Google |
|---|---|---|
| Agentic coding, FrontierCode 2.0 | 49.1% at xhigh effort | 41.1% at medium effort in Anthropic table |
| Agentic coding, FrontierSWE v2 | 69.9% | 61.2% in Google table |
| Business workflows, AutomationBench | 51.4% | 53.9% in Google Zapier-variant table |
| Terminal science, Terminal-Bench Science 0.1 | 74.6% | 64.8% in Google table |
| Computer use, OSWorld 2.1 offline | 89.3% | 81.5% in Google table |
| Cybersecurity trio, SEC-bench Pro, ExploitGym, CWE-bench | Not the focus of Anthropic table | 88.2%, 54.8%, 71.0%, evaluated without guardrails |
Anthropic’s announcement is Introducing Claude Fable 5.2 and Claude Mythos 5.2. Google’s announcement is Gemini 4 Carbon: our next frontier model. The independent pages to watch for updates are the Vals Index leaderboard and the Artificial Analysis LLM leaderboard.
What the gap means in plain practice
A narrow overall gap will not be visible in casual use. The component pattern is more useful. Fable is the stronger pick for building software from scratch and for moving older code to a new setup. Carbon is the stronger pick for cited lookup work in legal and tax material within this particular Vals release.
Neither result says much about everyday chat, writing tone, or image creation. Both models accept image input, but the two leaderboards compared here score text, code, and research work rather than picture quality. For assistant-style tradeoffs such as speed, memory, and ecosystem fit, our broader Gemini vs ChatGPT guide remains the better starting point.
Price, speed, and access in brief
Token prices favor Carbon on paper, while longer reasoning traces change the per-task math. The tables above carry the task-cost figures; the takeaway is that Fable costs more per finished task in both harnesses despite the small headline gap.
Access remains uneven. Carbon was initially rolled out to cyber defenders and trusted partners rather than general availability, so many readers will encounter these scores before they can use the model directly.
Frequently asked questions
Is Fable 5.2 ranked first?
Yes. Fable 5.2 is 1st on the Vals Index and 1st on the Artificial Analysis Intelligence Index at the max effort setting.
Where does Gemini Carbon rank?
Carbon is 3rd on the Vals Index, behind Fable 5.2 and Sonnet 5.5. On Artificial Analysis it is the next distinct model family by top score, behind Fable 5.2.
Which model is better for coding?
Fable 5.2 leads the coding components shown here, especially app building and code migration. Carbon remains strong on terminal coding but does not lead these particular Vals coding splits.
Which model is better for legal and tax research?
Carbon leads the cited legal and tax components in this Vals release. That is a test of finding and citing sources, not legal advice.
Should I choose by the headline score?
No. Match the component to your task, then check effort, cost per completed task, and availability. Headline scores hide those tradeoffs.
Will these ranks stay fixed?
No. Vals and Artificial Analysis update versions, add runs, and revise methods. Re-check the live leaderboard pages linked above before making decisions.
The bottom line
Fable 5.2 holds the top spot in both independent rankings checked here, with Carbon close behind overall and ahead in cited research lanes. The honest summary is narrow at the top and split underneath: Fable for building and migrating code, Carbon for source-backed legal and tax lookup, with cost and access deciding many real choices. Verify the current leaderboard date before acting, because these tables shift.
