Signals
Can Gemini 4 Carbon Dethrone Fable 5.2? Not on Intelligence, Where Fable Leads 65 to 60. On Cost It Is Already Ahead. Every Board Where Both Have Been Tested, Google's Claims Checked, and Which One to Use
By the Top5Apps editorial team · Published November 10, 2026 · Updated November 10, 2026 · 12 min read
Short answer: not on raw capability, and it didn't need to. Claude Fable 5.2 leads Gemini 4 Carbon 65 to 60 on the Artificial Analysis Intelligence Index, and leads on most boards where both have been tested. But Carbon went on general sale today at a quarter of Fable's token price until November 30 and half after that. It matches Fable's best score on Cursor's coding benchmark for 28% of the cost, and it sits first on the Arena text leaderboard, where Fable 5.2 hasn't been ranked yet.
So there are two thrones. Fable 5.2 holds the one for the hardest problems. Carbon has the stronger claim on the one for work per dollar near the top. We read Google's launch post, model card, safety report and pricing pages, Anthropic's launch table, and every independent board that lists both models as of this afternoon. Nothing below is our own test; each number is linked to whoever ran it. (Disclosure, per The Receipts Standard: this site is produced with Claude-family tools, including Fable 5.2, one of the two models compared here.)
The short version
- On the index, no. Artificial Analysis has Fable 5.2 at 65 on max effort and 62 on high. Carbon is at 60 on its highest setting. Carbon held first place for twelve days, until November 9.
- Board by board it is nine to five for Fable. Fable wins where the job is to know or reason: hard code, expert questions, physics. Carbon wins where the job is to operate something: a terminal, an automation, a real codebase in Cursor.
- At top settings, Carbon is far cheaper. $7.76 a task against $26.54 on Artificial Analysis. On CursorBench, 60.4% for $9.22 against Fable's 60.0% for $32.79.
- At equal spend it is a split. Fable 5.2 at high effort costs $9.08 a task, scores 62 and runs faster than Carbon. On Cursor's board at about the same cost, Carbon leads by three and a half points.
- Google's claims check out, with a twist. On the one benchmark where its post admits second place, first place is Fable 5.2. Its model card still compares Carbon with Fable 5.1.
- Access: paid API today at $2 and $10 per million tokens, doubling December 1. In the Gemini app, Google AI Ultra only, from $99.99 a month. The public model is not the security version in Google's cyber charts.
- Not yet tested: Arena's text board has Carbon first and no Fable 5.2. LiveBench, ARC and Epoch have Fable 5.2 first and no Carbon. Nobody has published a hands-on comparison, and we have not run one.
- Our read: use Fable 5.2 for the hardest reasoning, Carbon for agents that run tools, and don't switch a $100 subscription over five points.
What Google shipped today
Google's post, signed by DeepMind's Koray Kavukcuoglu, says 'developers, enterprises, scientists, and consumers can now try our latest frontier model.' Two weeks after a release limited to security teams, this is the general one. The details, from Google's documentation:
- API: the model ID is gemini-4-carbon, in the Gemini API and AI Studio for paying developers. Google's pricing page is blunt: 'Gemini 4 Carbon is not available on the Free tier.'
- Price: $2 per million input tokens and $10 per million output 'through November 30, 2026,' then $4 and $20 'starting December 1, 2026.' Cached input is $0.10, then $0.20. Batch is half price. Argon's prices are identical, including the doubling.
- Specs: a 1,048,576-token context window and 65,536 tokens of output by default. Text, image, video, audio and PDF in; text out. Three thinking levels, low, medium and high, with medium the default. Knowledge cutoff March 2026.
- Gemini app and AI Mode: rolling out to Google AI Ultra subscribers only. The plans page lists Carbon under Ultra and nowhere else. The $19.99 AI Pro plan stays on Argon.
- Antigravity: Carbon for Ultra subscribers, plus a preview of /swarm, which runs a coordinator, 'parallel workers in isolated git worktrees' and 'adversarial verifiers.' Google warns that swarms 'consume quota much faster than a standard chat.'
- Not the security version. The model card says Carbon's 'cybersecurity benchmark results reflect performance when accessed through the Fairwind Program,' with 'specialized capabilities available to verified defenders.' The safety report says the public model 'blocks 100% of end-to-end exploit milestones' in Google's tests. The headline of the October launch was state-of-the-art cyber. That is not what the API sells.
Notice what the post doesn't do. It names no competitor. Sundar Pichai's post today says only that 'Gemini 4 Carbon is rolling out today.' On October 28, he, Kavukcuoglu and Logan Kilpatrick each led with first place on the Artificial Analysis index. We found no Google executive repeating that claim today. It stopped being true yesterday.
The paperwork shows its age too. Carbon's model card compares it with GPT-6 Astra, Claude Opus 5.5 and Claude Fable 5.1, with results 'as of October 2026.' Fable 5.1 was replaced a day before the card's model went on sale. Carbon beats Fable 5.1 on every row. That tells you nothing about the question in the headline.
What 'the throne' was, and when Carbon held it
When people say one model dethroned another this autumn, they nearly always mean one chart: Artificial Analysis's Intelligence Index, a composite of nine evaluations that the lab runs itself on every major model. It is the board Google's own launch leaned on. When Carbon was first announced on October 28, limited to security teams, the pitch from Google's side included 'ranks #1 on Artificial Analysis V5.'
That was true for twelve days. Carbon went to the top at 59, one point over Claude Opus 5.5. Then on November 9 Anthropic released Fable 5.2, which scored 65. The Latent Space roundup that day described Anthropic as 'taking back Opus 5.5's #1 spot from Gemini 4 Carbon,' which had its 'brief moment in the sun.' Carbon reached the public one day after losing the title it launched on.
So the question in the headline is really whether today's release changes that. Carbon is not a new model today. It is the same model, now on sale. What changed is who can use it and what it costs, and those turn out to matter more than the five points.
Head to head: every board where both have been tested
On the index itself the order is not close. Artificial Analysis today has Fable 5.2 first, second and third, at max, extra high and high effort. Carbon at its highest setting is fourth at 60. Opus 5.5 is fifth at 59 and OpenAI's GPT-6 Astra eighth at 57.
| Board | Fable 5.2 | Gemini 4 Carbon | Ahead | Who ran it |
|---|---|---|---|---|
| AA Intelligence Index v5.0.1 | 65 (max) / 62 (high) | 60 (high) | Fable | Artificial Analysis |
| FrontierCode 2.0 (hard coding) | 48.8% | 40.6% | Fable | Artificial Analysis |
| HLE-Diamond (expert questions) | 80.2% | 69.1% | Fable | Artificial Analysis |
| AA-Omniscience (knowledge and honesty about gaps) | 67 | 54 | Fable | Artificial Analysis |
| CritPt v2, hard subset (physics) | 76.5% | 69.0% | Fable | Artificial Analysis |
| Terminal-Bench-Science | 73.3% | 64.8% | Fable | Artificial Analysis |
| AA-Briefcase (office work, Elo) | 1785 | 1611 | Fable | Artificial Analysis |
| Multimodal Knowledge Agent | 57% | 54% | Fable | Artificial Analysis, on Google DeepMind's benchmark |
| Vals Index | 71.31% | 70.45% | Fable, inside the margin | Vals AI |
| Terminal-Bench 4.0 (command-line agents) | 59.6% | 65.2% | Carbon | Artificial Analysis |
| GDP.pdf (document work) | 24.2% | 28.4% | Carbon | Artificial Analysis |
| AutomationBench-AA (workflow automation) | 76.4% | 77.8% | Carbon | Artificial Analysis |
| AutomationBench 1.0.6 | 51.45% | 53.88% | Carbon | Zapier |
| CursorBench 4.0 (real coding sessions) | 60.0% (max) | 60.4% (high) | Carbon, by 0.4 | Cursor |
Nine for Fable, five for Carbon. Two cautions before reading that as a score. The first row is built from the nine Artificial Analysis rows beneath it, so those ten lines are one lab's view, not ten votes. And AutomationBench appears twice, once run by Artificial Analysis and once by Zapier, which wrote it.
Still, the split has a shape. Fable 5.2 wins where the task is to know something or reason through something hard: expert questions, physics, difficult code, science at the command line. Its widest lead is on AA-Omniscience, 67 to 54, a test that rewards knowing the answer and punishes confident guessing. Carbon wins where the task is to operate something: a terminal, a document pipeline, a chain of business tools, a real codebase in Cursor. Three of its five wins are by under three points. Terminal-Bench 4.0, at 65.2% to 59.6%, is the convincing one.
That fits how each company describes its model. Google's post says Carbon is for 'complex multi-step coding, debugging, and open-ended scientific research,' and that it 'uses more tokens' than Argon 'to explore, reason, and run verification loops.' It is built to keep going. Fable 5.2 is built to be right.
The part Google is actually selling: cost
Now the other throne. Carbon's API price is $2 per million input tokens and $10 per million output through November 30, then $4 and $20. Fable 5.2 is $8 and $40. Until the end of the month Carbon's tokens cost a quarter of Fable's. After that, half.
Token price is only part of what a task costs, because the models use different amounts of tokens to get there. Artificial Analysis measures the whole bill, and it prices Carbon at the standard $4 and $20, not the introductory rate.
| Model and setting | AA index | Cost per task | Time per task | CursorBench 4.0 | Cursor cost per task |
|---|---|---|---|---|---|
| Fable 5.2, max | 65 | $26.54 | 44.4 min | 60.0% | $32.79 |
| Fable 5.2, high | 62 | $9.08 | 17.4 min | 56.9% | $9.70 |
| Gemini 4 Carbon, high | 60 | $7.76 | 25.6 min | 60.4% | $9.22 |
| Gemini 4 Carbon, medium | 58 | $6.59 | 23.6 min | 60.1% | $8.49 |
| Opus 5.5, max | 59 | $10.04 | 25.1 min | 57.8% | $13.43 |
| Opus 5.5, high | — | — | — | 56.0% | $3.97 |
Read the top line against the third and you get Google's pitch. Fable 5.2 at max costs 3.4 times as much per task as Carbon for five more index points, and takes 73% longer. On Cursor's board it is starker: Carbon's 60.4% costs $9.22 a task, and Fable's 60.0% costs $32.79. Same coding score, 28% of the price. That is the comparison behind the launch post's line about 'more than 2X lower cost per task than the next best external model.' The claim checks out.
One caution on the Cursor figure. Cursor says it applies 'each model's published per-million-token pricing' and doesn't say which Carbon rate that was. Its own price list shows Carbon at the introductory $2 and $10. If the benchmark used that, Carbon's $9.22 becomes roughly $18 on December 1: still 44% under Fable at max, but almost twice Fable at high. Artificial Analysis's figures don't carry that doubt, which is why we lean on them.
But max is not how most people should run Fable, as we found yesterday. Read the second line against the third. At high effort Fable 5.2 costs $9.08 a task to Carbon's $7.76, scores two points higher on the index, and finishes in 17 minutes instead of 26. At nearly the same spend, on the broad index, Fable is still ahead. On Cursor's coding board at nearly the same spend, Carbon is ahead by three and a half points. Two honest boards, two answers, and they split the same way the table above did: reasoning to Fable, operating to Carbon.
One more row deserves a look. Opus 5.5 at high effort scores 56.0% on CursorBench for $3.97 a task. That is four points behind the leader for 43% of the money. The frontier fight is being held at $9 a task and up. A lot of real work is still best done below it.
Google's launch claims, checked
| The claim | What we found |
|---|---|
| #1 on CursorBench 4.0 | True, by 0.4 points over Fable 5.2 at max (60.4% to 60.0%). Cursor's own page: 'small differences in scores may not be statistically meaningful.' |
| #1 on Zapier AutomationBench | True. Zapier's board has Carbon at 53.88% and Fable 5.2 at 51.45%. Anthropic's launch table concedes this one. |
| #1 on Perplexity WANDR | Not checkable against Fable 5.2. No public board lists Carbon. Press reports put Carbon at 85.7%; Anthropic reports 82.1% for Fable 5.2. Different sources, not one run. |
| 'More than 2X lower cost per task than the next best external model' | True at the prices used: 3.6 times on Cursor ($9.22 to $32.79) and 2.2 times on Zapier ($2.12 to $4.64). Zapier notes that Claude's published costs 'use a different basis.' |
| #2 on the Multimodal Knowledge Agent benchmark at 'more than 6X lower' cost | True, and the model in first place is Fable 5.2, 57% to 54%. The costs are from Google's own chart: $5.63 against $35.28. Google DeepMind designed the benchmark; Artificial Analysis ran it. |
| 'Leads' on SEC-bench Pro, ExploitGym and CWE-bench | Those scores are the Fairwind version, not the public model. Google ran two of the three itself. No Fable 5.2 figures exist for any of them. |
So the cost claims hold and the rankings hold where they can be checked. The one place Google admits second place, it is second to Fable 5.2, on a benchmark Google wrote.
Anthropic's side is no cleaner. Its launch page does include Carbon, and reports Fable 5.2 ahead on FrontierSWE (69.9 to 61.2), on computer use in OSWorld (89.3 to 81.5) and on science at the terminal (74.6 to 64.8). Those are Anthropic's numbers. On FrontierCode it sets Fable at extra-high effort against Carbon at medium, which isn't a like-for-like setting. Each company's table is a list of the benchmarks it wins. That is why the table above uses neither.
What hasn't been tested yet
A day after one launch and hours after the other, the boards are half full. This is where we would be most careful about anyone's headline, including ours.
- Arena's text leaderboard has Carbon first and no Fable 5.2 at all. Carbon leads at 1526, flagged 'Preliminary' on 4,282 votes, four points over Gemini 4 Argon. Fable 5.2 has not been ranked. Arena scores are human preference votes, so it takes days of traffic to place a new model.
- Three boards have Fable 5.2 first and no Carbon. LiveBench has Fable 5.2 at 84.9. The ARC Prize board lists it at 93.8% on ARC-AGI-2. Epoch's capabilities index has it first at 174. Carbon was not available to these labs until today.
- Nobody has published a hands-on comparison. We looked for one from the testers we normally cite and found none as of this afternoon. The first-day write-ups that pick a winner are built from the same leaderboards as this page.
- We have not run one either. This post is a reading of other people's measurements, each linked. When we test, we will say so.
The Arena gap matters for a reason beyond pride. A Polymarket contract asking which company will have the best AI model at the end of November has Google at 76% and Anthropic at 23%, on $1.18 million of volume. It resolves on Arena's text board. Traders are not saying Carbon is the better model. They are pricing the fact that Carbon is on the deciding chart and Fable 5.2 is not there yet. If you see that number quoted as a verdict this week, that is what it measures.
Early reaction from people with access was more about access than quality. On the Hacker News thread, one commenter said Carbon's answer on a paper draft 'surprised me with some interesting points all the other AI's missed.' Another wrote: 'We have an enterprise account and we can't use this.' A third asked the question Google's own post invites: 'It uses more tokens than Argon?'
Who can use each one today, and for how much
| Gemini 4 Carbon | Claude Fable 5.2 | |
|---|---|---|
| API price per million tokens | $2 in, $10 out to Nov 30; $4 and $20 after | $8 in, $40 out |
| Cached input | $0.10, then $0.20 | $0.20 |
| Context and output | 1M context; 65,536 output by default | 1M context; 128,000 output |
| Settings | Low, medium (default), high | Low, medium, high, extra high, max |
| Cheapest subscription that includes it | Google AI Ultra, from $99.99 a month | Claude Max, $100 a month, for up to half the weekly limit |
| On the $20 plan | No. AI Pro has Argon | Paid with usage credits (included up to half the week through Nov 16) |
| Where it runs | Gemini app, AI Mode, AI Studio, Antigravity, Android Studio, the API | Claude apps, Claude Code, the API |
For developers paying by the token, the gap is as wide as it looks. For everyone else it nearly vanishes: the cheapest way to get either model in an app is about $100 a month. Both meter you inside that. Google's plan page says 'your limit refreshes every 5 hours until you reach your weekly limit,' and Anthropic caps Fable at half the week, which we measured against one heavy user's logs earlier today.
So which one?
- Research, analysis, hard debugging, anything where a wrong answer is expensive: Fable 5.2, starting at high effort. It leads the broad index, the hard-coding test and the knowledge test, and at high it costs about what Carbon does.
- Agents that run in a terminal, in Cursor, or across business tools: Carbon, at medium or high. Medium is the default and gives up almost nothing: 60.1% on CursorBench against 60.4% at high, for less money.
- You pay by the token and volume matters: Carbon until November 30, while it is a quarter of Fable's price. After that, test Opus 5.5 at high effort before paying for either flagship. 56% on CursorBench for $3.97 a task is hard to argue with.
- You already pay for a $100 plan: stay where you are. Five index points is not worth moving your files, habits and history to another company's app, and the lead has changed hands twice in two weeks.
- Security teams: the public Carbon is not the one in Google's cyber charts. The Fairwind Program is still the door, as we covered in October.
- Not in a hurry: wait a week. Arena, LiveBench, ARC and Epoch will each have both models, and somebody will have run them on the same real task.
Our read
No, Carbon has not dethroned Fable 5.2, and it didn't need to. Fable 5.2 leads by five points on the index both companies have used to claim first place, and by more on the tests of hard reasoning. Carbon went on sale today a point ahead of everything except Fable, at half the token price or less, and first on the two boards closest to everyday agent work. For the first time this cycle, the two things people mean by 'best' belong to different companies: the highest ceiling is Anthropic's, and the best price near the ceiling is Google's.
Three things would change this page, and we will update it for each: Fable 5.2 arriving on Arena's text board; Carbon arriving on LiveBench, ARC and Epoch; and December 1, when Carbon's price doubles. Rankings: no movement today. Claude and Gemini keep their places, the Gemini review's Carbon answer is updated, and scores are due at the weekly refresh. The Token Counter now carries Carbon alongside Fable 5.2 if you want to price your own prompts on both.
Where these apps rank
