Gemini 4 Carbon vs Claude Fable 5.2: Benchmarks, Price and Which One to Pick

The short answer

Claude Fable 5.2 is the smarter model on the independent leaderboards, and Gemini 4 Carbon is the cheaper and faster one that wins Cursor. As of November 11, 2026, Artificial Analysis scores Fable 5.2 (max) at 65 on its Intelligence Index v5 against 60 for Carbon (high), with Fable ahead on science, knowledge and hard coding. Cursor’s own CursorBench 4.0 flips the order slightly, 60.4% for Carbon at high effort against 60.0% for Fable 5.2 at max, and Fable’s run costs $32.79 per task against $9.22.

So the decision is mostly a price-for-accuracy trade. Pick Fable 5.2 when the task is hard science, expert-level knowledge, or a difficult coding problem where an extra few points are worth roughly three times the spend. Pick Carbon when you want near-top agentic coding, terminal work or spreadsheet-style automation at a lower cost per finished task, or when you need audio and video input. And before you pay for either, read the part both vendors put in their own announcements: Anthropic recommends starting with Claude Opus 5.5, and Google says Gemini 4 Argon is the better default for most work.

This is post three of a cluster. For one model at a time, see the Claude Fable 5.2 complete guide and the Gemini 4 Carbon complete guide. Every benchmark cell below comes from an independent leaderboard we fetched on November 11, 2026, with the board named. Anything a vendor published about itself is labelled vendor-reported, and the YouTube video below is treated as a secondary source and attributed.

Video: a creator roundup of 58 test videos and the public leaderboards, published right after both launches. Treated here as commentary, not as a data source.

Which one for which job

The table gives the pick first and the evidence second. “Split” means the independent boards disagree, and the honest answer is to test on your own work. All scores are as of November 11, 2026 unless the row says otherwise.

Specs, access and prices

The two models are matched on context and mismatched on price. Both offer a 1M-token context window. Fable 5.2 lists at $8 per million input tokens and $40 per million output tokens, down 20% from Fable 5.1’s $10 and $50. Carbon lists at $4 and $20 but sells at an introductory $2 and $10 through November 30, 2026, the same schedule as Gemini 4 Argon. On December 1 the Carbon gap to Fable narrows from four times to two times on raw token prices.

Sources for the table are Anthropic’s announcement and model overview, Google’s availability post, what’s new page and pricing page, plus Artificial Analysis for modalities. Two details deserve care. First, Fable 5.2’s launch promo and exact cache-write rates sit on Anthropic’s pages and we do not repeat them here. Second, Artificial Analysis lists Carbon at $4 and $20, so its Carbon cost-per-task figures appear to reflect the post-December list price, not the introductory one. That is our reading of its page, not a statement from Artificial Analysis.

Independent benchmarks

Both vendors published charts that favor themselves. The video below notes that Google’s chart compares against the older Fable 5.1, and that Anthropic’s table runs Carbon at medium effort on one coding test (we confirmed the second point in Anthropic’s table). So the tables here use boards that run both models on the same tasks. Each cell names its board and date.

Artificial Analysis Intelligence Index v5.0.1

Artificial Analysis combines nine evaluations into one index. Its head-to-head page (fetched November 11, 2026) compares Fable 5.2 at max effort with its default fallback enabled against Carbon at high effort, which is Carbon’s top setting. Fable 5.2 takes the index 65 to 60. Google’s October 28 post put Carbon at 59 on the index at launch, so Carbon’s own score has moved up by a point since then.

Fable 5.2 wins six of the nine component evaluations and Carbon wins three. The pattern is clean: Fable leads wherever the task rewards deep reasoning, recall and scientific work (HLE, Omniscience, CritPt, Terminal-Bench-Science), plus FrontierCode and Briefcase. Carbon leads on the shell-and-spreadsheet side (Terminal-Bench 4.0, AutomationBench-AA, GDP.pdf). Artificial Analysis also noted that Fable 5.2’s high and xhigh settings score 62, ahead of every other model at any effort, using about 98k output tokens per task against 173k for Opus 5.5 at max, so Fable does not need max effort to lead.

CursorBench 4.0

CursorBench 4.0 uses ambiguous, multi-file tasks from real Cursor sessions. It is run by Cursor, which sells access to both models, so it is independent of the two labs but not of Cursor’s harness. Cursor’s own board was fetched on November 11, 2026.

Two things stand out. First, the 0.4-point gap between Carbon high and Fable max is tiny and should not be read as a real winner; the real separation is cost, $9.22 against $32.79, and steps, 71 against 218. Second, at the same price Carbon wins clearly: Carbon high costs $9.22 for 60.4% while Fable 5.2 high costs $9.70 for 56.9%. Carbon’s weakest run, low effort at 58.6%, still beats every Fable 5.2 setting except max. Cursor’s own post says Fable 5.2 at default effort scores 56.9%, up from 49.2% for Fable 5.1.

Terminal-Bench 4.0 on Vals

Vals’ Terminal-Bench 4.0 board, updated November 9, 2026, uses the Mini-SWE-agent harness. Fable 5.2 and Opus 5.5 tie for first at 65.15%, Sonnet 5.5 sits at 64.14%, and Carbon scores 63.13%. Carbon’s task cost was $7.45 against $14.56 for Fable. Artificial Analysis, running its own harness, has Carbon ahead 65.2% to 59.6%. Two boards, two orders, same conclusion: this is a statistical tie that Carbon reaches for less money.

One caveat from RidgeBench’s summary of the Vals run: Vals says 17 of 198 Fable attempts were served by Opus 5 or Opus 4.8 because of Anthropic’s fallback safeguards. That is a secondary report of Vals’ note, so we label it as such, but it is a reminder that Fable scores on security-flavored tasks include some answers from older models.

Arena and other boards

Arena’s new-release panel on November 11, 2026 listed Claude Fable 5.2 as number one in WebDev at xhigh effort. RidgeBench, summarizing Arena+, reports Fable at 1538 against Carbon at 1528, a ten-point Elo gap that is within noise for casual use. We did not retrieve a full Arena text-board table for both models, so we do not give text-arena scores. Arena’s agent board still showed Opus 5.5 first and Carbon (high) second at that date, with Fable 5.2 not yet listed.

Two newer Artificial Analysis boards add color. On the Multimodal Knowledge Agent benchmark that Google DeepMind built with Artificial Analysis, Fable 5.2 (max) leads with a 57% Climb score, Carbon (high) is second at 54% and Opus 5.5 third at 53%. Cost per task is $31.97 for Fable and $6.29 for Carbon, about five times more for roughly three extra points. Artificial Analysis also reports that no model leads all ten industry verticals: Fable leads five, Carbon leads Healthcare. On VIALS-AA, a visual reasoning test for life-science images, Fable 5.2 scores 67% against 64% for Carbon and 63% for GPT-6 Astra. We saw both results on Artificial Analysis’ X account on November 10 and 11, 2026.

Vendor-reported claims

These numbers come from each lab’s own launch material. They are useful for seeing what each company wants you to notice, but none of them is a head-to-head under matched conditions.

When a vendor number and an independent number measure the same thing, they mostly agree here. Terminal-Bench-Science is the cleanest example: Anthropic’s 64.8% for Carbon matches Artificial Analysis’ 64.8% to the decimal, and the Fable number differs by only 1.3 points. That is a good sign for the quality of the published tables, with the standing exception of effort-level choices like the medium-effort Carbon run.

Cost per task, not per token

Token prices mislead because the two models spend very different numbers of tokens. Artificial Analysis measured 274k output tokens per index task for Fable 5.2 at max, 193k of them reasoning, against 123k output and 72k reasoning tokens for Carbon at high. Fable is both more expensive per token and wordier at max effort, so the two effects multiply.

Cursor’s run tells the same story with a different harness: $32.79 per task for Fable 5.2 max against $9.22 for Carbon high, a 3.6x gap for essentially the same score. Vals’ Terminal-Bench run has $14.56 against $7.45, about 2x.

Worked example: one token-heavy job

Take a job that reads 1 million input tokens and writes 100,000 output tokens, with no caching. At Fable 5.2 list prices that is $8.00 for input plus $4.00 for output, $12.00 total. On Carbon at the introductory price it is $2.00 plus $1.00, $3.00 total, one quarter of the cost. After December 1, Carbon’s $4 and $20 gives $4.00 plus $2.00, $6.00, half of Fable’s. Fable’s fast mode ($16 and $80) would double that job to $24.00 in exchange for up to 2.5x faster output. These are arithmetic on list prices. Real jobs differ because Fable at max tends to think longer, so the practical gap is usually larger than the per-token gap.

The effort-level twist

Cursor’s board shows a useful pattern. Fable 5.2 at medium effort scores 54.8% for $4.66 per task and at low scores 53.2% for $3.94. Those are cheap, but Carbon at low (58.6%, $11.89) costs more than Fable at low and scores higher, and Carbon takes 124 steps against Fable’s 30. So Carbon is not always the cheaper model per task: at low effort, Fable can finish with fewer steps and lower spend. The video says Artificial Analysis found Fable cheaper at low effort where both score the same, at $310 against $719 for the index. We could not match those figures on the pages we fetched, so treat that as the video’s claim, not a measured fact.

Speed and latency

Artificial Analysis’ latency numbers favor Carbon at top settings. Carbon (high) takes 25.6 minutes per index task against 45.2 for Fable 5.2 (max), and its time to first answer token is 67 seconds against 234. The mean end-to-end time for single-request workloads is 156 seconds against 411. Fable has the higher raw output speed, 101 tokens per second against 80, so once it starts answering it streams faster. The wait comes from how long it reasons before the first answer token.

If latency matters, Fable has two levers Carbon lacks. Lower effort settings cut its thinking time, and fast mode on Anthropic’s platform delivers up to 2.5x faster speed at twice the price ($16 and $80). For interactive use, expect Carbon to feel snappier at high effort and Fable to feel deliberate at max.

The launch posts

Five first-party posts frame the launch. Anthropic announced Fable 5.2 and its restricted sibling Mythos 5.2 on November 9, 2026. Artificial Analysis posted the index result the same evening and called Fable 5.2 number one at 65, five points ahead of Carbon at 60.

Cursor added Fable 5.2 the same day. Its post says Fable 5.2 is the top model on CursorBench at max effort, 60%, and scores 56.9% at default effort, up from 49.2% for Fable 5.1. It also advises starting at the default effort, because effort levels behave differently from Fable 5.1’s.

Google answered a day later. Google DeepMind’s November 10 post pitches Carbon as its strongest model for complex, long-horizon tasks, and Cursor then announced that Carbon at high effort overtook Fable on its board at 60.4%, with 50% off through November 30.

What the video says

The video, Gemini 4 Carbon vs Claude Fable 5.2: The Internet Tested Both (Who Actually Wins?), is the most useful commentary we found. Its creator reviewed 58 test videos and scored the matchup in rounds. We treat it as attributed commentary and have checked each number against the primary boards above.

The video’s own conclusion is two rounds each and two ties: Fable is the sharper brain, Carbon wins code by a hair and costs less at full power, and both companies tell most people to start with their everyday models. Its caveat is the most important line: Carbon was a day old, and only one creator had given both models the same task. We agree with that framing. The tally is a story, not a measurement.

Both vendors say start elsewhere

Neither company positions its flagship as the default. Anthropic’s announcement says that for most everyday coding and knowledge work, it recommends starting with Claude Opus 5.5, and its API docs tell developers to start with the latest Opus model and move to Fable 5.2 when evals on Opus at high effort fall short. Google’s post says Argon remains the ideal choice for the majority of software engineering and knowledge work, and describes Carbon as the model for tasks that test the limits, using more tokens than Argon to explore, reason and verify.

The independent numbers back that up. On CursorBench 4.0, Opus 5.5 at high scores 56.0% for $3.97 per task, which is nearly the score of Fable 5.2 at high (56.9%) at 41% of the cost. Gemini 4 Argon at high scores 52.9% for $4.57. If a 4 to 8 point lift on Cursor’s tasks does not change your outcomes, the everyday model is the better buy. See our Opus 5.5 vs Sonnet 5.5 coding comparison, the Gemini 4 Argon guide, and the broader best AI model for coding roundup for where each fits.

For prompting and benchmark-reading habits that carry over from the previous generation, see the Claude Fable 5 benchmark and prompting guide and the earlier Claude Fable 5.1 guide (English only). For price tiers outside this pair, the GPT-6 Astra vs Sol vs Luna pricing guide is the closest comparison. GPT-6 Astra (max) scores 57 on the Artificial Analysis index, behind both models here.

Cyber and safeguards

Security work is where the two models differ most in how you can access them, not in what they score. Fable 5.2 ships with safeguards for cybersecurity and biology. Anthropic says many queries in those domains are automatically routed to less capable models when flagged, and teams with legitimate work can apply to its Cyber Verification Program or Life Sciences Verification Program. The stronger cyber model is Claude Mythos 5.2, which only approved organizations can reach. Anthropic’s own table shows Mythos 5.2 at 76.3% on CyScenarioBench and 89.6% on CVE Rediscovery, labelled vendor-reported.

Google takes the opposite route for cyber. Carbon for cyber is available to defenders through the Fairwind Program, and Google reports SEC-bench Pro at 88.2%, ExploitGym at 54.8% and CWE-bench at 71.0% for that access, all vendor-reported. The public Carbon model on the API is a different access tier, and RidgeBench notes it has found no public cyber score for it. Artificial Analysis’ cyber board has no Fable 5.2 entry yet either. So for security research, there is no fair public head-to-head, and the practical question is which vendor’s verification program you can get into.

One more difference to plan around is data retention. Anthropic says using Fable models requires 30-day data retention for safety monitoring by default, with Enterprise Frontier Safeguards as the route to zero-data-retention compatibility. Check that against your compliance rules before you pick Fable for sensitive data.

Run your own bake-off

A benchmark average hides your workload. A one-afternoon test is cheaper than a wrong default.

1. Collect 20 real tasks from your repository or documents, including five that your current model fails.
2. Run each task on Opus 5.5 or Argon first, at the vendor-recommended default, and record pass or fail, tokens and wall-clock time.
3. Run only the failures on Carbon at high and Fable 5.2 at high, using the same prompts and tools.
4. For anything Fable passes and Carbon fails, repeat once at xhigh or max to see whether effort or model is the real difference.
5. Divide total spend by tasks passed. That cost per solved task is the number to compare, not the price per million tokens.
6. Re-run in December. Carbon’s price doubles on December 1 and both models will keep changing.

If you work in Cursor, you can pick effort levels per run, and Cursor’s public board lets you compare every effort level of both models before you spend anything.

Limitations

• Day-old data. Carbon was generally available for about a day when these numbers were collected. Leaderboards are updated, and Carbon’s Artificial Analysis score has already moved from 59 on October 28 (Google’s post) to 60.
• Different settings. Artificial Analysis compares Fable at max with Carbon at high because Carbon has no max setting. Cursor shows no Carbon max run for the same reason. Matched-effort comparisons are only partial.
• Different harnesses. Artificial Analysis, Vals and Cursor each use their own agent harness, and their orderings on Terminal-Bench and coding differ.
• Fallback contamination. Fable’s safeguards can route some security-adjacent attempts to older models, which affects score interpretation on cyber-flavored tasks.
• Gaps. We did not retrieve a full Arena text-board table, no independent cyber score exists for either public model, and there is no matched-effort head-to-head on games and 3D.
• Pricing changes. Carbon’s introductory price ends November 30, 2026. Fable 5.2’s launch promo terms are on Anthropic’s page and may change.
• Press and creator sources. RidgeBench, VentureBeat and the YouTube video are secondary sources. We use them for context and label every number we take from them.

Verdict

If you can only pay for one frontier model today, the answer turns on what a mistake costs you. Fable 5.2 is the stronger reasoner: it leads the Artificial Analysis index by five points and wins science, expert knowledge and hard coding, and its high and xhigh settings already lead without needing max. Carbon is the better-value agent: it matches Fable on CursorBench, ties it on terminal work, wins three of nine Artificial Analysis components, and finishes tasks in about a third of the spend and roughly half the time.

A practical split works for most teams. Keep Opus 5.5 or Argon as the daily driver. Use Carbon when a coding or automation agent needs more headroom at a cost you can predict, and reach for Fable 5.2 when a hard scientific or knowledge task justifies the premium. Revisit the call after November 30, when Carbon’s price doubles and the gap on raw token cost shrinks from four times to two. For each model on its own, the Claude Fable 5.2 complete guide and the Gemini 4 Carbon complete guide cover access, API usage and migration in detail.

FAQ

Sources

Sources checked November 11, 2026. Leaderboard values change; check the live boards before you decide.

• Artificial Analysis: Claude Fable 5.2 (Max, Default Fallback) vs Gemini 4 Carbon (High), with the Fable 5.2 and Gemini 4 Carbon model pages (index, component scores, cost, speed, modalities)
• Cursor: CursorBench 4.0 (scores, cost and steps per task)
• Vals AI: Terminal-Bench 4.0 leaderboard, updated November 9, 2026
• Arena leaderboard (new-release and agent panels)
• Anthropic: Claude Fable 5.2 and Claude Mythos 5.2 and the Claude API model overview (vendor-reported benchmark table, prices, safeguards, plan access)
• Google: Gemini 4 Carbon is now available, Introducing Gemini 4 Carbon, what’s new in Carbon, model page, pricing and the DeepMind model card (vendor-reported)
• X posts: @AnthropicAI, @ArtificialAnlys, @cursor_ai (Fable 5.2), @GoogleDeepMind, @cursor_ai (Carbon)
• RidgeBench: Fable 5.2 takes the top of the index the day Carbon goes on sale and VentureBeat: Gemini 4 Carbon is now publicly available (press and secondary analysis)
• Gemini 4 Carbon vs Claude Fable 5.2: The Internet Tested Both (Who Actually Wins?) (YouTube, creator roundup; secondary source)