Claude Fable 5.2 tops Artificial Analysis AA-3D benchmark

Claude Fable 5.2 tops Artificial Analysis AA-3D benchmark

Claude Fable 5.2 has taken first place on Artificial Analysis' AA-3D benchmark with a score of 1,271 Elo, Crypto Briefing reported. Anthropic unveiled the model on November 9, 2026.

The ranking matters because AA-3D assesses models on their ability to create 3D experiences, according to Crypto Briefing. The publication said two other Claude models occupy the next two places, putting Anthropic's models across the top three.

The benchmark is one of several measures on which the new model has been evaluated, and its results do not establish how it will perform on every project. Crypto Briefing reported that AA-3D covers 55 tasks and cautioned that a strong test score does not automatically carry over to all real-world work.

How Claude Fable 5.2 leads AA-3D

The 1,271 Elo result puts Fable 5.2 ahead of Claude Sonnet 5.5 by 67 points, according to Crypto Briefing. The publication said Sonnet 5.5 and Claude Opus 5.5 are separated by two points on the same benchmark.

Crypto Briefing reported that Artificial Analysis gives Fable 5.2 a 95% confidence interval of 1,246 to 1,297 and lists the result as 1,271 ±25. The reported interval is part of the score's presentation, rather than a separate test result.

The new model also scored 132 points more than its predecessor, Fable 5.1, Crypto Briefing reported. Fable 5.1 had an AA-3D score of 1,139, placing it below GPT-6 Astra and Gemini 4 Argon, according to the publication.

In the current comparison, GPT-6 Astra is 103 points behind Fable 5.2, while Gemini 4 Argon trails by 124, Crypto Briefing reported. The publication said Artificial Analysis serves as an independent evaluator of the models.

The tasks in AA-3D span areas that include games and natural phenomena, according to Crypto Briefing. The publication said the benchmark uses Elo ratings to compare performance within the model pool.

Other benchmark results

Claude Fable 5.2 also led Artificial Analysis' Intelligence Index v5 with 65 points, THE DECODER reported. The publication put it five points ahead of Gemini 4 Carbon and 15 ahead of Fable 5.1 on that index.

THE DECODER reported that the 65-point result was recorded at the highest compute tier. It said Gemini 4 Carbon scored 60 at the High tier and Claude Opus 5.5 scored 59 at the Max tier.

Fable 5.2 led five of the index's nine individual tests, according to THE DECODER. The publication reported scores of 80% on HLE-Diamond, which tests expert-level academic questions, and 73% on Terminal-Bench-Science, which tests independent scientific workflows in a terminal.

On AA-Briefcase, which measures multistep knowledge work, Fable 5.2 scored 1,785 Elo, below Sonnet 5.5's 1,813, THE DECODER reported. The publication said Fable's score was within the confidence interval of Opus 5.5, which scored 1,800.

The model also led a finance and accounting capability index among 27 models compared, according to LavX News. The outlet reported that Fable 5.2 at Max ranked first for intelligence among 74 models, while placing 73rd of 73 on cost and 72nd of 73 on both speed and verbosity.

Pricing and access

Standard pricing for the new models is $8 per million input tokens and $40 per million output tokens, less than their predecessors. THE DECODER reported that an Intelligence Index task at the highest tier cost an average of $26.50 for Fable 5.2, compared with $10 for Opus 5.5 and $12.99 for Fable 5.1.

The models generate output up to 30% faster than their respective 5.1 predecessors, THE DECODER reported, citing Anthropic. The publication said an optional fast mode can increase output speed by up to 2.5 times at twice the cost, but is limited to select organizations and certain platforms.

Claude Fable 5.2 is available across Claude's platforms as well as Amazon Web Services, Google Cloud and Microsoft Azure, according to TestingCatalog AI News. The outlet said developers can access it on the Claude Platform using the identifier claude-fable-5-2.

Fable 5.2 and Mythos 5.2 share a model foundation but have different access arrangements, THE DECODER reported. Claude Platform Docs says Mythos is available to organizations verified through Anthropic programs and has the same specifications and pricing as Fable.

According to TestingCatalog AI News, organizations working in cybersecurity and life sciences that have been vetted may seek access to Mythos through the verification programs run by Anthropic. Claude Platform Docs says Fable requires 30-day data retention and cannot be used under zero data retention without express authorization from Anthropic.

Safeguards and technical work

Fable 5.2 is available to all users with cybersecurity and biology safeguards against misuse. Crypto Briefing reported that Fable has stronger safeguards against misuse than Mythos 5.2.

THE DECODER reported that Fable blocks most security-related tasks, including penetration tests and exploit code generation. In early testing, its safeguards produced about 55% fewer blocks than those of Fable 5.1, TestingCatalog AI News reported, citing Anthropic.

Claude Platform Docs describes Claude Fable 5.2 as designed for demanding reasoning and long-running agentic work. THE DECODER reported that Anthropic continues to recommend Opus 5.5 for everyday use while positioning the new models for more demanding technical tasks.

Anthropic has also described results from its own testing of the new models. According to Mashable SEA, Anthropic said they "designed three working electronic devices from scratch, solved two open problems in the physics of black holes…and formalized the proof of Fermat’s Last Theorem in 30 hours."

THE DECODER reported that Anthropic used about 1,000 Fable 5.2 agents to translate the theorem's proof into the Lean programming language for step-by-step computer verification. The publication said Anthropic reported that the agents organized their own collaboration after brief instructions and did not use mathematical references.

Conclusion

Crypto Briefing's reported AA-3D standings place Fable 5.2 first, ahead of two other Claude models, on a test of 3D experience generation. Other results reported by THE DECODER put the model first on Artificial Analysis' Intelligence Index, though the same publication recorded a higher average task cost than for Opus 5.5 or Fable 5.1 at the highest tier. For access, TestingCatalog AI News reports that developers can use Fable through the Claude Platform, while Anthropic's verification programs offer an application route for vetted organizations in cybersecurity and life sciences.

Frequently Asked Questions

What score did Claude Fable 5.2 get on AA-3D?

Crypto Briefing reported that Claude Fable 5.2 ranked first with 1,271 Elo on Artificial Analysis' AA-3D benchmark. The publication said the benchmark measures models' ability to build 3D experiences and covers 55 tasks. It also reported a 95% confidence interval of 1,246 to 1,297 for Fable's score.

How far ahead of Sonnet 5.5 is Fable 5.2 on AA-3D?

Fable 5.2 leads Sonnet 5.5 by 67 Elo points on AA-3D, according to Crypto Briefing. The publication reported that the two positions immediately below Fable belong to other Claude models and that Sonnet 5.5 and Opus 5.5 are separated by two points.

How much does Claude Fable 5.2 cost?

Anthropic charges $8 for each million input tokens, with the output rate set at $40 for the same volume. The corresponding rates for the preceding models were $10 and $50, respectively, according to THE DECODER. The publication also reported that a fast mode for select organizations costs twice as much.

Who can access Claude Mythos 5.2?

Claude Platform Docs says Mythos 5.2 is available to organizations verified through Anthropic's programs and shares Fable 5.2's specifications and pricing. TestingCatalog AI News reported that vetted cybersecurity and life sciences organizations can apply through those programs, while developers can access Fable on the Claude Platform.

Sources

Read Also

Gemini 4 Carbon Tops AI Arenas as Google Claims Frontier Lead

Gemini 4 Carbon Tops AI Arenas as Google Claims Frontier Lead

Google DeepMind introduced Gemini 4 Carbon on October 28, 2026, and the model immediately secured top positions in several competitive AI arenas, according to Crypto Briefing. Google described the system as our next frontier model and our strongest model, claiming gains across coding and agentic tasks as well as cybersecurity defense, scientific work, full-stack web development and multimodal understanding, Shattered reported.

The debut matters because Gemini 4 Carbon is the second entry in the fourth numbered generation of Gemini, following Gemini 4 Argon, and Google is positioning it at the top of that line, according to Shattered. The model ranked first in the Text Arena, second in the Agent Arena and fourth in the Code Arena: WebDev, Crypto Briefing reported, results that establish it as a strong contender across general, agentic and coding evaluations.

Gemini 4 Carbon Arena Rankings and Index Score

Beyond the arena placements, Google says Gemini 4 Carbon holds the number one spot on version 5.0 of the Artificial Analysis Intelligence Index with a score of 59. On WebDev Arena specifically, the model gained 89 Elo points to reach fourth place at 1769, according to MarkTechPost.

The rankings carry a caveat: Arena updates its standings continuously, so Gemini 4 Carbon's initial showing may shift as ongoing assessments accumulate, Crypto Briefing reported. Market participants are likely to watch future Arena updates for any movement in the model's positions, the outlet said.

The composite index score also needs context. Google did not state the gap between Gemini 4 Carbon and the runner-up, so first place could reflect a wide lead or a narrow one, Shattered reported, and only the Artificial Analysis ranking comes from a named outside index, with Google itself doing the citing.

Security Benchmarks Lead the Launch

Google reported one benchmark for each stage of a vulnerability's life: finding it, exploiting it and fixing it. Gemini 4 Carbon posted 88.2% on SEC-bench Pro for discovery, 54.8% on ExploitGym for exploitation and 71.0% on CWE-bench for patching, according to figures reported by MarkTechPost and Shattered.

The 88.2% discovery score is the highest of the three security figures, and Shattered reported it suggests the model can triage large volumes of code, which matters to teams sitting on backlogs of unreviewed dependencies. Google frames Gemini 4 Carbon around the full defense lifecycle of find, validate, then fix, according to MarkTechPost.

Against rivals, Google says Carbon's 88.2% on SEC-bench Pro compares with 85.4% for GPT-6 Astra, while its 54.8% on ExploitGym stands against 42.4% for GPT-6 Astra and 34.5% for Opus 5.5, MarkTechPost reported. On Collinear's CWE-bench, Carbon's 71.0% edges the 68.0% posted by both Argon and GPT-6 Astra, the outlet said.

Google also claims Gemini 4 Carbon delivers more than twice the success-rate uplift of Argon on secure-code generation in its internal benchmarks. That figure is an internal Google measurement, so outside teams cannot yet verify it, Shattered reported, though MarkTechPost corroborated the claim of more than 2x higher success at generating secure code.

The Google team also states that the Chrome team found Gemini 4 Carbon more than 10 times better than another lab's model at producing high-quality patches, and security firm Wiz reports Carbon leading its web penetration testing benchmarks, according to MarkTechPost.

Agentic Coding and Long-Horizon Work

Google says its internal teams watched Gemini 4 Carbon work for days on complex, multi-step engineering tasks. MarkTechPost reported the same, noting examples such as root-causing and patching subtle bugs across large codebases and low-level systems.

In agentic coding benchmarks, Gemini 4 Carbon scores 65.2% on Terminal-Bench 4.0 (AA), ahead of Opus 5.5 at 59.6%, and reaches 77.4% on Agents' Last Exam v1.1, according to MarkTechPost. On FrontierSWE, however, GPT-6 Astra leads at 65.5% versus Carbon's 61.2%, the outlet reported.

On Vending-Bench 2, a simulation developed by Andon Labs in which each model operates a vending machine business across 365 simulated days from a starting balance such as $500, Gemini 4 Carbon finished with a bank balance of $16,484, the highest listed. OfficeChai reported that the result topped rival models including GPT-6 Astra and Claude Opus 5.5 and marked the first time in over a year that a Google model took the top position on that benchmark.

OfficeChai said the $16,484 balance shows Google maintaining long-term context without degrading performance, optimizing supplier negotiation strategies and executing compounding financial decisions over months of simulated time. The outlet added that the performance signals Google is ready to compete at the frontier of enterprise automation as businesses look to deploy autonomous agents on complex workflows.

On integrity, Google says successful cheating on Terminal-Bench 4.0 dropped to 0.0% while cheating attempts fell to 0.4%, with overall accuracy improving, according to Shattered. Google also reports that high-severity misalignment alerts in its internal coding environments fell by almost 80%, a figure reported by both MarkTechPost and Shattered.

Staged Access Through the Fairwind Program

Gemini 4 Carbon starts with a vetted group of cyber defenders rather than the public API, with Google saying the model would first reach those defenders during the announcement week. Access then expands gradually to trusted partners through the Fairwind Program, Shattered reported.

For everyone else, Google's line is that Gemini 4 Carbon is "coming soon to developers, businesses, and consumers," a sentence that carries no date, according to Shattered. MarkTechPost likewise reported that developers, businesses and consumers get access later, with no date given.

The gated approach mirrors a rival's playbook: Anthropic limited early Mythos access to a set of organizations before widening the group further, Shattered reported. When two of the largest labs both stage their strongest security-capable models this way, the practice starts to look like an industry norm rather than a one-off caution, the outlet said.

What Google Did Not Disclose

The announcement confirms capability claims but leaves logistics open. There is no price, no context window, no parameter count and no general release date for Gemini 4 Carbon, Shattered reported, and MarkTechPost confirmed the parameter count and context window were not disclosed.

For comparison, the model's predecessor Gemini 4 Argon has a 1,048,576-token context, according to MarkTechPost. Gemini 4 Carbon was run at the highest thinking setting in Google's published comparisons, and most rival numbers in those tables are self-reported by each provider under Google's evaluation methodology, the outlet said.

The commercial stakes are real: enterprises choose cloud providers partly on model access, so a leading Gemini model can pull workloads toward Google's platform, Shattered reported. But contracts renew on annual cycles, and Gemini 4 Carbon does not yet have a price or a general release date, which means anyone planning a migration or procurement cycle around it today is planning around a promise rather than a ship date, the outlet said.

Google has not set a price, and how it positions Gemini 4 Carbon against rivals will shape adoption, according to Shattered. The most likely disappointment, the outlet reported, would be a price that places the model out of reach for small teams.

Conclusion

Gemini 4 Carbon arrives with top arena placements, a number one ranking on version 5.0 of the Artificial Analysis Intelligence Index and vendor-reported security scores of 88.2%, 54.8% and 71.0% across discovery, exploitation and patching benchmarks. What it does not yet have is a price, a disclosed parameter count or a general release date.

The next concrete step is the staged rollout: cyber defenders get access during the announcement week, trusted partners follow through the Fairwind Program, and Google says the model is coming soon to developers, businesses and consumers without committing to a date. Continuous Arena updates will show whether the debut rankings hold.

Frequently Asked Questions

What is Gemini 4 Carbon and when did Google announce it?

Gemini 4 Carbon is Google DeepMind's latest frontier model, announced on October 28, 2026. Google called it our next frontier model and our strongest model, claiming gains in coding, agentic tasks, cybersecurity defense, scientific work, full-stack web development and multimodal understanding. It is the second model in the Gemini 4 family, following Gemini 4 Argon.

How did Gemini 4 Carbon rank in the AI arenas?

According to Crypto Briefing, Gemini 4 Carbon ranked first in the Text Arena, second in the Agent Arena and fourth in the Code Arena: WebDev at debut. Arena updates its rankings continuously, so those positions may shift over time. Google also says the model holds the top spot on version 5.0 of the Artificial Analysis Intelligence Index with a score of 59.

What security benchmark scores did Google report for Gemini 4 Carbon?

Google reported 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for exploitation and 71.0% on CWE-bench for patching, according to MarkTechPost and Shattered. Google also claims more than twice Argon's success-rate uplift on secure-code generation in internal benchmarks, a figure outside teams cannot yet verify.

When will developers and consumers get access to Gemini 4 Carbon?

Google said Gemini 4 Carbon would first reach a vetted group of cyber defenders during the announcement week, with access expanding gradually to trusted partners through the Fairwind Program. For everyone else, Google says the model is "coming soon to developers, businesses, and consumers," but gave no date, price or general release timeline.

How does Gemini 4 Carbon compare to GPT-6 Astra and Claude Opus 5.5?

In Google's published comparisons, Carbon leads on several security tests, posting 88.2% on SEC-bench Pro versus 85.4% for GPT-6 Astra and 54.8% on ExploitGym versus 42.4% for GPT-6 Astra and 34.5% for Opus 5.5. GPT-6 Astra leads on FrontierSWE at 65.5% versus Carbon's 61.2%. Most rival numbers are self-reported by each provider.

Sources

Read Also