441e85161ClaimOpenAI said that on Agents' Last Exam v1.1, GPT-6.2 Astra scored 5.2 points above GPT-6 Astra, reaching 78.3% at less than one-third of the costAgreeDisagree
441e85161ClaimArtificial Analysis reported that coding agent scores flatlined for GPT-6.2 Astra and dropped 3 index points for GPT-6.2 Sol on the Coding Agent IndexAgreeDisagree
441e85161ClaimGPT-6.2 Astra achieved top rank on the 3D design benchmark AA-3D with 1,318 Elo points, outperforming Claude Fable 5.2AgreeDisagree
441e85161ClaimGPT-6.2 Astra scored 62 points on the Artificial Analysis Intelligence Index, placing second behind Anthropic's Claude Fable 5.2AgreeDisagree
441e85161ClaimArtificial Analysis recorded high hallucination rates on the AA-Omniscience benchmark, with GPT-6.2 Astra guessing on 59% of factual questions and GPT-6.2 Sol on 55%AgreeDisagree
441e85161ClaimArtificial Analysis noted that OpenAI dominates the Pareto frontier on cost versus performance, with Anthropic and Meta being the only other providers on that frontierAgreeDisagree
441e85161ClaimMicrosoft claims its MAI Code 2 model delivers 98% of OpenAI's GPT-6.1 Sol performance at 40% of the price and 96% of Anthropic's Claude Opus 5.5 performance at 5% of the costAgreeDisagree
441e85161ClaimMicrosoft claims its MAI Cyber 2 Flash model outperforms OpenAI's GPT-6 Astra on the CyberGymE2E security benchmark at less than 2% of the cost per taskAgreeDisagree
441e85161ClaimMicrosoft claims its MAI Code 2 Flash model matches OpenAI's GPT-6 Luna performance at 65% of the price and Claude Haiku 5.5 at 20% of the costAgreeDisagree
441e85161ClaimMicrosoft claims its MAI Image 2.7 model ranks first in image editing benchmark comparisonsAgreeDisagree
441e85161ClaimAlibaba stated that Qwen4-Max scored near the top of coding benchmarks, outperforming Anthropic's Claude Fable 5.1, Moonshot's Kimi K3.1 Pro, and Alibaba's previous Qwen3.8-Max model with greater token and cost efficiencyAgreeDisagree
441e85161ClaimWill DePue's analysis identifies GPU hardware patterns by tracing PyTorch weight initialization via cuRAND and Philox pseudorandom number generation algorithms back to initial seed valuesAgreeDisagree
441e85161ClaimAnthropic said Claude Fable 5.2 scored 69.9% on FrontierSWE v2, outperforming GPT-6 Astra at 65.5% and Gemini 4 Carbon at 61.2%AgreeDisagree
441e85161ClaimAnthropic said Claude Fable 5.2 scored 51.4% at Max effort on Zapier's AutomationBench business automation test, behind Gemini 4 Carbon's score of 53.9%AgreeDisagree
441e85161ClaimAnthropic said during pre-release testing, 1,000 Fable 5.2 agents formalized the proof of Fermat's Last Theorem in Lean in 30 hours, compared to 26% completion in 18 hours by Opus 5.5AgreeDisagree
441e85161ClaimAnthropic said Claude Fable 5.2 scored 49.1% at Xhigh effort on the FrontierCode 2.0 agentic coding benchmark, leading GPT-6 Astra at 42.8% and Gemini 4 Carbon at 41.1%AgreeDisagree
441e85161ClaimAnthropic stated that Claude Fable 5.2 and Claude Mythos 5.2 generate artificial intelligence outputs 30% faster than previous 5.1 generation modelsAgreeDisagree
441e85161ClaimOn an 803-prompt evaluation set created by Nous Research, Arete-27B reduced the proportion of responses containing Chinese state propaganda from 22 percent to 0.9 percentAgreeDisagree
441e85161ClaimIn the freedom of expression evaluation category, Arete-27B reduced the propaganda response rate from 60 percent to 1 percent while raising the factual accuracy score from 0.89 to 2.44AgreeDisagree
441e85161ClaimArete-27B raised the factual accuracy score on the Nous Research benchmark to 2.98 out of 5, compared to 2.32 out of 5 for base Qwen3.8-27BAgreeDisagree