AI developer tools · What shipped

Claude Haiku 5.5 has been evaluated on Artificial Analysis Intelligence Index v4.3.2, which incorporates 10 benchmarks including AA-Briefcase v1.1, AutomationBench-AA, Terminal-Bench 4.0, SciCode, and others measuring capabilities across finance, legal, healthcare, engineering, and coding domains. The model shows cost-efficiency positioning in the Intelligence Index versus cost-per-task analysis, with performance metrics tracked across output speed, latency, and domain-specific capability scores. A new AA-Personality Benchmark has also been introduced for personality index evaluation. [Source: artificialan]

Claude Fable 5.1 and GPT-6 Astra benchmark comparison shows Claude Fable 5.1 leading in coding (76.8 vs 75.5) and agentic tasks (80.8 vs 75.1) with Supported evidence on both, while GPT-6 Astra has a higher overall public score estimate (84.03 vs 83.25) with overlapping confidence intervals. Claude Fable 5.1 offers lower cost for cache-heavy agent loops ($0.75 vs $0.9), while GPT-6 Astra has a larger context window (1.05M tokens). On specific coding benchmarks, results are mixed: GPT-6 Astra leads on SWE-bench fixes (45 vs 43) and DeepSWE (65.5% vs 56.3%), while Claude Fable 5.1 leads on multimodal tasks (78.92% vs 70.67%). [Source: benchlm]

GPT-6 Astra outperforms Claude Opus 4.6 across multiple benchmarks with a public score of 84.03 versus 62.3. On coding tasks, GPT-6 Astra scores 75.5 compared to Claude Opus 4.6's 46.3, with both measured on BenchAlign v5.8 and Supported evidence. For agentic work, GPT-6 Astra achieves 75.1 versus 45.9. GPT-6 Astra also leads on math (97.3 vs 69.1) and reasoning benchmarks (93.3 vs 51.2 on ARC-AGI-3). Claude Opus 4.6 offers lower costs for basic chat turns ($0.0175 vs $0.035) and repository review ($0.325 vs $0.65), while GPT-6 Astra has lower cache-heavy agent loop costs ($0.9 vs $1.35) and a larger context window (1.05M vs 1M tokens). Shared benchmark results between models include BrowseComp (agentic), FrontierCode 1.1 Main (coding), GPQA variants (knowledge), and FrontierMath v2 tiers. [Source: benchlm]