AI developer tools · What shipped

Anthropic released Claude Haiku 5.5 in October 2026, offering five model variants with different performance tiers. The top variant (Claude Haiku 5.5 Max) achieved an Intelligence Index score of 43, ranking 4th out of 17,443 models, with output speed of 240 tokens per second and cost of $0.12 per input token. The Intelligence Index v4.3.2 incorporates 10 evaluations including AA-Briefcase, GDPval-AA, AutomationBench-AA, Terminal-Bench 4.0, SciCode, Humanity's Last Exam, and others, measuring performance across specific capabilities including finance, legal, healthcare, and engineering domains.

GPT-6 Astra leads the HLE-Diamond web-and-code benchmark released September 22, 2026, with 82.9% accuracy on 1,000 expert questions using constrained web search, page fetching, and code execution at high reasoning effort. Claude Opus 5.5 places second at 73.9%, followed by Claude Fable 5.1 at 72.3%. Eight models total have published results on this frontier benchmark, which tests expert knowledge and reasoning with tool assistance; the benchmark is currently displayed for reference but excluded from overall model rankings.

Vals AI released CUA-bench, a computer-use benchmark where agents control video games using keyboard and mouse across six commercial titles including Minecraft, SUPERHOT, and eFootball. The October 13, 2026 snapshot shows GPT-6 Astra leading at 19.17%, followed by Claude Opus 5.5 at 14.00% and Claude Fable 5.1 at 13.17%. Nine models have been evaluated on the benchmark, which measures real-time keyboard-and-mouse computer use; the benchmark is currently display-only on BenchLM and excluded from weighted global rankings.

Tether AI Research released QVAC Genesis III, a 191.43-billion-token synthetic dataset designed to improve smaller AI models' performance on STEM problems. The dataset employs two novel training methods—Failure Analysis and Option-Level Reasoning—that teach models not just answers but explanations and reasoning verification. In benchmarks, models trained on Genesis III's Option-Level data outperformed open-source baselines, with a 1.7-billion-parameter model achieving 99.45% valid answers on STEM tests and showing gains of 28.57 percentage points on ARC-Easy and 21.35 on ARC-Challenge compared to equivalent Cosmopedia-v2 training across Qwen, Llama, SmolLM, and Gemma architectures. The research enables smaller models to run locally on devices with limited computing power, supporting applications in education, specialized technical assistance, and privacy-sensitive environments. The work will be presented at the Conference on Language Modeling (COLM) 2026.