GPT-5

2026-10-28

MODELS · 1 SOURCE · WIRED AI

Author deploys Codex-built AI agents to manage Premier League fantasy team; not yet outperforming humans

The author built a harness with OpenAI Codex that coordinates dozens of AI agents (news readers, modelers, critics) to select and manage a Premier League fantasy soccer team; despite improved capabilities versus years ago, the system has not yet outscored the human average. The article cites research by Haifeng Xu and others showing frontier models (reported ranking: GPT-5, Grok 4, Claude Sonnet, Gemini, Llama) improving at prediction tasks and benchmarks like Prophet Arena, while other experiments such as Prediction Arena and ForecastBench show mixed returns and evidence of models improving over time.

5.0

2026-10-28

RESEARCH · 1 SOURCE · arXiv cs.AI

Counting the Bandages: On measuring the time that AI saves

New arXiv paper (2610.35942v1) argues for a 'ledger' view of AI time-savings that pairs measured time saved with counterfactual excess escaped defects, defines model-free quantities (time saved, excess escaped defects) and a break-even defect-cost decision number, presents a baseline checking model, analyzes identifiability in randomized studies, replicates METR's developer study (same 246 issues show a 19%–24% slowdown depending on estimand; 34% without imbalance adjustment; AI helped on 3%–82% of issues), reports that GDPval's review-every-output workflow caps GPT-5 speedup at 3.6x, and tests model predictions in a small pre-specified experiment where reduced checking (informed by optimistic drafter error rates) produced faster but worse outputs.

7.0

2026-10-26

RESEARCH · 1 SOURCE · arXiv cs.AI

Comparative study: FISER vs. CEMIR for pragmatic intent recognition

This arXiv study benchmarks FISER (structured social/embodied reasoning) against CEMIR (context-enriched multimodal intent recognition that synthesizes narratives and photorealistic images) across four LLM backbones, from Qwen2.5-Omni-3B to GPT-5, on the HandMeThat v2 benchmark. Using multiple-comparison-corrected statistics and ablations isolating narrative and visual components, the authors find no statistically significant performance difference between the two approaches; performance correlates mainly with backbone capability and problem difficulty.

6.0

2026-10-22

RESEARCH · 1 SOURCE · TechCrunch AI

Microsoft’s MAI-DxO outperformed generalist doctors on simulated complex cases, study in Science

Microsoft published a Science paper reporting that its AI orchestration system MAI-DxO correctly diagnosed 83% of complex New England Journal of Medicine cases in a simulation, compared with 36% for individual generalist physicians using reference materials; specialists or physician teams (who typically handle such cases) were not included in the comparison. The study tested multiple base models (best results used OpenAI’s GPT-5 and GPT-5.5), found DxO added most value on complex cases and weaker models, and Microsoft said it will train a Mayo Clinic-backed healthcare foundation model and run real‑world evaluations at Beth Israel Deaconess.

8.0

2026-10-19

RESEARCH · 1 SOURCE · arXiv cs.AI

Hierarchical feedback-guided LLM planning for multi-component cyber-physical systems (arXiv:2610.22441v1)

This arXiv v1 paper proposes a hierarchical feedback framework that prompts LLMs to produce a structured task plan plus signal temporal logic (STL) constraints, detects definite contradictions between them, and feeds diagnostic information from failed control synthesis back to the planner. The authors report experiments on seven heterogeneous multi-component CPS scenarios and state a reported success rate of 87.9% when using GPT-5 to generate executable plans and feasible controls that meet the original natural-language missions.

7.0

2026-10-14

RESEARCH · 1 SOURCE · WIRED AI

Thompson study: OpenAI GPT‑5 series and other models increasingly solve theoretical problems across sciences

A study by Neil Thompson with Gabriel Manso and Claudio Mayrink Verdun tested OpenAI models (o1, o3, GPT‑5, GPT‑5.1, GPT‑5.4, GPT‑5.5) on theory problems drawn from papers posted Dec 2025–May 2026; models were blocked from internet access and many ground‑truth solutions were published after the models' training cutoffs. The best models' success rate on problems in probability, statistics theory, economic theory, high‑energy theory, gravitation and cosmology, and quantum physics rose from roughly 2% in mid‑2024 to about 35% by mid‑2026.

8.0

2026-09-27

MODELS · 1 SOURCE · The Decoder

OpenAI says most research is focused on GPT‑7 and beyond

Boris Power, OpenAI's Head of Applied Research, told THE DECODER that roughly 80–90% of the company's research is aimed at future model generations such as GPT‑7, GPT‑8 and beyond, because those generations deliver the largest value; incremental within-generation updates are treated as short-term, tactical bets. Power also said user onboarding remains a major challenge — future models (GPT‑6 and later) should better surface capabilities and accept goals more naturally than current versions.

7.0

2026-09-18

RESEARCH · 1 SOURCE · WIRED AI

Experts argue AI-enabled bioweapons risk is real but not an imminent existential threat

The article surveys recent concerns—including an Anthropic report flagging efforts to use Claude for biological misuse and research showing AI-designed viral genomes—but presents several scientists who say that while AI can accelerate access to biological information, major technical, material, and human checks still make AI-driven creation and deployment of novel bioweapons unlikely today. It also notes calls from some AI leaders for regulations on synthetic DNA and frontier capabilities.

6.0