Images 2.5 trails Nano Banana, Multiverse cuts refusal to 5% on saturated XSTest

Sources

Introducing ChatGPT Images 2.5 simonwillison.net

Introducing ChatGPT Images 2.5 OpenAI’s image generation models are apparently used “more than 3 billion images across ChatGPT Images and the GPT‑Image models in the API”. This latest release improves their instruction-following ability across multiple turns, responds faster, and “is better at preserving the subjects in your reference photos”. There are two new model IDs in the API: gpt-image-2.5-sunburst and gpt-image-2.5-flare . Based on this I think Sunburst is the stronger option: Choose Su…

Safety for Whom? Refusing the Right Subset of a Topic, Not the Whole Topic huggingface.co

References

Hacker News discussion news.ycombinator.com

Flare and Sunburst are horrible names — no imagination or coherence, just a forced attempt to mirror Anthropic’s Haiku/Sonnet/Opus branding.

OrcaRouter — GPT-Image-2.5 vs Nano Banana 2 orcarouter.ai

Nano Banana 2 is roughly 2.4–2.8× faster than GPT-Image-2 and wins on photorealistic portraiture, while GPT-Image-2.5 is markedly better at legible multilingual text and poster/infographic layouts.

Kompozy hands-on review kompozy.io

Fixing a minor detail — like adding a single person — still triggers a cascade of unintended global changes, re-rendering backgrounds and shifting the established lighting; it’s the ‘1-step-forward, 3-steps-back’ problem.

WaveSpeed — GPT-Image-2.5: What We Know wavespeed.ai

The edits endpoint now accepts up to 16 reference images per request, with roles assignable in-prompt (e.g. image[0] as subject, image[1] as style), and adds native transparent-background output missing in v2.0.

GSMDome — Firefly integration & GPT-Image-1 sunset gsmdome.com

GPT-Image-1, gpt-image-1-mini and gpt-image-1.5 are scheduled for full removal on December 1, 2026 — barely six months after developers were pushed off DALL·E 3.

arXiv 2026 C2PA security analysis arxiv.org

Current C2PA specifications fail to meet several key security goals and may mislead users if relied upon for high-stakes verification such as legal evidence or journalism.

arXiv preprint (Multiverse Computing) — Boundary-Aware Self-Distillation arxiv.org

A civics tutor and a public-sector assistant may share the same base model while requiring distinct refusal boundaries… datasets [including politics and religion] and generation code will be released for safety research purposes only, as the harmful split contains persuasive and manipulative prompts.

ICML 2025 poster — OR-Bench: An Over-Refusal Benchmark for LLMs icml.cc

OR-Bench… 80,000 prompts across ten categories… Claude 3.5 Sonnet shows a high over-refusal rate of 43.8% on OR-Bench despite its reputation for safety, while models saturate XSTest at near-100%.

Röttger et al., XSTest paper (ResearchGate) researchgate.net

A test suite for identifying exaggerated safety behaviours in large language models — the same XSTest that Multiverse’s paper uses as its primary over-refusal yardstick.

Anomify 2025 political bias benchmarks vertexaisearch.cloud.google.com

Claude Opus 4.6 refused approximately 32% of questions when given a formal ‘prefer not to answer’ option, compared to zero refusals in forced-choice settings — refusals are strategic avoidance of subjective bias rather than hard-coded censorship.

Quantum Zeitgeist — Multiverse Computing funding coverage quantumzeitgeist.com

Multiverse Computing secures $215M to scale AI model compression technology… some technical observers argue the technique closely resembles classical low-rank decomposition or SVD and question whether the ‘quantum’ terminology is primarily marketing.

Sebastian Raschka — State of LLMs 2025 magazine.sebastianraschka.com

A growing concern among AI developers in 2025 is ‘benchmaxxing’—labs over-optimize models for public leaderboards at the expense of real-world utility… high scores on academic benchmarks like XSTest often fail to reflect performance in production environments.