AI Technology Observations | 31 October 2026, 07:00

This observation covers technical material published around the period from 07:41 on 30 October to 07:41 on 31 October 2026 in Australia/Sydney. It includes three verified changes.

GPT-Rosalind-6 opens to approved life sciences research customers

OpenAI began rolling out GPT-Rosalind-6 Research to a limited set of approved Research customers on 30 October. It is available in Work, Codex and the API under the model identifier gpt-6-rosalind, but not in Chat. Enterprise workspaces enter a two-week Admin Preview with the model off by default. Administrators can enable it only for organisations and authorised users with Rosalind Research approval. Discovery approval does not grant Research access.

Rosalind-6 uses the same weights as GPT-6 Astra. OpenAI added no training data and performed no additional post-training for this release; the changes lie mainly in behaviour boundaries for biological research and deployment controls. The model refuses fewer legitimate dual-use biology questions while continuing to refuse requests that would materially enable severe biological weaponisation. Under the Preparedness Framework, OpenAI classifies its Biological and Chemical capability as High and carries over Astra’s Critical Cybersecurity assessment. Applicants undergo reviews of their use case, governance, security controls and identity. Customers must apply least-privilege access, insider-risk controls and qualified human review. All capability and safety results come from the developer’s system card, with no public independent reproduction of this restricted version.

Direct sources: OpenAI Deployment Safety Hub, “GPT-Rosalind-6 System Card Addendum”, 30 October 2026, https://deploymentsafety.openai.com/gpt-rosalind-6; OpenAI Help Center, “Model Release Notes”, updated 30 October 2026, https://help.openai.com/en/articles/9624314-model-release-notes

GPT-6 Luna October 28 version enters use

OpenAI deployed an updated GPT-6 Luna on 29 October to provide more concise responses and reduce overly agreeable language. The change applies to Free and Go users and people using ChatGPT without signing in. The main GPT-6 model for Plus and Pro users is unchanged; only the model to which the system may switch after they reach GPT-6 message limits is affected. The system card calls the deployed model GPT-6 Luna (October 28), so the version date and the publication date of the supporting material are distinct.

OpenAI reports improved handling of gore in production-style safety tests, a Pareto improvement between safe handling of harmful requests and helpfulness on legitimate requests, and 99.91 per cent robustness on instruction-hierarchy evaluations conducted with automated red teaming. Cybersecurity and Biological and Chemical capabilities remain classified as High, and the safeguards stack remains the one documented for the 7 October version. The disclosed data and training types are also unchanged. These figures are developer evaluations without public independent reproduction. The confirmed change is a model replacement for the stated user groups, not a universal API upgrade.

Direct sources: OpenAI Deployment Safety Hub, “Addendum to GPT-6 Luna System Card: October 28, 2026 Update”, 29 October 2026, https://deploymentsafety.openai.com/gpt-6-luna-october-28-2026-addendum; OpenAI Help Center, “Model Release Notes”, updated 30 October 2026, https://help.openai.com/en/articles/9624314-model-release-notes

Pika H3 Flex releases generator adapters and local inference code

Pika released Pika H3 Flex on 29 October, a video-and-audio generation model post-trained on MiniMax H3. The public package includes two generator adapters, inference code and Hugging Face files for text-to-video, first- and last-frame conditioning and multimodal reference-to-video generation. Native output is 480p or 768p at 24 frames per second with 32 kHz stereo audio. The model uses six sampling steps, accepts up to 12 image, video or audio references, and supports local output durations from five to 15 seconds.

The open-weight component consists of Pika adapters for MiniMax H3 rather than a standalone base model. The adapters use the MiniMax H3 Community License Agreement, while the inference and loader code uses Apache-2.0. Pika tested one request on one H100 80GB GPU. Downloading both tasks requires about 210GB of base files plus 5.5GB of adapters, and the current package cannot shard or accelerate one request across multiple GPUs. Upsampling to 1080p, 2K or 4K, five-way audio-stem separation, prompt enhancement and accelerated serving remain Pika-hosted API features. Pika’s real-time speed and quality figures are developer tests; local speed depends on the hardware and implementation, and the public package has not received a complete independent reproduction.

Direct sources: Pika, “Introducing Pika H3 Flex”, 29 October 2026, https://pika.art/blog/pika-h3-flex; Pika Labs, “Pika-H3-Flex”, as at 30 October 2026, https://github.com/Pika-Labs/Pika-H3-Flex; Pika Labs, “Pika H3 Flex model card”, as at 30 October 2026, https://huggingface.co/Pika-Labs/Pika-H3-Flex

Sources

OpenAI Deployment Safety Hub, “GPT-Rosalind-6 System Card Addendum”, 30 October 2026, https://deploymentsafety.openai.com/gpt-rosalind-6

OpenAI Help Center, “Model Release Notes”, updated 30 October 2026, https://help.openai.com/en/articles/9624314-model-release-notes

OpenAI Deployment Safety Hub, “Addendum to GPT-6 Luna System Card: October 28, 2026 Update”, 29 October 2026, https://deploymentsafety.openai.com/gpt-6-luna-october-28-2026-addendum

Pika, “Introducing Pika H3 Flex”, 29 October 2026, https://pika.art/blog/pika-h3-flex

Pika Labs, “Pika-H3-Flex”, as at 30 October 2026, https://github.com/Pika-Labs/Pika-H3-Flex

Pika Labs, “Pika H3 Flex model card”, as at 30 October 2026, https://huggingface.co/Pika-Labs/Pika-H3-Flex