AI News Daily 2026-10-16
Contents
- Anthropic has given six staff members of the nonprofit evaluator METR the same level of access to its hardware, offices, and internal tools that its own employees have, with narrow carve-outs for customer data and attorney-client privileged material.
- METR is designated a "stage three" evaluator, meaning it can publish its findings without needing Anthropic's sign-off first — a notable step for independent verification of safety claims.
- OpenAI has updated the model behind ChatGPT Voice to GPT-Live-1.1 (and a GPT-Live-1.1 Quip variant), adding new voices in several languages and improving accent naturalness.
01 Anthropic gives outside evaluator METR staff-level access to its systems
Published: 2026-10-15
Facts
Anthropic announced that it has granted six people from METR (Model Evaluation & Threat Research), a nonprofit organization that evaluates frontier AI systems, the same level of access to its hardware, offices, and internal tools that its own staff receive. The only carve-outs are for certain categories of information such as customer data and attorney-client privileged material. According to Anthropic, this puts METR's evaluators on a similar information footing to the company's own internal alignment and risk personnel. Anthropic also states that METR holds "stage three" evaluator status, meaning it is permitted to publish its findings without first obtaining Anthropic's approval.
Background
METR is a nonprofit organization focused on evaluating frontier AI models, and it has previously worked with multiple AI developers on capability and safety assessments. The move to grant staff-level access goes beyond the more limited, time-boxed access that outside evaluators are typically given for individual model testing engagements, extending instead to the company's broader internal environment and tooling.
Implications
Granting an external evaluator this depth of access — combined with the freedom to publish without prior approval — is a concrete test of whether AI developers can support independent verification of their own safety claims rather than relying solely on self-reporting. For organizations and decision-makers who weigh AI safety assurances when choosing which models or vendors to deploy, this kind of externally verifiable access arrangement is the type of evidence that can make safety claims more credible, though its practical significance will depend on what METR actually finds and publishes going forward.
Source: Anthropic — An update on independent evaluators / CNBC
02 OpenAI rolls out GPT-Live-1.1 to ChatGPT Voice for more natural speech
Published: 2026-10-15
Facts
OpenAI has updated the underlying model powering ChatGPT Voice to GPT-Live-1.1, along with a GPT-Live-1.1 Quip variant. The update is intended to produce more natural-sounding accents across multiple languages and to expand the variety of short filler phrases the assistant uses to indicate it is still searching for information. OpenAI also added new voices for French, German, Hindi, Japanese, Korean, and Mexican Spanish, along with a new English-language voice named "Bubbie."
Background
ChatGPT Voice is OpenAI's spoken-conversation feature within ChatGPT, and the company has periodically updated its underlying voice model to improve naturalness and language coverage. This update was published in OpenAI's official changelog for ChatGPT release notes.
Implications
Because the changes focus specifically on accent naturalness and multilingual voice variety, they have a direct bearing on everyday users of ChatGPT Voice, including those who rely on it for multilingual customer support or day-to-day spoken interactions. The addition of new voices across six non-English languages in particular signals continued investment in making the voice experience feel native across a broader set of markets.
03 Editor's note
Today's two confirmed stories both point, from different angles, to the ongoing effort to make AI safety and product claims verifiable rather than merely asserted. Anthropic's decision to give METR staff-level access is part of a broader pattern of AI developers trying to demonstrate the substance of their safety practices to outside parties, rather than relying only on their own internal assessments. OpenAI's ChatGPT Voice update is a smaller, user-facing change, but it reflects the same competitive backdrop of continued, incremental investment in core assistant products even as the industry's attention is often drawn to larger model-training races.
It is also worth noting what did not make today's edition. A wide field of candidate stories — including reported large-scale in-house model pretraining at major providers, AI-investment overheating concerns, and regulatory scrutiny of leading AI developers — was reviewed but set aside because the available sourcing could not meet this report's independent-verification bar (for example, a single-outlet report without independent corroboration, or an official source outside the pre-approved domain list). As a result, this edition favors a short but well-sourced set of items over a padded one.