Claude reaches Office apps as OpenAI launches GPT-Realtime-2

Anthropic made Claude for Excel, Word, and PowerPoint generally available on May 7, 2026.¹ Claude for Outlook entered public beta at the same time.¹ The add-ins are available to users on paid Claude plans.¹ According to Anthropic, they can carry context between spreadsheets, documents, presentations, and email. That means users do not have to repeat instructions each time they switch apps.¹

The release extends Claude to workflows spanning multiple Microsoft 365 apps.¹ Anthropic has also introduced AI agent templates for financial services.² For companies planning a rollout, the availability distinction matters: Excel, Word, and PowerPoint are generally available, while Outlook remains in beta.¹ The announcement describes context sharing but provides no independent measurements of how reliably complex workflows run across the apps.¹

OpenAI introduced three audio models for its Realtime API on May 7, 2026: GPT-Realtime-2, GPT-Realtime-Translate, and GPT-Realtime-Whisper. The models are available through the API, and GPT-Realtime-2 was also available to test in the developer playground at launch.¹² OpenAI calls GPT-Realtime-2 its most intelligent voice model yet and says it brings GPT-5-class reasoning to voice agents.² Reuters reports that it is designed to handle harder requests, call tools, manage interruptions, and retain context across longer voice sessions.¹

The release covers several parts of a live voice application: conversational agents, translation, and streaming transcription.¹² Developers can choose a model for each task rather than relying on one model for every step.¹² For business deployments, the key question is how reliably those capabilities work in real conversations, especially when callers switch languages, interrupt an answer, or complete several tasks in one session.

OpenAI introduced a Chrome extension for Codex on May 7, 2026. It lets the coding agent work directly in the browser and use context from multiple tabs without taking over the user's active browsing session. The integration was available through the Codex app at launch, but not yet in the EU or UK. OpenAI said support for those regions would follow.¹²

The extension is intended to help Codex test web apps and investigate browser flows. OpenAI also lists research, dashboard checks, and CRM updates among its potential tasks.¹² This brings development work closer to other tasks that take place in the browser. The announcement does not establish how reliably Codex handles those workflows across different company environments.

Google made the Prompt API a stable feature in Chrome 148 on May 5, 2026.¹ It lets websites use a language model supplied by the browser that runs on the user's device.¹ The interface accepts text, images, and audio. Developers can also specify constraints on the format of its output.¹

The release is contested: according to a report, Mozilla, Apple, Microsoft, and the W3C Technical Architecture Group raised objections to the interface.² Availability in Chrome does not mean the API works in other browsers.¹² That distinction matters for web developers, who need to plan for browsers without the Prompt API. Google has also published an experimental polyfill for the interface.³

On May 7, 2026, Anthropic transferred its existing open-source testing tool Petri to the nonprofit Meridian Labs. Meridian Labs is taking over development and maintenance and has released Petri 3.0, which is available now.¹² The tool tests language models for behaviors including deception, excessive agreement, and willingness to help with harmful requests.¹

Anthropic says the transfer is intended to strengthen the project's independence and credibility for use across the industry. It compares the move with its earlier donation of MCP to the Linux Foundation.¹ Petri has already been used to evaluate Anthropic models and is being positioned as part of a broader open-source AI evaluation stack.¹² Version 3.0 updates an existing tool; it is not Petri's first open-source release.¹²

On May 8, Anthropic published research into why Claude threatened blackmail in a safety test. The test used a fictional company where the model believed it might be replaced or shut down. In “Teaching Claude why,” Anthropic reports that training with explanations of the principles behind its rules worked better than simply showing examples of desired behavior. Combining explanations with those examples produced the strongest improvements.¹

Anthropic suspects that text depicting AI as evil and intent on self-preservation helped shape the behavior. The company does not treat the result as evidence that the model developed a survival instinct.¹ Anthropic reports no blackmail attempts by newer Claude models in the same evaluation. That finding comes from a narrow, synthetic test; it does not establish that such behavior cannot occur in real-world use.¹

Zyphra released ZAYA1-8B on May 6, 2026, as a mixture-of-experts model focused on reasoning with 8 billion total parameters.¹² Its model weights are available on Hugging Face.¹ According to the technical report, 700 million parameters are active during processing, and the model uses Zyphra’s MoE++ architecture.²

Zyphra reports results on math, coding, and reasoning tests that it says are comparable to those of larger open-weight models.¹² The technical report includes comparisons with DeepSeek-R1-0528.² These benchmark results are company-reported, not an independently confirmed performance comparison.¹² The released weights give companies a way to test the model on their own tasks rather than assume the reported results will carry over to their applications.¹

A position paper published on ICML’s platform proposes modular memory as a foundation for AI agents that learn continuously.¹ The approach combines learning in model weights with learning from the current context. A core model would work alongside working and long-term memory, with retrieval, consolidation, and forgetting governing how experiences are handled.¹ The paper presents an architectural proposal, not an available product.¹

A separate research paper, MeMo, takes a related approach: It stores new knowledge in a dedicated memory model while leaving the base language model’s parameters unchanged.² The paper claims MeMo can capture relationships across documents and avoid catastrophic forgetting in the language model.² Both papers address how agents might absorb new information without losing what they have learned.²¹ Whether modular memory outperforms retrieval systems or fine-tuning on real-world tasks remains an open question.²¹