Inference — Friday, May 8, 2026

Friday, May 8, 2026

New Releases

OpenAI Adds Three Voice Models to Realtime API

thenextweb.com · May 8

OpenAI shipped GPT-Realtime-2, GPT-Realtime-Translate, and GPT-Realtime-Whisper to its Realtime API, bringing GPT-5-class reasoning and live multilingual audio to developer-built voice apps.

Bottom line

Voice workflows that previously required assembling three separate vendor APIs — transcription, translation, reasoning — can now run through a single OpenAI endpoint.

OpenAI Releases GPT-5.5-Cyber for Security Defenders

helpnetsecurity.com · May 8

OpenAI launched GPT-5.5-Cyber in limited preview, a model tuned with more permissive safety thresholds to support offensive-security research and defensive workflows for vetted infrastructure teams.

Bottom line

Security teams that have been jailbreaking general models to get useful threat-research output now have a sanctioned alternative with no policy workarounds required.

Claude Reaches GA Inside Microsoft 365 Suite

finance.biggo.com · May 8

Anthropic moved Claude for Microsoft 365 to general availability in Excel, Word, and PowerPoint, with Outlook remaining in beta, positioning Claude as a direct competitor to Microsoft's own Copilot inside Copilot's host platform.

Bottom line

Microsoft now distributes a rival AI inside its own productivity suite, a structural tension that will define enterprise AI procurement conversations for the rest of 2026.

GitHub Copilot Sets June 1 Deadline to Drop GPT-4.1

github.blog · May 8

GitHub confirmed GPT-4.1 will be removed from all Copilot experiences on June 1, 2026, forcing enterprise admins to update model policies before the cutoff or lose access for users who have not migrated to GPT-5.5.

Bottom line

Any enterprise running automated Copilot integrations against GPT-4.1 endpoints has less than four weeks to test and migrate before those workflows silently break.

Anthropic's Natural Language Autoencoders Decode Claude's Internal Reasoning

marktechpost.com · May 8

Anthropic published a method called Natural Language Autoencoders that converts Claude's internal model activations into human-readable text, replacing the opaque numerical representations that have made AI auditing impractical for compliance teams.

Bottom line

For the first time, an enterprise deploying Claude could produce human-readable records of what the model was attending to during a decision — a prerequisite for AI governance frameworks in finance, healthcare, and public sector.

Funding

Nvidia Commits $3.4B to IREN for 5 GW AI Infrastructure Buildout May 8

Nvidia will spend up to $3.4 billion to use and expand IREN's data centers filled with Blackwell GPUs, a circular arrangement that reveals how Nvidia is now financing the infrastructure required to sell its own chips at scale.

HubSpot Posts 23% Revenue Growth as AI-Agent CRM Bet Splits Wall Street May 8

HubSpot reported $881M in Q1 revenue and raised full-year guidance, but shares slipped as CEO Yamini Rangan warned that its pivot to agentic pricing and AI-native sales cycles will extend deal timelines — a signal that AI-driven CRM transformation carries measurable near-term revenue risk.