AI briefing, 19 November: Gemini Neon access and an open speech model

Checked 19 November 2026. Two announcements from 18 November with practical implications for business users. Prepared with AI assistance from linked primary sources. No hands-on testing or independent benchmarking. Ordinary source links; no affiliate links.

Gemini 4 Neon: a paid option for document-heavy work

Google announced Gemini 4 Neon on 18 November. Rollout includes the Gemini API paid tier, Gemini for business, and the Gemini app for Google AI Pro and Ultra subscribers. Paid Pro and Ultra subscribers also get access through Google Antigravity. This is a rollout, so check the actual model selector before planning work around it.

The announced API rates are $3 per million input tokens, $0.15 for cached input and $15 per million output tokens, with a one-million-token context window. App subscription access and API billing are separate routes.

Tecrave analysis: a firm comparing supplier proposals could trial a fixed set of PDFs and spreadsheets, checking extracted figures and omissions against its existing process. Google’s claims of stronger analysis and lower fabrication are vendor claims, not Tecrave findings. A larger context window does not guarantee correct conclusions.

Original Google announcement, 18 November.

Rumble-2-Flash: multilingual transcription for teams with technical help

Reson8 released Rumble-2-Flash on 18 November: a 0.5-billion-parameter speech-recognition model supporting more than 40 languages. Downloadable weights carry an Apache 2.0 licence, and the documentation supplies Transformers and vLLM setup examples.

Tecrave analysis: a multilingual service business with a developer could assess it for transcribing authorised customer recordings in its own environment. Downloadable weights do not remove hosting, maintenance or review costs. This needs technical deployment rather than simply opening a meeting-notes app.

Reson8 reports strong speed and accuracy results, but these are vendor-reported evaluations. The model card identifies H200 hardware for its leaderboard run; do not treat those figures as a laptop-speed promise. Check names, amounts and accents on representative audio before relying on transcripts.

Original Reson8 release, 18 November and model card and licence.