TL;DR — As of 2026-09-16, Google launched Gemini 3.8 Live and 3.8 Live Extended Thinking in the Live API on 2026-09-15: native speech-to-speech, 97-language mid-call switch, async tool calls, and visual grounding ($0.005/min in, $0.018/min out). Extended Thinking leads Artificial Analysis Speech-to-Speech at 82.6. Pair Google's real-time voice intelligence with BibiGPT's long-context notes workflow to turn lectures, podcasts, and recordings into searchable chapter summaries.
Paste a podcast, lecture, or meeting URL — BibiGPT turns it into timestamped chapters, a full transcript, and interactive Q&A. Pair cutting-edge models with structured knowledge extraction.
As of 2026-09-16, Google put Gemini 3.8 Live and 3.8 Live Extended Thinking in the Live API on 2026-09-15: native speech-to-speech, 97-language hot-swap, async tool calls, visual context. Audio in $0.005/min, out $0.018/min. Extended Thinking leads Artificial Analysis Speech-to-Speech at 82.6. This is a live conversation product, not a summary engine.
Public numbers from Google's Live API announcement and developer docs — not a claim that BibiGPT runs this model.
Gemini 3.8 Live reasons over incoming and outgoing audio in one model. Callers can interrupt, switch language mid-turn, or keep talking while a tool runs. That is a live call, not a recap.
Google reports Gemini 3.8 Live Extended Thinking at 82.6 on Artificial Analysis Speech-to-Speech, ahead of GPT-Live-1-Astra. Audio input is $0.005/min and output $0.018/min. Those scores and list prices are Google's, restated here.
Live models understand and generate speech in 97 languages and can switch mid-conversation. Inputs include text, images, audio, and video. Same day, Gemini 3.5 Transcribe — a separate speech-to-text SKU, 85+ languages — landed in the Gemini API.
A smoother interruption does not timestamp last week's decision. Knowledge work still needs a file you can search next Tuesday.
Gemini 3.8 Live can keep speaking while a tool call finishes in the background. After the call, you still need chapters, a transcript, and a quote you can find again. That job is async.
Laxis State of Meetings 2026: knowledge workers sit through 21.7 meetings a week, about 30% of the work week. A cheaper live minute does not shrink that pile. A timestamped recap does.
Edison Infinite Dial 2026: US podcast listeners average 8 hours 24 minutes a week across 6.8 shows. The leftover episodes need a triage desk, not another live voice. A summary decides which hour is worth playing.
Headline shifts from Google's 2026-09-15 Gemini Live API launch.
One model reasons over incoming and outgoing audio together. Interruptions, language switches, and backchannels stay in the same session. Older stacks that transcribe, then think, then speak wait for a turn to finish.
Google reports Gemini 3.8 Live Extended Thinking at 82.6 on Artificial Analysis Speech-to-Speech, #1 overall, ahead of GPT-Live-1-Astra. Live itself placed second on Speech Agent Arena. Those scores are Google's evals, restated here, not a test we ran.
Audio input is $0.005 per minute and audio output $0.018 per minute on the Live API. GPT-Live-1's front-end layer is $0.05/min. List prices are Google's and OpenAI's, not a BibiGPT SKU.
Tool calls run in the background while audio keeps streaming. Inputs: text, images, audio, video. The models understand and generate speech in 97 languages and can switch mid-conversation. Developer docs list gemini-3.8-live as the low-latency default.
Same day, Gemini 3.5 Transcribe landed in the Gemini API: dedicated speech-to-text, 85+ languages, streaming WER 4.0% / non-streaming 2.6% (Artificial Analysis, cited by Google). After the session you still need chapters you can search next week. That is the BibiGPT path, not a Live integration.
Where a live voice layer helps — and where a notes workflow still does the work.
US listeners average 8 hours 24 minutes a week across 6.8 shows (Edison Infinite Dial 2026). A live voice API does not triage the backlog. Generate chapters first, then decide which hour is worth a full listen.
Laxis 2026: 21.7 meetings a week, about 30% of the work week. A cheaper live minute is useful on the call. Next Tuesday you still need the sentence that was decided. File the recording; search the transcript.
Paste the URL. Get chapters before you press play, jump to a timestamp, export Markdown. Gemini 3.8 Live does not have to be in the loop. The notes workflow already runs on files you own.
The notes workflow is the product. The blog owns the long comparison. This page owns the event.
Triage a long episode into chapters before you spend the commute hour.
Turn a finished recording into a searchable recap, not a live agent.
The OpenAI full-duplex voice API this release is measured against.
The 2026-09-16 comparison this landing sits next to — two live voice APIs, one notes gap.
Launch claims come from Google's announcement and independent coverage. Last updated 2026-09-16. Not an integration claim.
On 2026-09-15 Google put Gemini 3.8 Live and 3.8 Live Extended Thinking in the Live API: native speech-to-speech, async function calling, visual context. Audio input $0.005/min, audio output $0.018/min. Same day, Gemini 3.5 Transcribe landed in the Gemini API (85+ languages).
Google — Build real-time voice applications with Gemini 3.8 Live and 3.5 Transcribe (2026-09-15) ↗Gemini 3.8 Live Extended Thinking leads Artificial Analysis Speech-to-Speech Quality Index at 82.6, with 68.6% on τ-Voice, 35.1% on Sierra τ-Voice-banking, and 97.7% on Big Bench Audio. Gemini 3.8 Live placed second on Speech Agent Arena. Google's numbers, restated.
Google — Gemini 3.8 Live & Gemini 3.8 Live Extended Thinking (2026-09-15) ↗Developer card: model code gemini-3.8-live. Inputs text, images, audio, video. Outputs text and audio. Live API supported. Migrating from Gemini 3.1 Flash Live. Latest update September 2026.
Google AI for Developers — Gemini 3.8 Live model docs ↗Independent recap: 97-language understand-and-generate with mid-conversation switch; visual grounding; early verbal cues such as 'let me check that'. Extended Thinking 82.6 surpasses GPT-Live-1-Astra on Artificial Analysis Speech-to-Speech.
SiliconANGLE — Google's new speech model Gemini 3.8 Live (2026-09-15) ↗Gemini 3.5 Transcribe (2026-08-26, API 2026-09-15): streaming WER 4.0%, non-streaming WER 2.6% (Artificial Analysis, cited by Google). 85+ languages. File endpoint: speaker labels and word-level timestamps. Live endpoint: neither. As of 2026-08.
Google — Gemini 3.5 Transcribe launch (2026-08-26) ↗US podcast listeners average 8 hours 24 minutes a week and follow 6.8 shows. The leftover catalog needs a triage desk, not another live voice.
Edison Research — The Infinite Dial 2026 (as of 2026-03) ↗Knowledge workers attend 21.7 meetings a week, about 30% of the work week (Laxis State of Meetings 2026, as of 2026-01).
Laxis — State of Meetings 2026 ↗Gemini 3.8 Live is Google's native speech-to-speech model in the Live API, released on 2026-09-15. One multimodal model reasons over incoming and outgoing audio simultaneously, handling natural interruptions and mid-turn language switches without chaining separate transcription and synthesis steps.
Gemini 3.8 Live Extended Thinking leads Artificial Analysis Speech-to-Speech benchmarks at 82.6, ahead of GPT-Live-1-Astra. Live API audio pricing is $0.005 per minute for audio input and $0.018 per minute for audio output, supporting async tool execution and visual grounding across 97 languages.
Gemini 3.8 Live powers low-latency, real-time voice conversations, but does not archive meetings or generate searchable knowledge notes. In contrast, BibiGPT pairs with models like Gemini 3.8 Flash to transform hours of recorded video and audio into timestamped chapters, mind maps, and interactive QA.
Why people use BibiGPT to turn videos into text every day.
Trusted by 50,000+ users worldwide
“I paste a link and get clean captions in seconds — it saves me hours of retyping every single week.”
Maya R.
Content Creator · Repurposes short videos
“Exporting the transcript lets me review new words at my own pace instead of pausing the video constantly.”
Daniel K.
Language Learner · Studies with real videos
“Accurate, timestamped text I can quote directly. It has quietly become part of my daily workflow.”
Priya S.
Researcher · Cites public talks
FAQ'S
Ask us anything!
1 2026 student AI study workflow playbook: compare Microsoft Study and Learn Agent, Laxu AI, NotebookLM Education, and BibiGPT — plus a 5-step walkthrough turning lecture videos into Anki flashcards.
2 Best Bilibili AI video summary tool 2026? Compare 5 options, then paste a Bilibili link for a free AI recap, mind map, and transcript-style takeaways — no login needed.
3 Install a video-summary skill into DeepSeek Harness in one copy-paste — SKILL.md matches Claude Code. Or skip dsh: paste a Bilibili or YouTube link in the browser and get a timestamped summary.
Paste a podcast, lecture, or meeting recording into BibiGPT. You get chapters, a searchable transcript, and follow-up Q&A. Gemini 3.8 Live handles the interruption. The notes still need a timestamp.
