Google dropped two versions in a single blog post yesterday: Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking. One targets real-time conversation, the other targets deep reasoning.
On the surface it looks like the usual "one model, two modes" play. Reading the product page and the Hacker News thread more closely, it's more complicated than that.
The 3.8 version jump is unusual
Gemini has moved from 3.0 to 3.5 along a fairly linear path, so 3.8 is a noticeable skip. Google's blog doesn't explain the numbering. Some on Twitter speculate it's an internal sub-version pushed to production; others think it's a branch of the broader Gemini 3 family.
My read: Google is running an independent release cadence on its real-time voice and multimodal track, decoupled from the text model releases. The Live line (descended from Project Astra) has always iterated faster than the pure-text models, so an independent version number makes sense.
But skipping the explanation creates a perception problem. Developers who see 3.0 → 3.5 → 3.8 will default to assuming a step-change in capability, when in practice it might just be a two-point bump on some benchmark.
Standard Live vs Extended Thinking: not two products, but two API paths
The announcement presents them side by side, but they target different use cases:
- Gemini 3.8 Live: low-latency conversation, aimed at voice agents, real-time translation, video understanding
- Gemini 3.8 Live Extended Thinking: multi-step reasoning, selling the idea of "spend a few more seconds for a better answer," directly competing with Claude's extended thinking and o3's reasoning mode
Same underlying model, two routing paths. The user picks an API endpoint based on scenario.
This pattern isn't new. OpenAI splits o3 and GPT-4o across two separate models. Anthropic uses a single Claude with a thinking toggle. Google is now bundling both under the "Live" product line, which muddies the semantics. A developer who sees "Live" first will instinctively think low latency — not deep reasoning.
The real question: can Extended Thinking actually work in a Live context?
This is the contrarian point worth examining.
"Live" usually means sub-second response. Extended Thinking's value proposition is "think longer for a better answer." Bundling them in one release leaves Google silent on the latency budget: does the API switch by scenario, or can a single request allow 30 seconds of thinking plus 5 seconds of output?
That distinction determines whether the thing is usable in a voice agent.
Concrete scenario: a user tells a voice agent, "Book me an 8 AM car to the airport tomorrow. Cheaper than the last one." That requires Extended Thinking (price lookup, comparison), but every second the user waits is dead silence in the conversation — the experience falls apart.
If Extended Thinking can be triggered selectively — say, the user says "think about this carefully," and the model switches into deep mode while playing a "thinking..." audio cue — that's a genuinely useful product shape. Otherwise it's just Google piling up SKUs that developers can't choose between.
What HN is debating at 110 points / 62 comments
The original HN thread is 49715947 and made the front page. The comments split into a few camps:
The latency camp: One developer ran a simple back-to-back test comparing Gemini 3.8 Live with GPT-4o Realtime API under equivalent network conditions. Conclusion: Live's TTFB (time to first byte) in text mode is comparable to GPT-4o, but the audio stream chunk interval is slightly longer, about 80-120ms more. That gap is perceptible in conversation.
The pricing camp: Multiple people complained the pricing page hasn't been updated. The new model has been live for 24 hours, but the docs still point to the old per-character billing — nobody knows if it's token-based or audio-seconds-based. This complaint is real and familiar: Google's documentation update speed has been a sore point for a while.
The reasoning camp: Several people pushed on whether Extended Thinking's thought process can be interrupted. If the model is mid-think and the user changes their mind, can the request be canceled? The API docs don't say explicitly — there's only a best-effort cancel field.
All three classes of issues are the kind developers hit only after actually building with the product, not the kind that show up in press-release talking points about how strong the model is.
What this means for multimodal agents
If Extended Thinking can be triggered selectively within Live mode, this becomes a genuinely useful product shape.
A specific use case: a meeting agent. The user joins a video call, the agent listens in real time and generates summaries on screen. That part is Live (standard latency is fine). But when the user asks, "What was that Q3 number he mentioned? How does it compare to last quarter?" — that needs Extended Thinking, pulling from meeting history, retrieving and computing.
If the switch is one line of API code:
stream = client.live(model="gemini-3.8-live-extended-thinking", ...)
vs.
stream = client.live(model="gemini-3.8-live", ...)
developers will use it.
If instead it's the same endpoint with a thinking_mode: "deep" parameter, the backend routes the same way, and the latency behavior is identical — that's just marketing language.
Google didn't clarify the API shape for these two versions, possibly on purpose — leaving room for third-party interpretation before releasing details slowly. Or maybe it's genuinely undecided internally.
Editorial take
Google's dual-version strategy is essentially an admission: one model can't do both low-latency conversation and deep reasoning well at the same time. Not yet.
OpenAI splits o3 and GPT-4o across two models. Anthropic uses a thinking toggle on a single Claude. Google has now arrived at the same crossroads — but bundling both under the Live product line makes positioning murkier and raises the cost of picking wrong.
What's actually different this time: the Live line originally served only voice and video scenarios, as Astra's continuation. Adding Extended Thinking means Live is starting to absorb text reasoning work, and the boundary is blurring.
My prediction: within six months, Google will merge these into a single API with a thinking_mode parameter, matching Anthropic's approach. Releasing them separately is just a transitional state.
The open question: Extended Thinking's token pricing. If deep-thinking output costs 5-10x the standard tier, developers will be forced into a "should I or shouldn't I" optimization mindset instead of "leave it on" by default. That directly shapes product decisions.