Gemini 3.8 Live Extended Thinking vs Gemini 3.8 Live: Paying 4x for the 38 Points That Matter

Two models, one launch, one rate card, and a decision that most write-ups get wrong in the same direction. Gemini 3.8 Live Extended Thinking and Gemini 3.8 Live shipped together on September 15, 2026, both built on Gemini 3 Pro, both handling 97 languages with automatic mid-conversation switching, both accepting near real-time visual input, both watermarking generated audio with SynthID. On the composite Speech to Speech Index that Artificial Analysis publishes, they sit 6.6 points apart — 82.6 against 76.0. On the hourly audio cost that same board measures, they sit a little over four times apart.

Neither of those is the number that should decide your architecture. The number that should is 38: the gap between the two on τ-Voice, the agentic task-completion component, where the reasoning variant resolves 68.6% of replica customer-service scenarios and the standard variant resolves 30.1%. You are not choosing between a good model and a better one. You are choosing between a conversational interface and a reasoning system that happens to talk, and the price difference is the least interesting thing about that choice.

The price fork, in the units you actually get billed in

Start with the bill, because it is the part people get right and then over-weight. Both models carry the same published rate through the Live API: $0.005 per minute of audio input and $0.018 per minute of audio output, and on the Extended Thinking side those output tokens include the thinking. Same card, same rates, no separate premium tier for reasoning.

The divergence shows up once you measure work rather than minutes. Artificial Analysis's cost-per-hour-of-input-audio column — the cost of completing a fixed 40-question Big Bench Audio subset, normalised to an hourly figure — puts the two models in different brackets entirely:

• Gemini 3.8 Live Extended Thinking (High) — $3.50 per hour of input audio, on Artificial Analysis's own measurement

• Gemini 3.8 Live — $0.84 per hour, the lowest paid rate on that board

• GPT-Live-1 (Astra backend, medium effort) — $5.83 per hour, so the reasoning variant still undercuts the model it beats by roughly 40%

• Grok Voice Think Fast 2.0 High — $4.80 per hour

• GPT-Realtime-2.1 High — $10.75 per hour

That is the fork: four times the hourly audio cost, on the same published per-minute rate, because the reasoning variant spends more tokens doing the same amount of listening. The standard model's $0.84 is not a discount, it is the floor of the entire board.

What the 38 points buy

Pull the composite apart and the two models stop looking like a pair:

• Speech to Speech Index — Gemini 3.8 Live Extended Thinking (High) 82.6 vs Gemini 3.8 Live 76.0

• Agentic performance (τ-Voice task completion) — 68.6% vs 30.1%

• Speech reasoning (Big Bench Audio) — 98% vs 92%

• Conversational dynamics — 91.9% vs 96.1%

• Arena preference (Elo) — 990 vs 1083

• Task success rate — 89.1% vs 93.2%

• Time to first audio — 1.35s vs 1.18s

• Cost per hour of input audio — $3.50 vs $0.84

Read that honestly and the cheaper model wins four of the eight lines. It is faster to first audio, preferred by the arena, more likely to complete a shallow task, and rated better on conversational dynamics — the last of which is not a consolation prize, because conversational dynamics is what a caller notices. What it cannot do is finish a job that has steps. Thirty-point-one percent against 68.6% is the largest single gap anywhere in this release, and it lands on the one axis that separates an agent from an interface.

Two caveats on those rows. The arena preference figure inside the index is frozen at the point a model becomes eligible for publication, so it is a snapshot rather than a running Elo — read it as a point-in-time rating and not as the live Speech Agent Arena chart. And the top of the τ-Voice board is tight: the reasoning variant's 68.6% leads GPT-Live-1 at 67.9% (Astra backend, medium effort) by 0.7 points, with GPT-Live-1 (Sol, low effort) at 59.3% and Grok Voice Think Fast 2.0 High at 56.5% behind. The 0.7-point lead over GPT-Live-1 is within noise. The 38-point gap inside this pair is not.

The other 4x: what the reasoning tier costs you in code

There is a second fork in this release, and it does not appear on any leaderboard. The two models are not drop-in interchangeable, because the reasoning variant changes the contract your client has to honour.

On the standard model, Gemini 3.8 Live behaves the way a Live API client expects: background tool and API calls run while the model keeps talking, and turnComplete: true still marks the end of a turn. There is no thinking-level setting to pick, because there is nothing separate to configure — the model answers, and when it has answered, the turn is done.

On Gemini 3.8 Live Extended Thinking, three things change at once:

• Function calls are always asynchronous. A blocking tool declaration does not queue politely — it returns a hard error. Any tool schema you wrote for the standard model has to be re-declared as non-blocking.

• A new status field carries the real state. Clients have to read interaction_status instead of trusting turnComplete: the field reads IN_PROGRESS while the model is still reasoning or a tool is still running, and only settles to IDLE when the whole task is finished. A turn can end while the task has not.

• Thinking depth is a knob. The model exposes configurable reasoning levels — low, medium and high — and the 82.6 and 68.6 figures on the board are the (High) configuration. Dropping to low changes both the quality and the token bill, and nothing on the index tells you what low costs you in task completion.

That third point is the migration trap. Because Extended Thinking answers the same way on the wire as the standard model until a tool call is involved, a team can swap the model ID, see a working demo, and only discover the async contract in production when a booking tool returns an error instead of a result. Budget for the client refactor alongside the 4× audio rate; on a mature integration it is the larger of the two costs.

Which one your workload is actually asking for

The clean way to decide is to ask what the session is for, not how smart you want it to be.

Take Gemini 3.8 Live when the conversation is the deliverable. Answering, holding a thread, taking a message, qualifying a caller, being pleasant and quick and cheap. At $0.84 per hour of input audio it is the cheapest competent voice model anyone currently publishes, it is preferred by the arena, it is more reliable on shallow tasks, and it is 170 milliseconds faster to first audio — a difference a caller feels. The one thing to accept is the ceiling: 30.1% on multi-step task completion means it will not finish work that has steps, and no amount of prompt engineering moves that number.

Take Gemini 3.8 Live Extended Thinking when the session has a job. Booking, changing, cancelling, resolving an account issue, walking a caller through a multi-step process while they wait. That is the 38-point line, and it is worth $3.50 an hour — which, note, is still cheaper than both models it outscores on the composite. You also get the narration behaviour that makes the wait tolerable: this model reasons and speaks at the same time, so instead of silence while it looks something up, the caller hears "Let me check that…" and progress as it goes. That is the same problem full-duplex architectures solve by never stopping the audio; Google's answer is to keep talking over the work.

Two more rows worth weighing. Google also reports 35.1% for the Extended Thinking model on Sierra's τ³-Banking leaderboard, against 32.0% for GPT-Live-1 Astra — a vendor-reported figure with no independent reading behind it, and a sobering one in absolute terms, since 35.1% means the best model on that board fails roughly two of every three realistic banking task attempts. And the standard model's second place in the Speech Agent Arena, which Google cites in its own materials, is a real result on a different board measuring preference rather than task completion. Both statements hold. Together they say the cheap model is very good at being talked to and the expensive model is the only one of the two that can be given work.

Running both, without two integrations

The practical objection to all of this is that picking per workload means maintaining two paths. It does not have to. Both variants sit in front of the same class of backend — background tool execution and multi-step planning resolve into ordinary text-model calls — and that layer is where your cost variance lives and where switching is cheap.

OrcaRouter puts 190 models from a single key at provider list price with no markup, so a vendor price change is live on our side the same day rather than at the next contract renewal. For a stack that routes between a cheap conversational tier and an expensive reasoning tier, the two properties that matter are that the delegation target can be swapped on live traffic without touching the voice integration, and that automatic failover keeps a session alive when a backend errors or times out. To be precise about what we do and do not host: the Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking endpoints are not on our router — those come from Google's own API, in the Gemini API, the Live API and Google AI Studio. The text models these agents hand their work off to very often are, Gemini 3.8 Flash among them.

Where each model sits on the board

The capture below was taken on September 16, 2026, and the top of the Speech to Speech Index reads as follows:

• Gemini 3.8 Live Extended Thinking (High) — 82.6, first place

• GPT-Live-1 (Astra backend, medium effort) — 81.5

• Grok Voice Think Fast 2.0 High — 81.3

• GPT-Live-1 (Sol backend, low effort) — 80.1

• Gemini 3.8 Live — 76.0, fifth place

• GPT-Realtime-2.1 High — 73.9

• Gemini 3.1 Flash Live High — 71.5

One naming note while you read that list: the (High) suffix attached to this model in coverage is a reasoning-effort configuration label, not a separate release. It is the same convention the board uses for Grok Voice Think Fast 2.0 High and GPT-Realtime-2.1 High.

What is still uncommitted

One thing about this pair is easy to misread, so it is worth stating plainly: neither model is generally available in the contractual sense, even though both are priced and documented.

Developers get Gemini 3.8 Live in the Gemini API, the Live API and Google AI Studio under the ID gemini-3.8-live; enterprises get private preview in Gemini Enterprise, with Gemini Enterprise for Customer Experience listed as coming soon; consumers get it in Search Live. Gemini 3.8 Live Extended Thinking has the same developer surface under gemini-3.8-live-extended-thinking, plus a wider consumer path — Gemini Live, Docs Live for Google AI Pro and Ultra subscribers, and Gmail Live and Keep Live for all Google AI subscribers. Workspace business customers are listed as coming soon.

What is missing is what an enterprise needs before putting a revenue line on this: a general-availability commitment, a published uptime figure, and a rate Google has promised to hold. The prices are published, and preview pricing has a habit of moving. Prototype now, plan for the migration, and do not sign an availability target you have not been given.

The decision this pair actually presents is narrower than the index makes it look, and it is not a quality ladder. If your agent's job is to talk, the cheap model is not a compromise — it is the better product at the lower price, and the four-times-cheaper rate is a bonus rather than the argument. If your agent's job is to finish something, the reasoning variant is the only one of the two that can be given the task at all, and $3.50 an hour is a rounding error against a booking that otherwise fails. The mistake to avoid is splitting the difference: paying for reasoning depth on a workload that only ever needed a good conversation, which is what happens when a team reads the 6.6-point composite gap and never scrolls down to the 38.