Gemini 3.8 Live Extended Thinking vs GPT-Live-1: A 1.1-Point Tie and Two Different Bills

On the composite score, this is a tie. Gemini 3.8 Live Extended Thinking sits at 82.6 on Artificial Analysis's Speech to Speech Index; GPT-Live-1, in its Astra backend at medium reasoning effort, sits at 81.5. On the agentic component underneath — τ-Voice task completion — the gap is 0.7 points, 68.6% against 67.9%. On the reasoning component it is wider and runs the other way: 97.7% on Big Bench Audio for the Gemini model, 90.1% for GPT-Live-1. Nothing in that spread survives a second benchmark run, and none of it is what you should decide on.

What separates these two is not quality. It is that they disagree about what a voice agent is, who does the thinking, and — the part that hits a budget line first — how the session gets billed. Gemini 3.8 Live Extended Thinking charges by the audio token and reasons in the same model that talks. GPT-Live-1 charges a flat rate for the session clock, whether or not anyone is speaking, and hands the actual thinking to a separate text model you select and pay for separately. Those are two different products wearing the same benchmark score, and the one that fits depends on a question the leaderboard never asks: what does an average call look like on your line?

The 1.1-point tie, and where it actually breaks

Start with the numbers, because the thinness of the headline is the point:

• Speech to Speech Index — Gemini 3.8 Live Extended Thinking (High) 82.6 vs GPT-Live-1 (Astra backend, medium effort) 81.5

• Agentic performance (τ-Voice task completion) — 68.6% vs 67.9%, with GPT-Live-1 at 59.3% when it runs the Sol backend at low effort

• Speech reasoning (Big Bench Audio) — 97.7% vs 90.1%, the one component where the gap is not noise

• Complex banking workflows (Sierra τ³-Banking, vendor-reported) — 35.1% vs 32.0%

• Time to first audio — 1.35s vs 1.24–1.34s through the API

• Measured cost per hour — $3.50 vs $5.83 for the Astra configuration, both on Artificial Analysis's input-audio measurement

The honest reading is that the two models are indistinguishable on the composite, distinguishable on speech reasoning where the Gemini model leads by nearly eight points, and distinguishable on cost where it leads by about 40%. Where GPT-Live-1 pulls ahead is in the parts of the experience a benchmark struggles to score: it is genuinely full-duplex, listening while it speaks, emitting backchannels, deciding several times a second whether to talk or yield. Gemini 3.8 Live Extended Thinking is turn-based. It solves the same underlying problem — a caller left in silence while the agent works — by narrating instead, reasoning and speaking at once with cues like "Let me check that…" while the task runs.

Both approaches keep the line alive. They feel different. Only one of them shows up on an index.

Two billing models, and why the rate card misleads

This is the section that decides most deployments, and it is the one the coverage skips.

Gemini 3.8 Live Extended Thinking is billed the way a token model is billed. Through the Live API the published rates are $3.00 per million audio input tokens — roughly $0.005 per minute of audio input — and $12.00 per million audio output tokens, roughly $0.018 per minute, with thinking tokens counted in that output. You pay for audio that moves. A caller who pauses, thinks, or gets put on hold costs you nothing while they are quiet.

GPT-Live-1 is billed the way a phone line is billed. It is a flat $0.05 per minute of session time, charged per second, regardless of who is speaking or whether anyone is. The reasoning backend is not included: the text model doing the thinking, and any tools it calls, bill separately on top. That is how a $0.05 headline becomes an all-in figure of roughly $4.47 per hour on the Sol backend and $5.83 per hour on Astra medium, measured by Artificial Analysis.

Put those side by side and the naive comparison — $0.018 against $0.05 per minute, a 2.8× gap — is wrong in both directions. The measured hourly figures put the real gap at $3.50 against $5.83, closer to 1.7×. But the session-clock model also changes the shape of your bill rather than just its level: on GPT-Live-1 your cost tracks call duration, so a line with long, quiet, low-content calls — a support queue, a verification step, an IVR handoff — pays the same as a dense one. On the Gemini side, cost tracks audio, so silence is free and verbosity is not. Run the arithmetic on your own average call length before you run it on a per-minute rate, because that is the number that decides which of these is cheaper for you, and it is not the same answer for a booking line and a triage line.

Where the thinking happens

The second structural difference is delegation, and it is the one that determines how much of this stack you can actually swap.

GPT-Live-1 is a front-end. It handles the duplex audio, and when a query needs real reasoning it hands the work to a separate backend model — GPT-5.5 and GPT-6 Astra are both documented as backends — with reasoning-effort selectors that let you choose how much of that backend you want to buy. This is why the board lists GPT-Live-1 twice at different scores: 81.5 on Astra at medium effort, 80.1 on Sol at low. That 1.4-point spread between its own two configurations is larger than the 1.1-point gap between the two models in this matchup, which tells you that on the OpenAI side the configuration you pick matters more than the vendor you pick.

Gemini 3.8 Live Extended Thinking reasons in the same model that speaks. There is no separate backend to select and no separate bill for one; the trade is that you tune depth with thinking levels — low, medium and high — on the same model, and the 82.6 and 68.6 figures are the (High) configuration. Background tool and API execution runs asynchronously on both sides, so neither model hangs up while it works.

One practical consequence: because GPT-Live-1's intelligence is a purchased text model behind a voice front-end, its quality ceiling moves when OpenAI ships a text model, not when it ships a voice model. Gemini 3.8 Live Extended Thinking's ceiling moves when Google retrains the audio model. If you are placing a two-year bet on a voice platform, that difference in upgrade cadence is worth more thought than 1.1 index points.

What switching costs, whichever way you go

Neither of these is a model-ID swap, and teams that treat it as one find out in production. Moving onto Gemini 3.8 Live Extended Thinking means accepting a different turn contract: function calls are always asynchronous, so a blocking tool declaration returns a hard error rather than queueing, and clients have to read the interaction_status field — IN_PROGRESS while reasoning or a tool is still running, IDLE only when the whole task completes — instead of trusting turnComplete to mean the work is done. Moving onto GPT-Live-1 means adopting its backend-selection plumbing and a second billing surface, and living with an API that only became generally available to developers on September 10, 2026, after debuting in ChatGPT on July 8 — so third-party tooling, SDKs and observability around it are months old, not years. Whichever direction you move, budget the integration work next to the token bill. On a mature voice stack the refactor is usually the larger of the two.

How to run a comparison like this without betting on it

The sensible way to settle a 1.1-point question is to run both on your own traffic, and the awkward part is that doing so normally means two integrations, two contracts and two sets of credentials. It does not have to mean two architectures. In both of these systems the voice front-end is a thin layer over ordinary text-model calls — for GPT-Live-1 that is explicit, and for the Gemini side background tool execution resolves the same way — and that text layer is where your cost variance lives and where switching is genuinely cheap.

OrcaRouter carries 190 models from a single key at provider list price with no markup, so an upstream price change is live on our side the same day rather than at the next contract renewal. For a voice stack 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 call alive when a backend errors or times out — which is what lets you trial an unproven configuration on real traffic without putting a production line behind it. To be precise about what we do and do not host: neither the Gemini 3.8 Live Extended Thinking endpoint nor the GPT-Live-1 endpoint is on our router — those come from Google's and OpenAI's own APIs. The text models they delegate to very often are, Gemini 3.8 Flash among them.

The top of the board, and how little it settles

The capture below was taken on September 16, 2026:

• 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

• GPT-Realtime-2.1 High — 73.9

• Gemini 3.1 Flash Live High — 71.5

Three things worth noticing. First, first and third place are separated by 1.3 points and first and fifth by 6.6, so the top of this board is a scrum, not a ranking. Second, the model in this matchup's title is not the fifth-place Gemini entry — that is the standard Gemini 3.8 Live, a different and much cheaper product that should not be confused with the reasoning variant being compared here. Third, there is a precedent for treating any single index figure as provisional: xAI reported 82.9 for Grok Voice Think Fast 2.0 in July, and that model reads 81.3 on this board. Vendor-reported index scores do not always survive contact with a live leaderboard. The 68.6% and 67.9% τ-Voice numbers now sit on a public board run under Artificial Analysis's own harness, so they are corroborated rather than merely claimed — but a 0.7-point difference between two models on the same harness is not a difference. Verify it on your traffic or ignore it.

One more figure worth pricing into the decision, because it is the least comfortable number in this comparison: Google reports 35.1% for Gemini 3.8 Live Extended Thinking on Sierra's τ³-Banking leaderboard, against 32.0% for GPT-Live-1 Astra. Those are vendor-reported, unreproduced, and they describe a world in which the leading voice agent on that board fails roughly two of every three realistic banking task attempts. If your agent has to complete regulated, multi-step work, neither of these models is finished — and a 3.1-point lead on that benchmark is not a reason to pick a vendor, it is a reason to keep a human in the loop.

So: if conversation naturalness is the product and your calls are short and dense, GPT-Live-1's full-duplex behaviour is a real advantage that the index cannot see, and the session-clock billing is tolerable at that call length. If calls are long, quiet, or variable, or if speech reasoning and per-hour cost dominate, Gemini 3.8 Live Extended Thinking is ahead on the components that matter and about 40% cheaper per hour while it does it — with the caveat that it is turn-based, still in preview for enterprise, and requires a client refactor before it will run your tools. What none of this justifies is picking either one on the strength of 1.1 index points. On the evidence available, the honest reason to choose between these two is the billing model and the architecture underneath it, and both of those are things you can measure on your own traffic this week.

On OrcaRouter, automatic failover keeps a call alive when a backend errors or times out.