Phonic 1 vs Gemini 3.8 Live: One Board, Four Numbers, and the Question of Which Gemini You Mean

The first thing to settle in a Phonic 1 versus Gemini 3.8 Live comparison is that "Gemini 3.8 Live" is not one model. Google shipped two on September 15, 2026 — gemini-3.8-live, tuned for cost and latency, and gemini-3.8-live-extended-thinking, which reasons while it talks and runs tools in the background — and they behave like different products. On Artificial Analysis's speech-to-speech board the standard model scores 74.5 and the extended-thinking variant scores 82.6. Phonic 1, launched in late October 2026, scores 87.1. Which Gemini you line up against Phonic changes the margin from 12.6 points to 4.5, and changes which half of the argument you are having.

So the useful question is not "who wins". It is "which of these four measured things do you actually need", because the composite index hides a split decision underneath it. Phonic 1 wins the component that governs whether a call gets something done. Google's models win the components that govern whether the call feels good and whether it completes, and Google charges less per hour while doing it. A reader who only sees 87.1 against 82.6 will pick the wrong model for their job about half the time.

Two Gemini models, and only one of them is on the board against Phonic

Gemini 3.8 Live replaced gemini-3.1-flash-live-preview and is generally available through the Gemini API and AI Studio, with Gemini Enterprise and Search Live still in private preview. The standard model optimises for exactly what a phone agent needs at scale: fast, cheap, natural. The extended-thinking variant spends reasoning before it answers, which buys accuracy on hard multi-step requests and costs latency. They are separately named, separately priced at the token level, and separately scored.

• Gemini 3.8 Live (standard) — speech-to-speech index 74.5, reasoning 92, dynamics 96.1, agentic 30.1, arena preference 1,074, task success 93.2%, time to first audio 1.18s.

• Gemini 3.8 Live Extended Thinking (High) — index 82.6, reasoning 98, dynamics 91.9, agentic 68.6, arena 966, task success 89.1%, time to first audio 1.35s.

• Phonic 1 (High) — index 87.1, reasoning 99, dynamics not reported, agentic 79.9, arena 1,012, task success 92.3%, time to first audio 0.61s.

Notice the internal contradiction inside Google's own line-up. Turning on extended thinking raises the index by 8.1 points, but it lowers arena preference by 108 Elo and task success by 4.1 points. Reasoning harder makes the model score better on the composite and worse on two of the four inputs. That is not a flaw in the models; it is a reminder that a single weighted average can move in a direction no user would vote for.

The component arithmetic, one line at a time

Artificial Analysis builds its speech-to-speech index from four inputs — speech reasoning, agentic performance, arena preference and task success rate — and labels it a weighted average. Read the four separately and the matchup resolves into something far more specific than a 4.5-point gap.

• Speech reasoning — Phonic 1 99 vs Gemini 3.8 Live Extended Thinking 98 and standard 92. Effectively a tie at the top; this component is not deciding anything.

• Agentic performance — Phonic 1 79.9 vs 68.6 extended thinking and 30.1 standard. This is the whole argument. Extended thinking is the only Gemini configuration that competes here, and it still trails by 11.3 points.

• Arena preference — Phonic 1 1,012 vs 966 extended thinking and 1,074 standard. The standard model beats Phonic by 62 Elo; the thinking model loses by 46. Human listeners prefer the cheap Gemini.

• Task success rate — Phonic 1 92.3% vs 89.1% extended thinking and 93.2% standard. Again the standard model wins, narrowly, and the thinking model loses.

Set those four in a row and the story is coherent. Phonic 1 is the most capable agent and the least preferred voice among the three. Gemini 3.8 Live standard is the most preferred voice and the best task completer, and the worst agent. Gemini 3.8 Live Extended Thinking sits between them and is arguably the worst of both — it pays for reasoning with latency and preference and still does not catch Phonic on agency.

What a minute of each costs, and why the headline rates mislead

On Artificial Analysis's normalised cost-per-hour-of-input-audio measure, Gemini is the cheaper model: $3.50 an hour for Gemini 3.8 Live Extended Thinking against $4.20 for Phonic 1. That is the number Phonic's own launch material concedes — its comparison table shows Gemini undercutting it — and it is worth stating plainly, because a newcomer beating a Google model on agentic score and losing to it on price is a more interesting story than a clean sweep.

Google's published token rates are $3 per million audio input tokens and $12 per million audio output tokens, which works out to roughly $0.005 a minute for input and $0.018 a minute for output at typical densities. Phonic's billing is one flat $0.07 a conversation minute, $0.065 on the Pro tier with a $200 monthly minimum, and the vendor states explicitly that there is no LLM pass-through and no per-token charge.

• Cost per input hour — Gemini 3.8 Live Extended Thinking $3.50 vs Phonic 1 $4.20, per Artificial Analysis's normalisation.

• Billing shape — Gemini bills separately for audio in and audio out at token rates; Phonic bills one rate per conversation minute covering the model layer.

• Where the shapes diverge — a call where the agent talks a lot and listens little costs more on Gemini's output meter and the same on Phonic's flat rate; a call where the agent mostly listens costs less on Gemini and the same on Phonic.

• Free tier — Gemini offers free tokens on both input and output with unpublished free-tier rate limits for the Live models; Phonic has no free tier but ran a $250 credit promotion for sign-ups before November 30, 2026.

This is exactly the sort of decision a router is built for, with one honest caveat: neither of these two conversational models is on OrcaRouter's catalogue, so they cannot be compared or failed over here. What is on the catalogue is the text-side sibling — Gemini 3.8 Flash, routable today alongside 200+ other models on a single key at list price with no markup. If your voice agent delegates reasoning to a text model rather than doing it inline, that leg is the one you can route, price-compare and fail over, and it is usually the leg that dominates the bill.

Latency, sessions, and the caps nobody reads

Phonic 1 responds in 0.61s to first audio. Gemini 3.8 Live takes 1.18s and the extended-thinking variant 1.35s. On the index's own speed chart that puts Phonic in the same band as Grok Voice 3.0 at 0.81s, with both Gemini configurations roughly a third of a second behind — the difference between a caller waiting and a caller noticing they are waiting.

Google has published hard session limits that matter more than the latency gap for some products: Gemini 3.8 Live sessions cap at fifteen minutes of audio-only interaction, or two minutes once video is in the stream. For a support call that is fine. For a scheduled intake interview, a coaching session or a long-running voice assistant, a fifteen-minute ceiling is a design constraint you have to build around, and one Phonic does not state. Phonic's public material also lists 50+ languages against Google's 97, which is the reverse of the usual newcomer deficit.

• Context and output — Gemini 3.8 Live carries 131K context with 65,536 max output tokens; Phonic does not publish a context figure.

• Multimodality — Gemini 3.8 Live accepts visual input at up to about one frame per second and ships 30 named voices; Phonic 1 is audio-native with no stated visual input.

• Provenance — Google applies SynthID watermarking to generated audio; Phonic makes no equivalent claim.

• Compliance — Phonic states SOC 2 Type II, HIPAA BAA availability and zero data retention, plus cloud, single-tenant, VPC and self-host deployment; Gemini 3.8 Live is a Google-hosted API.

Which decision each reader is actually making

If your agent books, verifies, refunds or escalates — if the success of the call is a database row that changed — Phonic 1's 79.9% agentic score is worth the higher hourly rate and the unproven track record. The 11.3-point lead over the best Gemini configuration is the largest genuine gap in this matchup, and it is the one that shows up as a completion-rate difference rather than a preference difference.

If your agent mostly needs to hold a good conversation, answer accurately and not sound like a robot doing it, the standard Gemini 3.8 Live is the better buy on three of four components and on cost. It wins task success outright, wins arena preference outright against Phonic, and costs 17% less per input hour. What it does not do is act — 30.1% agentic is a model that can talk about your booking system but not use it.

And if you are the team that would have reached for extended thinking by default because "thinking" sounds like more model: the numbers say don't. Gemini 3.8 Live Extended Thinking costs latency, arena preference and task success relative to its own cheaper sibling, and buys 38.5 points of agentic performance — still less than Phonic, and less than it needs to justify a call that has to complete a transaction.

What makes this matchup worth reading instead of skimming is that the winner changes depending on which of four numbers you weight. Phonic 1 takes the composite because agency is a third of the story and it leads it decisively. Google takes preference, task completion and price, and does it with a model that costs half as much to run. The index is a fair summary and a poor instruction manual. Read the components, then choose the one you can measure in your own traffic within a fortnight.