Phonic 1 vs GPT Live 1: The 1.1-Second Demo, the 5-Cent Sticker, and the Bill Nobody Measures

The most repeated fact about Phonic 1 is a demo comparison: in Phonic's own test, its agent responded 1.1 seconds after the caller stopped speaking, against 15.7 seconds for OpenAI's GPT Live 1. It is a spectacular number, it is vendor-run, and it is the least useful thing in this matchup. The most useful thing is almost the opposite — a five-cent per-minute sticker price on GPT Live 1 that looks like it beats Phonic's seven, right up until you add the second meter OpenAI attaches to it. Between those two facts sits the comparison that actually matters, and it has one clean answer: on Artificial Analysis's speech-to-speech board, Phonic 1 (High) scores 87.1 and GPT-Live-1 in its Astra-medium configuration scores 83.2, a 3.9-point gap concentrated almost entirely in one component.

Both models are full-duplex conversational systems that call tools mid-call, both launched into a market that had no serious tool-calling voice model eighteen months ago, and both split their reasoning away from the speech layer in some fashion. That last similarity is what makes the cost comparison so easy to get wrong, and it is where this piece spends most of its time.

The demo, taken apart

Phonic's 1.1-second versus 15.7-second demonstration is a real observation of a real behaviour, and it is also a comparison with two unknowns. The first is configuration: GPT-Live-1's latency depends heavily on which backend model it delegates to, and Artificial Analysis measures time to first audio at 1.34s for the Astra-medium configuration and 1.24s for Sol-low — neither of which is fifteen seconds. A 15.7-second figure implies either a backend under heavy reasoning load or a task whose tool chain is long. The second unknown is the task: the two numbers describe different conversations, and there is no published protocol behind the pair.

The structural claim underneath the demo is sound, though, and here it is worth separating from the number. Phonic 1 fires tool calls while the caller is still speaking; its vendor material says so explicitly and its API exposes strict synchronous tools declared inline over the session. GPT-Live-1 is full-duplex in the sense that it listens while it speaks, and it delegates tool work across an asynchronous boundary to a separate backend model. Both designs move work off the critical path, by different mechanisms. Which one wins on your traffic is a measurement you take, not a demo you watch.

GPT-Live-1 is a configuration, not a score

Every public number about GPT-Live-1 needs a qualifier attached, because the model's score moves with its backend. On the same board:

• GPT-Live-1 (Astra, medium) — speech-to-speech index 83.2, reasoning 90, dynamics 94.9, agentic 74.5, arena preference 1,027, task success 87.4%, time to first audio 1.34s.

• GPT-Live-1 (Sol, low) — index 80.1, reasoning 89, dynamics 97.3, agentic 59.3, arena 1,037, task success 90.9%, time to first audio 1.24s.

• 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.

Three things fall out of that. First, the lower-effort Sol configuration scores better on dynamics, task success and arena preference than the higher-effort Astra one, and worse on reasoning and agency — which is the same trade-off Google's extended-thinking Gemini exhibits, and it means "the GPT-Live-1 score" is a meaningless phrase without a configuration attached.

Second, the agentic gap between Phonic 1 and the best GPT-Live-1 configuration is 5.4 points. That is smaller than Phonic's launch material implies — the 8.3-point lead it claims is against Grok Voice 3.0, not OpenAI — and small enough that a different task mix could plausibly close it.

Third, GPT-Live-1 leads on arena preference in both configurations, by 15 and 25 Elo respectively, and leads Phonic on conversational dynamics by a number Phonic has not published. On the axis a caller would describe as "sounds like a person", OpenAI is ahead and the gap is not measured on both sides.

The five-cent sticker and the second meter

GPT-Live-1 is priced at $0.05 per voice minute, billed per second with no rounding up, and a fifteen-second session-initiation charge that is credited back. Phonic 1 is $0.07 per conversation minute, $0.065 on the Pro tier. On those two lines OpenAI is 29% cheaper and the case is closed.

It is not closed, because GPT-Live-1's per-minute fee covers the voice layer only. Backend reasoning, tool calls and web search bill separately at the backend model's standard token rates — OpenAI's own guidance suggests a small model for high-volume turns and a frontier model for complex reasoning. Artificial Analysis measured real-world totals of $4.47 to $5.83 per hour of input audio depending on reasoning configuration, against $4.20 for Phonic 1.

• Voice layer — GPT-Live-1 $0.05/min vs Phonic 1 $0.07/min, $0.065 Pro.

• What else bills — GPT-Live-1 adds backend model tokens, tool calls and web search; Phonic 1 states one flat per-minute rate with no LLM pass-through.

• Measured hourly input cost — GPT-Live-1 $4.47–$5.83 vs Phonic 1 $4.20, per Artificial Analysis.

• Cost predictability — GPT-Live-1's total depends on your routing of the backend; Phonic 1's does not.

That inversion is the entire cost argument, and it cuts the opposite way from the sticker price. A team that keeps its backend turns cheap can beat $4.20 an hour on GPT-Live-1 comfortably; a team that sends every turn to a frontier model will land above it. Phonic's flat rate is a fixed price for a variable workload, which is valuable or overpriced depending on the mix you actually run.

The backend leg is where OrcaRouter fits without pretending to be something it is not. Neither Phonic 1 nor GPT Live 1 is on our catalogue — each is available through its vendor's own API — but the model that answers GPT-Live-1's hard turns is an ordinary text model, and those are the 200+ models we route on one key at 0% markup over provider list price. Both backends OpenAI names, GPT-5.6 Luna and GPT-6 Astra, have model pages here. A routing rule can hold easy turns on the cheap backend and escalate only the ones that need it, with automatic failover if a provider wobbles mid-call — which is precisely the control that turns GPT-Live-1's variable bill into a bounded one.

What OpenAI has that Phonic does not

The maturity gap is not a marketing point, it is a production point. GPT-Live-1 has been in ChatGPT since July 8, 2026, in the API since September 10, and received an incremental update, GPT-Live-1.1, on October 15. It has run at a scale no week-old model has, and its transport story reflects that: Live Sessions over WebRTC, WebSocket or SIP, with telephony handled rather than left to the integrator.

Adoption is the other signal. Perplexity moved its voice mode onto GPT-Live-1 on October 28, 2026, and Yelp and Speak are named early partners. Those are not proofs of quality, but they are proofs that the model survives contact with consumer traffic.

• Transport — GPT-Live-1 offers WebRTC, WebSocket and SIP; Phonic 1 offers a WebSocket endpoint, with webhooks, a held-open socket or LiveKit as integration styles.

• Concurrency — GPT-Live-1 publishes tiered concurrent-session limits (roughly 25–50 at Tier 1, up to 500 at the top tier) and has no free tier; Phonic 1 publishes no session cap in its public material.

• Boundaries — GPT-Live-1 accepts no image or video input and offers no structured outputs on the voice layer; Phonic 1 is audio-native with tool calls as the structured surface.

• Deployment — Phonic 1 offers cloud, single-tenant, VPC peering and self-host with SOC 2 Type II and HIPAA BAA availability; GPT-Live-1 is an OpenAI-hosted API.

What Phonic 1 has that OpenAI does not

Speed, on the board that measures it identically: 0.61s to first audio against 1.34s and 1.24s. That is the one measured gap in this matchup that is large enough to be felt by a caller rather than inferred from a chart, and it is consistent with Phonic's claim of sub-600ms speech-in to speech-out.

Agency, more narrowly: 79.9% against 74.5% on the better OpenAI configuration. And one bill, which is less a capability than a procurement property — but for a company that has been running a spreadsheet with a voice line, a backend line and a tools line on it, collapsing three meters into one is worth something real.

• Time to first audio — Phonic 1 0.61s vs GPT-Live-1 1.34s (Astra, medium) and 1.24s (Sol, low), per Artificial Analysis.

• Agentic performance — Phonic 1 79.9 vs 74.5 (Astra, medium) and 59.3 (Sol, low).

• Arena preference — GPT-Live-1 1,027 and 1,037 vs Phonic 1 1,012.

• Task success — Phonic 1 92.3% vs 87.4% (Astra) and 90.9% (Sol).

OpenAI's own performance claims deserve their own labelling, since none of them have been independently reproduced: 80.1% on Full Duplex Bench v1.5 against 45.4% for GPT-Realtime-2.1, a turn-taking latency of 0.798s against 1.41s, and about 87% tool-calling success. Those are vendor figures. They are plausible and they are unverified, and the same sentence has to be written about Phonic's 15.7-second comparison.

The decision, stated as a bet

Choose GPT Live 1 if your calls are conversational more than transactional, if you want the transport and telephony handled, and if you are willing to own the backend routing that makes its bill competitive. The 5.4-point agentic deficit is real but not fatal for an agent that mostly answers questions and occasionally calls a lookup — and OpenAI's larger multi-configuration line-up, from minimal-effort voice to frontier backends, gives you more knobs than Phonic currently publishes.

Choose Phonic 1 if the call has to complete a transaction, if a third of a second to first audio changes your abandonment rate, and if a flat per-minute rate is worth more to you than the theoretical savings of a well-tuned backend. Accept that you are buying a week-old model with one independent board, no published session limits and no consumer-scale deployment history.

What would settle it is neither vendor publishing anything. It is a fortnight of your own calls, split between the two, with the completion rate for the same task measured side by side. Phonic's 5.4-point agentic lead is exactly the size of gap that a single task domain can reverse — which is an argument for measuring, not for picking the board leader.