A generated title card for Realtime 1 vs GPT Live 1 showing $0.01 against $0.05 per audio minute and a note that the 'feels very close' comparison is untested, plus an OrcaRouter logo.
Engineering & Research

Realtime 1 vs GPT Live 1: Five Cents Against One, and a Claim Nobody Can Check Yet

Author

jinhao.song

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

Five cents a minute. That is what O​penAI charges for G​PT Live 1 — the full-duplex voice model that "can listen and speak at the same time," billed by the second with no rounding up, backend model and tool usage charged separately. MAI Realtime 1, announced by Microsoft on 17 November 2026, is a cent a minute before its reasoning tier. On the rate card the difference is five to one, and one person who has used both told the internet that Realtime 1 "felt very close to the G​PT Live model." That sentence is the most interesting thing in this matchup and the least verifiable, because there is exactly one way to settle it and no independent party has run the test that would.

So treat this as two questions rather than one. Is a cent-a-minute model five times worse? And what does G​PT Live 1's measured profile actually buy that a cheaper model would have to match? The second has answers. The first does not, yet.

The architectures are not the same shape

G​PT Live 1 is unusual and its documentation says so plainly. It is a full-duplex model — listening and speaking simultaneously, rather than alternating turns — and it does not try to do the reasoning itself. It "hands off reasoning and tool work to a backend agent," reachable only through v1/live/sessions, with no Chat Completions, Responses or Realtime endpoint support. It takes audio and text in and out, supports streaming and function calling, and does not accept images or video.

• G​PT Live 1 — full-duplex front end, reasoning delegated to whatever backend agent you attach. Concurrent-session limits of 50 on Build, 300 on Launch, 500 on Grow. No free tier.

• MAI Realtime 1 — speech-to-speech with a fast path and an internal reasoning tier Microsoft bills at $1.00 per million input and $5.00 per million output tokens. Billed per audio minute on the conversational side. Preview unbilled until public preview.

Both are, in effect, two-part systems. The difference is who owns the second part. With G​PT Live 1 it is explicitly yours: you choose the backend agent, you pay its tokens separately, you see the seam. With MAI Realtime 1 the handoff is inside the product and the pricing page, and the routing decision is not something you configure.

That single design difference explains most of the price gap. O​penAI is charging $0.05 a minute for a very good ear and mouth. Microsoft is charging $0.01 for a decent ear and mouth plus an internal dispatcher, and monetising the thinking on the token meter.

What G​PT Live 1's measurements actually show

G​PT Live 1 is on Artificial Analysis' speech-to-speech board twice, because the backend agent you attach changes the result:

• Speech-to-speech index — 83.2 with an Astra-class backend at medium, 80.1 with a Sol-class backend at low. Both sit above G​emini 3.8 Live's 74.5 and well above Eleven Realtime Max's 77.3, which makes G​PT Live 1 the highest-indexed model in this comparison set.

• Speech reasoning (Big Bench Audio) — 90.1% and 89.0% respectively. Not the top of the board — Eleven Realtime Max scores 98.8% — but solidly competitive.

• Conversational dynamics — 94.9% and 97.3%. The Sol-configured variant is the best dynamics score in the set, which is the number that corresponds to "feels good to talk to."

• Agentic task completion (τ-Voice) — 74.5% and 59.3%. Both clear G​emini 3.8 Live's 30.1% by a wide margin, which is what you would expect from a model whose entire design premise is handing work to a capable backend.

• Arena Elo — 1027 and 1037.

• Task success rate — 87.4% and 90.9%.

• Time to first audio — 1.34 and 1.24 seconds. Slower than Eleven Realtime Max's 0.57 and slower than G​emini 3.8 Live's 1.18, but the full-duplex design means the perceived wait is not the same thing as the measured one; a model that acknowledges while it thinks does not feel like it paused.

• Blended cost — $5.83 and $4.47 per hour on Artificial Analysis' normalisation, against a $0.05-per-minute rate card that is $3.00 an hour for the voice line alone. The higher figures include the backend token spend, which is the honest way to read a model that delegates reasoning.

Two things follow from those numbers. First, configuration matters as much as the model: the same G​PT Live 1 scores 83.2 or 80.1 depending on which backend you point it at, and 74.5% or 59.3% on agentic tasks. A review that quotes one number without saying which backend it used has told you almost nothing. Second, the index leader here is not the arena leader — G​emini 3.8 Live wins Elo, G​PT Live 1 wins index. Those measure different things, and neither is a proxy for your workload.

A screenshot of Artificial Analysis's Speech to Speech AI Model & Provider Leaderboard page captured 18 November 2026, showing the page heading, the methodology summary covering reasoning quality, conversational dynamics, generation time and price, the Speech to Speech Index / Speed / Cost per Hour chart panels, and the Related Links rail; the individual model rows in the charts are not legible at this capture size.

The claim worth testing, and what testing it costs

"It felt very close to the G​PT Live model." Taken at face value, that is the strongest thing anyone has said about MAI Realtime 1, and it comes from someone with no published methodology, no recordings, and no task set. It is also, to be fair, exactly how most early impressions of voice models arrive — full-duplex conversation is hard to score and easy to feel, which is why the arena exists in the first place.

The trouble is that the claim is only half checkable. Microsoft publishes one accuracy figure for MAI Realtime 1 — 98.8% on Big Bench Audio, on the model page, vendor-run and unreproduced — alongside a P50 time-to-first-audio card and a "P50 Latency on BigBench Audio" caption. It publishes no τ-Voice score, no latency distribution, and no language list. Artificial Analysis does not carry MAI Realtime 1 as of 18 November, so there is no independent reading of that 98.8% to set beside G​PT Live 1's 90.1%: one is a vendor's own evaluation, the other is a third party's. Microsoft's own comparative claim names G​PT Realtime 2.1, a different model from G​PT Live 1 with a completely different price structure, and even that comparison is published as a frontier position rather than as numbers.

So the useful version of the claim is a test plan, not a verdict. Four things to run, all of them cheap while the preview is unbilled:

• Back-to-back conversation — same twenty recorded calls through both models, scored blind. This is the only way to test "feels close," and it is the test the original claim skipped.

• Interruption under overlap — G​PT Live 1's full-duplex design lets it speak and listen at once. Test whether Realtime 1 actually handles simultaneous speech or merely handles clean interruptions. Those are different capabilities and only one of them shows up in a demo.

• Escalation latency — the moment Realtime 1 hands a turn to its reasoning engine, measure the pause. Microsoft's claim is "without awkward pauses," which is a latency claim with no number attached.

• Cost per completed call — not per minute. A model that costs a fifth as much per minute and needs two more turns to resolve a request is not a fifth the price. This is the number that decides the matchup and neither vendor publishes it.

Where the five cents goes

If G​PT Live 1's extra four cents a minute buys you nothing, this is not a decision. But it buys something specific: a measured 74.5% agentic score against an unmeasured claim, and a full-duplex architecture that is a different capability from turn-taking rather than a better version of it.

What it costs you beyond the rate is control and plumbing. G​PT Live 1 requires you to supply and pay for a backend agent — which is real work and a real bill, but also the reason its configuration changes its scores. You choose the reasoning model; you see the token spend; you decide what escalates. MAI Realtime 1 makes that decision inside the product and charges $1.00 and $5.00 per million tokens for it, without telling you how often it happens.

That trade — visibility against convenience — is where a gateway earns its place. Both models end up calling out to reasoning models, G​PT Live 1 by design and MAI Realtime 1 internally, and that reasoning layer is the part OrcaRouter routes: 200+ models behind a single key at provider list price with 0% markup, so the backend that decides your G​PT Live 1 configuration is swappable without a second contract, and automatic failover so a stalled reasoning call does not turn into dead air mid-sentence. Model fusion is the other lever worth knowing about here — for a voice agent whose hard turns are the expensive ones, a panel answering together on the escalated path is a way to buy accuracy where it matters without paying for it on every turn.

To be exact about coverage: neither G​PT Live 1 nor MAI Realtime 1 is in the OrcaRouter catalogue. G​PT Live 1 is O​penAI's own v1/live/sessions endpoint; MAI Realtime 1 is MAI Labs and the Foundry limited-access preview. What sits behind them is what we carry.

The integration detail that decides your timeline

G​PT Live 1 has the more awkward integration and the better documentation. Its dedicated v1/live/sessions endpoint means nothing you have built against the Realtime API carries over — this is a new session model, and the fact that it does not support Chat Completions, Responses or Realtime is the first thing in its docs for a reason. Rate limits count concurrent sessions, not tokens, so capacity planning is a different exercise entirely: 50 concurrent on Build, 300 on Launch, 500 on Grow, no free tier.

MAI Realtime 1 is the opposite: an O​penAI Realtime API-compatible interface, which is the single most useful sentence in Microsoft's announcement. If your stack already speaks that protocol, evaluation is a model string and a base URL. The catch is everything around the interface — no language list, no latency distribution, no τ-Voice score, no GA date, and a two-minute demo in MAI Labs that was reporting "MAI Realtime is unavailable right now" when we checked.

So the timeline asymmetry is real. G​PT Live 1 you can integrate this week against published documentation and a production endpoint. MAI Realtime 1 you can prototype against compatibility you understand, and then wait on a preview that is free precisely because it is not finished.

A screenshot of OpenAI's GPT-Live-1 documentation captured 18 November 2026, showing the model page headline, the $0.05 per minute billed-per-second pricing, the v1/live/sessions endpoint, and the note that reasoning and tool work are handed off to a backend agent.

The verdict, stated as a bet

Pick G​PT Live 1 if the voice experience is the product and you want a backend you control. It has the highest speech-to-speech index in its class, the best conversational dynamics with a Sol-class backend, an agentic score that clears every non-OpenAI model in the comparison except Phonic 1 and G​rok Voice 3.0, and documentation that tells you exactly what it does. Five cents a minute is a real number and you can forecast it.

Bet on MAI Realtime 1 if the economics are the constraint and you have engineering time to spend on measurement rather than procurement. A cent a minute with an unbilled preview and an OpenAI-compatible interface is the cheapest possible way to find out whether Microsoft's two-tier design beats your application-side one. What you are accepting is that you will be the one who produces the latency distribution, the τ-Voice number and the language list, because Microsoft has published none of them — and that its one accuracy figure, 98.8% on Big Bench Audio, arrives from the vendor with nothing behind it.

The one thing not to do is repeat the "feels very close" claim as though it were a result. It might be true. The person who said it may well be right, and if they are, this matchup stops being interesting because the cheaper model would be the better one on the axis that matters most to users. But a feeling from an afternoon is not a task-completion rate, and the number that would settle it — what fraction of real calls each model actually finishes — is the one number neither vendor has published.

A generated two-column scoreboard titled 'MAI Realtime 1 vs GPT Live 1 — the scoreboard'. Left column MAI Realtime 1: voice price $0.01 plus tokens, speech index not published, agentic not published, dynamics not published, first audio P50 only. Right column GPT Live 1: $0.05 per minute, speech index 83.2 / 80.1, agentic 74.5% / 59.3%, dynamics 94.9% / 97.3%, first audio 1.34 s / 1.24 s. Footer reads 'GPT Live 1 per Artificial Analysis, two backends; MAI index and agentic scores unpublished.'