GPT-Live-1 vs Grok Voice Think Fast 2.0: Two Voice APIs Moving in Opposite Price Directions

The most useful fact about this matchup is a direction, not a number. When x​AI released Gr​ok Voice Think Fast 2.0 in late July 2026, it raised the per-minute audio rate 60%, from the $0.05 that Think Fast 1.0 charged to $0.08. Six weeks later, on September 10, 2026, Ope​nAI put GPT-Live-1 into the developer API at exactly the price x​AI had just abandoned. That gives you the rare situation where the two leading full-duplex voice APIs can be compared on price directly, and where the newer entrant is the cheaper one — while the older, more expensive model is the one with third-party benchmark scores behind it.

The ledger, before any benchmarks

Both of these are speech-to-speech models built for phone-shaped workloads: a live conversation with a customer, an interruption, a tool call, and a bill measured in minutes. Beyond that they diverge quickly.

• Price per minute of audio — GPT-Live-1 $0.05 vs Gr​ok Voice Think Fast 2.0 $0.08

• Billing unit — GPT-Live-1 bills by the second with no rounding up to the next whole minute; Gr​ok Voice Think Fast 2.0 is priced per minute of audio and works out to roughly $4.80 an hour

• Release — GPT-Live-1 shipped inside ChatGPT on July 8, 2026 and reached the API on September 10, 2026; Gr​ok Voice Think Fast 2.0 was announced around July 29–30, 2026, roughly three months after Think Fast 1.0

• Endpoints — GPT-Live-1 is Live Sessions only, at v1/live/sessions, over WebRTC; Gr​ok Voice Think Fast 2.0 runs on x​AI's Speech-to-Speech API over both WebSocket and WebRTC

• Tool surface — GPT-Live-1 delegates to a backend reasoning model through the Responses API; Gr​ok Voice Think Fast 2.0 has web search, X search, file search, MCP servers and client-side functions available inside the session

• Voice portability — GPT-Live-1 uses Ope​nAI's remastered set of twelve voices; Gr​ok Voice Think Fast 2.0 supports built-in and custom voices

The WebSocket line is smaller than it looks and matters more than it should. A large share of telephony infrastructure — the media-stream plumbing inside most CPaaS products — speaks WebSocket, not WebRTC. Gr​ok Voice Think Fast 2.0 meets that infrastructure where it is; GPT-Live-1 does not, and getting to WebRTC from a WebSocket-only call path is a real piece of engineering rather than a config flag.

The benchmark asymmetry is the actual story

Here is where the comparison stops being a wash, and where most write-ups get it backwards by treating both sets of numbers as the same kind of evidence.

Gr​ok Voice Think Fast 2.0's headline scores come from Artificial Analysis's Speech-to-Speech Quality Index, a third-party measurement that puts the model at 82.9%, up from 75.7% in version 1.0, ahead of GPT-Realtime-2.1 at 79.1% and Gemini 3.1 Flash at 69.5%. x​AI also reports first place on the τ-voice Bench for agentic behaviour at 56.5%, ahead of GPT-Realtime-2.1 at 45.7%. Those are externally run measurements of x​AI's model.

GPT-Live-1's headline scores are Ope​nAI's own. The company reports full-duplex interactivity of 80.1% against GPT-Realtime-2.1's 45.4%, turn-taking latency cut from 1.4 seconds to 0.8, and tool-calling accuracy up from 60% to 87%. Every one of those figures is vendor-reported, has not been reproduced by an independent lab, and compares Ope​nAI's new model against Ope​nAI's own previous model rather than against a competitor.

That is not a reason to dismiss the numbers — the direction of travel is consistent with what early deployers report — but it does mean the only cross-vendor comparison currently available favours x​AI's model on independently run tests, while the only available number for Ope​nAI's latency advantage is Ope​nAI's. Do not treat 0.8 seconds and 0.70 seconds as a real contest; only one of those two figures was measured by somebody with no stake in the result.

The alias trap, and why the price rise caught people out

The 60% increase would have been a rounding error in most release notes if x​AI had not also rolled it through an alias. Applications pointing at the grok-voice-latest alias inherited both the new model and the higher rate automatically on August 5, 2026, unless they had explicitly pinned grok-voice-think-fast-1.0. That is a design decision worth understanding rather than a complaint: floating aliases buy you free upgrades and they also move your unit economics without asking.

The obvious remedy is weaker than it looks. Gr​ok Voice Think Fast 1.0 — the model at $0.05 a minute, released April 23, 2026 — is now marked deprecated on x​AI's own model list. Pinning it protects you from the automatic upgrade, and it buys time rather than a permanent cheaper path, because a deprecated model has a retirement date somewhere ahead of it. Plan the migration on your own schedule instead of the alias's.

The pricing surface itself is worth reading carefully before you commit to a number. x​AI's models page lists the versioned rates explicitly — grok-voice-think-fast-2.0 at $0.08 per minute of audio, $4.80 an hour, plus $0.004 per text input — while the separate Voice API card on the same page advertises an agent price of "starting at $0.05 a minute", which is the entry point rather than the rate the 2.0 model bills at. If your capacity planning was done off the marketing number, it is roughly 60% light.

Ope​nAI's model does not have this problem in the same form, because there is nothing to float to. GPT-Live-1 has exactly one model ID, gpt-live-1, which is also its own default snapshot — there are no dated variants to pin, and therefore no alias that can silently change either the model or the price. The trade is that you also cannot freeze a known-good behaviour and ride it while something new beds in.

Both models bill the reasoning layer separately from the voice layer, and this is where the headline rate stops being the whole cost. GPT-Live-1's delegated Responses calls bill at the configured backend model's normal per-token rates. x​AI adds $0.004 per text input on the voice session, plus tool calls, a provisioned phone number at around $0.01 a minute where you need one, telephony and storage, on top of the per-minute audio rate. A voice agent that mostly listens and occasionally looks something up will track close to its headline rate; one that hits tools on every turn will not, and the two will not diverge in the same direction, because the delegated-backend design and the in-session tool design burn tokens in different places.

On our side, the reason that per-minute rate changes are worth watching closely is that OrcaRouter passes provider list price through with no markup. When a vendor moves a published rate, the number on our side moves the same day rather than at the next contract cycle.

What the delegation split costs you in engineering

If you are choosing between these two on architecture rather than on price, the deciding question is where the seam sits.

x​AI's model keeps the tools inside the session. That is simpler to reason about — one connection, one conversation, one place where a function call happens — and it means the model can decide mid-sentence that it needs to search and carry on talking while it waits.

Ope​nAI's model pushes that work out. The voice layer keeps the conversation alive and hands the task to a separate reasoning model through the Responses API, which means your tool definitions, your retrieval and your business logic live in a normal text-model integration rather than inside a session event stream. It is more moving parts, and it is also more portable: the delegated half is an ordinary API call that you can swap, price and fail over independently of the voice layer.

That matters for exactly one reason. Neither GPT-Live-1 nor Gr​ok Voice Think Fast 2.0 is served by anyone other than its own vendor — both are single-source, and a router cannot help you with the audio half. What a router can do is consolidate the half you can actually choose. GPT-5.6 Terra, the backend in Ope​nAI's own delegation example, is available through OrcaRouter at $2.00 per million input tokens and $12.00 per million output with both /v1/chat/completions and /v1/responses exposed, alongside roughly 190 other models on one key. If your voice agent calls a reasoning model per turn, that is the line item with routing leverage in it.

The verdict, split by workload

• Pick GPT-Live-1 if price per minute is the constraint, if your call path already speaks WebRTC, or if you want the reasoning brain behind the voice to be a swappable text model rather than a fixed part of the session.

• Pick Gr​ok Voice Think Fast 2.0 if you need independently verified quality on the table before you commit, if your telephony stack is WebSocket-native, or if in-session tools — X search in particular — are core to the product rather than incidental.

• Watch the alias if you are already on x​AI. Pinning a versioned model ID is the only thing standing between you and the next automatic rate change, and the same discipline is worth applying anywhere a floating alias appears.

The uncomfortable summary is that the cheaper model is the one without third-party proof behind it, and the better-documented model is the one that just got 60% more expensive. If you are early enough in the build that neither integration is locked in, the lowest-risk move is to prototype against both — the audio path is a few hundred lines either way — and let your own transcript quality decide, because the public evidence currently does not settle it.