Grok Voice Think Fast 2.0: xAI's Fastest, Priciest Voice Agent, Explained

xAI released Grok Voice Think Fast 2.0 on July 29, 2026 — its self-described "most capable speech-to-speech voice model" and the new flagship of its voice-agent line. It answers faster than any rival in the top five (0.70s to first audio), tops the agentic voice benchmark, and transcribes more accurately than the previous generation. It is also 60% more expensive per minute than the model it replaces, ships on a per-minute meter that quietly raises your bill, and comes with a version-alias switch that happens automatically in a week. This guide explains what Think Fast 2.0 actually is, what xAI changed under the hood, how the benchmark story splits into independent scores and vendor-reported claims, what it really costs with the arithmetic done, and how to call it — plus the caveats you'd be wise to read before routing a production call flow through it.

Accuracy note: the launch details, pricing, and the headline accuracy and business claims come from xAI's announcement and docs (2026-07-29). The Artificial Analysis composite scores are independent and third-party measured. xAI-reported claims — the Starlink A/B results, the transcription multipliers, the "60% fewer reasoning tokens," and the "nearly 5x faster" framing — have not published underlying data; treat them as directional vendor figures and run your own evals. Specs, prices, and the alias behavior can change; verify before you build.

TL;DR. Grok Voice Think Fast 2.0 is xAI's end-to-end speech-to-speech model for voice agents: audio in, audio out over a WebSocket, no separate ASR→LLM→TTS pipeline. It is genuinely fast (0.70s time-to-first-audio, the only top-five model under a second) and genuinely strong on agentic voice work, ranking first on Artificial Analysis's τ-Voice agent benchmark (56.5%) while sitting second overall on the speech-to-speech quality index (82.9%), behind Qwen Audio 3.0 Realtime Plus (84.1%). It reasons while it talks — cutting median reasoning-token use to about 40% of the old model — and it costs $0.08 per minute, up from $0.05. The catch is the meter: voice is billed per minute of wall-clock audio, so a 60% price rise on a model that is cheaper to serve lands directly on your invoice. And on August 5 the grok-voice-latest alias starts routing to 2.0 automatically, so teams on the old version need to pin it deliberately before then.

Key takeaways

• Native speech-to-speech: one model from microphone to speaker, no ASR→LLM→TTS chain — which is why first audio lands in 0.70s.

• Agentic leader: #1 on Artificial Analysis's τ-Voice agent benchmark (56.5%), well ahead of GPT-Realtime-2.1 (45.7%) and Gemini 3.1 Flash (37.7%).

• Not the overall leader: #2 on the speech-to-speech quality index (82.9%), behind Qwen Audio 3.0 Realtime Plus (84.1%); OpenAI still wins conversational dynamics (95.7% vs 95.1%).

• 60% pricier: $0.08/min (~$4.80/hour) vs $0.05/min for Think Fast 1.0 — despite the model reportedly using ~60% fewer reasoning tokens.

• Alias switch on August 5: grok-voice-latest routes to 2.0 automatically; pin grok-voice-think-fast-1.0 before then to stay on the cheaper version.

• Label every claim: the Artificial Analysis scores are independent; the Starlink A/B, transcription multipliers, and speed framing are xAI-reported and unpublished.

What Grok Voice Think Fast 2.0 is

Grok Voice is xAI's family of realtime, speech-to-speech models. "Think Fast" is the fast-response line, launched originally with the Voice Agent Builder in April 2026; Think Fast 2.0 is the second generation of that line, released July 29, 2026. The model is end-to-end: it consumes raw audio, reasons, and emits audio directly, rather than running a pipeline of a speech-recognition model feeding a text LLM feeding a text-to-speech model. That architecture choice is the root of its latency advantage — there are no serial hops and no intermediate text, so the model can begin a response while still listening.

xAI positions it explicitly for voice AI agents: customer-support lines, phone sales, reception, telephony, and other applications where a system talks to a person in real time and needs to actually get things done — book a call, qualify a lead, update a record, search the web — rather than just chat. The agentic emphasis, not raw transcription or synthesis, is the battleground the launch is aimed at.

What's new versus Think Fast 1.0

The upgrade from 1.0 is more than a numbers bump; xAI describes three structural changes:

• Parallel speech reasoning. The model reasons through a query while it is speaking, rather than thinking first and talking afterward. Practically, tool calls can fire before the agent has finished its first sentence — a big deal for latency-sensitive voice flows.

• Far fewer reasoning tokens. xAI reports median reasoning-token use dropped to about 0.4x the predecessor (roughly 60% fewer), which is part of why it says the model is cheaper to serve even though the price went up.

• Reinforcement-learned conversation style. RL reshaped how it talks: shorter sentences, one question at a time, no filler — a terser, more "human" call style that suits phone work where rambling burns minutes (and, on a per-minute meter, money).

On measurable speed, time-to-first-audio fell from 1.25s in 1.0 to 0.70s in 2.0. On transcription, xAI reports about a 1.4x improvement across 24 languages (vendor-reported, below).

The benchmarks, benchmark by benchmark

The launch is anchored on Artificial Analysis, an independent third-party benchmark house. That matters: unlike most of the other numbers in the announcement, these are measured by a neutral party, and they are the ones worth comparing like-for-like. The composite picture is better than a single headline.

Speech-to-Speech Quality Index: 82.9% (second place). This is the overall speech-to-speech composite, up from 75.7% for Think Fast 1.0. It beats GPT-Realtime-2.1 (79.1%) and Gemini 3.1 Flash (69.5%) — but it is not first. Qwen Audio 3.0 Realtime Plus leads at 84.1%. So "best overall voice model" is a stretch the data doesn't support; "fastest agentic voice model in the top tier" is accurate.

τ-Voice agent benchmark: 56.5% (first place). This is the metric xAI cares most about, and the ranking is real: 56.5% beats Think Fast 1.0 (52.1%), GPT-Realtime-2.1 (45.7%), and Gemini 3.1 Flash (37.7%). τ-Voice measures agentic voice performance — how well a model completes tasks over a voice channel, including tool use mid-conversation. On the narrow "can it do the job" axis, Grok leads. Musk promoted exactly this ranking.

Big Bench Audio: 97.2%. A speech-reasoning benchmark; strong, though Qwen posted a record 99.2%.

Full Duplex Bench: 95.1%. Conversational dynamics — interruption handling, turn-taking, barge-in. It is a big jump from 1.0's 77.8%, but OpenAI still edges it at 95.7%, and Qwen leads at 98.4%.

Time to first audio: 0.70s. The most consequential number for real voice products. It is down from 1.25s and is the only sub-second figure among the top five ranked models; independent testing broadly corroborates a sub-second response. Streaming TTS first-token latency is around 285ms. For phone and realtime use, latency is the metric users feel every single turn, and this is where 2.0 is decisively best.

Read across the rows, the profile is specific: fastest to speak, best at completing tasks, second-best overall, third on conversation finesse. That is an excellent voice-agent model, and a mediocre "best chatbot voice" claim. Choose it for the jobs the top row matches.

Transcription accuracy

Beyond conversation, xAI claims materially better transcription. Its internal evaluation across 24 languages and thousands of phrases reportedly shows about 1.4x the accuracy of Think Fast 1.0, and 1.5–2x the accuracy of dedicated transcription models like Deepgram Nova 3 and ElevenLabs Scribe v2. In noisy, real-world conditions — background noise, telephony compression, low-bandwidth calls — the reported accuracy gap over dedicated STT models widens to roughly 10x.

These are exactly the claims to treat with care. They are xAI-reported; the evaluation harness, the phrase sets, and the noise profiles are not published. An independent test of 2.0's noisy-phone transcription against Deepgram or ElevenLabs would be valuable and, as of this writing, hasn't been published. Directionally, end-to-end models absorbing more telephony in training is a plausible real improvement — but "10x" should be verified, not repeated as fact.

Pricing: $0.08 a minute and the 60% question

Here is where the launch gets contentious. Think Fast 2.0 costs $0.08 per minute of audio — about $4.80 per hour — plus $0.004 per text input message. Think Fast 1.0 was $0.05 per minute ($3.00/hour). That is a 60% increase, and it arrived the same month OpenAI cut prices on GPT-5.6 Luna by 80% and Terra by 20%, citing the same falling serving costs xAI claims for this model.

The meter is the whole story. Text models bill per token, so when a model gets more efficient, the customer's bill automatically shrinks. Voice models bill per minute of wall-clock audio, and the length of a conversation is set by the person speaking, not by the model. Efficiency gains in a per-minute product accrue to the vendor, not the buyer. That is why a model that reportedly uses ~60% fewer reasoning tokens — and is therefore cheaper for xAI to serve — costs 60% more to rent.

Do the arithmetic on a real call flow. A 4-minute call at 1.0 costs $0.20; the same call on 2.0 costs $0.32. Ten thousand such calls a month is 40,000 minutes: $2,000/month on 1.0 versus $3,200/month on 2.0 — a $1,200/month, or roughly $14,400/year, difference that comes from a routing change alone, not from call volume. Even a generous speed win barely dents it: shaving a full 10 seconds off every call saves about $0.013 while the price rise adds $0.12 — you recover roughly a ninth of the increase. Any business case that sells 2.0 as the cheaper model has confused "faster" with "cheaper."

For context, 2.0 is still roughly 2.2x cheaper than GPT-Realtime-2.1 High at about $10.75/hour. So it is not expensive in absolute terms — it is expensive relative to its own predecessor and to the direction every other model vendor was moving that month.

The August 5 migration trap

There is a deadline attached. On August 5, 2026, the grok-voice-latest alias will automatically begin routing to Think Fast 2.0. If you are live on Think Fast 1.0 through that alias, "doing nothing" upgrades you — and your bill — to the new version on that date. To stay on 1.0, you must explicitly pin the grok-voice-think-fast-1.0 model id before August 5.

The two defensible positions both require action before the deadline: either pin 1.0 deliberately because your scripted flows worked fine at $3.00/hour, or move to 2.0 deliberately and reforecast at $4.80/hour, justified on quality rather than savings. What's indefensible is discovering the change in a September invoice. A good rule: treat any alias ending in "latest" as a standing instruction to accept whatever the vendor ships next — including its price.

How to access and use it

Think Fast 2.0 is available through xAI's Speech-to-Speech API under the model id grok-voice-think-fast-2.0, with grok-voice-latest as the rolling alias. It is a realtime, full-duplex model over a WebSocket: wss://api.x.ai/v1/realtime?model=grok-voice-think-fast-2.0, authenticated with an Authorization: Bearer header on the handshake. For browser apps, mint ephemeral tokens server-side via the realtime sessions endpoint so the parent API key never reaches the client.

The API is OpenAI Realtime-compatible, so existing realtime-voice code generally needs only a base URL and key swap. Session setup happens through a session.update event: pick a preset voice (eve, ara, rex, sal, leo), set input and output audio format (PCM16 at 24kHz by default; μ-law for telephony), enable server-side voice activity detection, and register tools — including a built-in web search tool — so the agent can act mid-conversation. The event flow follows the standard pattern: the client appends audio buffer frames, the server streams response audio deltas, and tool-call argument events fire when the model decides to act, with results pushed back as function-call outputs. The model supports 25+ languages and custom voice cloning from roughly a minute of speech.

Note what it is not: it is a speech-to-speech agent model, not a general transcription or translation service. If your product's core job is converting audio to text cheaply, a dedicated STT model is a different tool — and on the per-minute meter you pay for the whole conversation, so transcription-only workloads get expensive fast.

How it fits a voice-agent stack

Think Fast 2.0 is a specialized realtime voice model accessed through xAI's dedicated API, not a general-purpose LLM that a model router hosts. The voice lane runs on xAI's WebSocket directly — that's where the sub-second latency lives, and it doesn't survive an intermediary hop.

Everything else around a voice product — the natural-language logic that isn't time-critical, the fallback reasoning when a conversation goes off-script, the summarization, the analytics, the email follow-up the agent triggers after a call — is general-purpose text and multimodal work. For that part, a vendor-neutral endpoint like OrcaRouter gives you one OpenAI-compatible API across many LLMs, so you're not locked into one vendor for the parts of the stack that don't need the realtime voice lane. The clean split: xAI's dedicated API for the streaming voice itself, OrcaRouter (or similar) for the text and multimodal reasoning around it.

The competition: agentic speed vs raw intelligence

Think Fast 2.0 lands in the middle of the most competitive market in AI right now — realtime voice agents — and the benchmark table makes the tradeoffs unusually visible.

Qwen Audio 3.0 Realtime Plus is the intelligence leader: 84.1% on the speech-to-speech quality index (first), a record 99.2% on Big Bench Audio, and 98.4% on full duplex. But its average time-to-first-audio is about 4.02 seconds — roughly 5.7x slower than Grok's 0.70s. For in-car, wearables, and phone conversations, that difference is not "fast vs slow," it's usable vs unusable. Qwen also splits into a Plus tier (quality) and a Flash tier (about 300ms first packet) for different scenarios, which is a thoughtful answer to the same tension.

GPT-Realtime-2.1 sits in the middle on most rows (79.1% quality, 45.7% agentic) and wins on conversational dynamics (95.7%). Its High tier is roughly 2.2x the per-hour price of Grok. For teams already deep in the OpenAI ecosystem, it remains the lowest-friction default.

Gemini 3.1 Flash trails on the quality and agentic composites (69.5%, 37.7%) but is part of a broader Gemini voice stack and Google ecosystem that many teams prefer for other reasons.

The practical read: pick by workload. If your voice agent must feel instant and must reliably complete tool-backed tasks (support, qualification, booking), Think Fast 2.0 is the strongest measured option today. If your priority is the deepest reasoning and you can absorb multi-second latency, Qwen Plus wins. If conversational polish and ecosystem fit matter most, GPT-Realtime-2.1 remains the incumbent to beat.

Three scenarios where it fits

1. Phone support and telephony

The combination that matters most: sub-second first audio, strong noisy-phone transcription (vendor-reported), and full-duplex interruption handling. A support line where the agent starts speaking almost immediately and can dig up account info mid-turn is exactly what 2.0 was tuned for. Its shorter, one-question-at-a-time RL speaking style is also well matched to telephony, where minutes are the billable unit — on either vendor's meter.

2. Lead qualification and post-call follow-up

Agentic voice is where 2.0 is genuinely first, and lead qualification is the canonical agentic voice task: qualify, route, and trigger follow-up while intent is fresh. Its tool calling can fire before the first sentence ends, so an agent can be creating a CRM record while still talking to the prospect. Plan for session-level attribution and strict tool permissions — a voice agent can err in real time, and the cleanup is on you.

3. Voice-agent products under active development

For builders shipping their own voice agents — the Tradecraft and GrokTerm integrations xAI cites are exactly this shape — 2.0's speed and OpenAI-compatible API make it a low-friction drop-in for GPT-Realtime code. Pin the version, run a narrow pilot with a clear success condition (e.g., "qualify and route pricing-page visitors"), and measure cost per completed outcome, not cost per minute.

Limitations and caveats

Think Fast 2.0 is five days old as this is written, and that shows in the caveats. The Starlink A/B results (higher conversion and support containment) are xAI-reported with no published figures; the transcription multipliers and the "nearly 5x faster" framing are likewise vendor claims, not independent benchmarks. The one article you'll see alleging "performance regression and fabricated test reports" refers to this same unverifiability — the numbers xAI published have not been independently reproduced, and the reliability-sensitive voice community is rightly suspicious of unpublished vendor metrics. Benchmarks also can't measure your worst call: a 0.70s response is worthless if the agent confidently routes a lead to the wrong team. Voice billing is per minute with a price increase already baked in, so the cost exposure is real. And as with any brand-new realtime model, expect API churn. Verify the accuracy claims on your own audio, pin versions in production, and put escalation and recording in place before the first live call.

FAQ

What is Grok Voice Think Fast 2.0?

xAI's flagship speech-to-speech model for voice agents, released July 29, 2026: native end-to-end audio-to-audio over a WebSocket, with a 0.70s time-to-first-audio and first place on the τ-Voice agentic benchmark.

How much does Grok Voice Think Fast 2.0 cost?

$0.08 per minute of audio (about $4.80/hour) plus $0.004 per text input message — a 60% increase over Think Fast 1.0's $0.05/minute.

Is Think Fast 2.0 the best voice model?

It's the fastest and best at agentic tasks (56.5% τ-Voice, first place) and second overall on the speech-to-speech quality index (82.9%), behind Qwen Audio 3.0 Realtime Plus (84.1%). "Best" depends on the workload.

What changed from Think Fast 1.0?

Parallel speech reasoning (it thinks while it talks), about 60% fewer reasoning tokens, RL-tuned shorter conversation style, a 1.4x transcription improvement (vendor-reported), and time-to-first-audio cut from 1.25s to 0.70s.

Will my Think Fast 1.0 traffic switch automatically?

Yes, on August 5, 2026, if you use the grok-voice-latest alias. Pin grok-voice-think-fast-1.0 before then to stay on 1.0, or take 2.0 deliberately and reforecast the cost.

How do I call Grok Voice Think Fast 2.0?

Through xAI's Speech-to-Speech API — a realtime WebSocket (wss://api.x.ai/v1/realtime?model=grok-voice-think-fast-2.0) compatible with the OpenAI Realtime API, with preset voices, server VAD, and tool calling.

Which voice models does it compete with?

Qwen Audio 3.0 Realtime Plus (more intelligent, far slower at 4.02s first audio), GPT-Realtime-2.1 (better conversational dynamics, ~2.2x pricier at High tier), and Gemini 3.1 Flash.

Are the benchmark claims reliable?

The Artificial Analysis scores are independent and solid. The Starlink A/B results, transcription multipliers, and speed framing are xAI-reported with no published data — treat them as directional and verify on your own audio.

Is Grok Voice good for transcription?

It transcribes within a conversation, and xAI claims large gains over dedicated STT models in noise — but it's billed per conversation minute, so dedicated transcription workloads are better served by a dedicated STT model.

Bottom line

Grok Voice Think Fast 2.0 is the strongest agentic voice model measured so far, and the fastest in the top tier by a wide margin — 0.70s to first audio is a genuine, independent, user-facing win, and first place on τ-Voice is a real ranking. The honest summary is specific: it's the best tool in its class for realtime, tool-backed voice agents, not the best voice model on every row — Qwen leads on intelligence and OpenAI on conversational finesse. The controversies are not about the speed. They're about the meter: a 60% price increase on a model xAI says is cheaper to serve, an automatic alias migration with a hard deadline, and headline claims that rest on unpublished vendor numbers. Use it for the agentic speed, pin your version, label the vendor claims, and verify the accuracy on your own calls — and keep the general-purpose text and reasoning around your voice product on a vendor-neutral endpoint like OrcaRouter.