Gemini 3.8 Live vs Grok Voice Think Fast 2.0: The 0.3-Point Gap and the 3.2× Bill

Both Gemini 3.8 Live and Grok Voice Think Fast 2.0 reason while they speak — or rather, both do so in the configuration that is actually competing for the top of the leaderboard, which for Gemini 3.8 Live means the Extended Thinking variant rather than the standard tier. Both have been described by their makers as first on Artificial Analysis's Speech to Speech Index. Only one of those claims is currently readable on the public board, and the gap between the two leaders is 0.3 points — barely larger than the margin that separates adjacent models elsewhere on the same chart, which is another way of saying it is not a real difference.

The cost difference between them is not 0.3 points. It is a factor of 3.2.

That asymmetry is the whole story, and it is the thing to understand before reading either vendor's announcement.

Two claims of first place

When xAI announced Grok Voice Think Fast 2.0 in late July 2026, the reported figures put it at 82.9% on the Artificial Analysis Speech-to-Speech Quality Index, ahead of GPT-Realtime-2.1 at 79.1% and Gemini 3.1 Flash at 69.5%. Several independent outlets repeated that framing and described the model as top-ranked worldwide on voice.

On the board captured on September 16, 2026, after Google's launch, the same model appears as Grok Voice Think Fast 2.0 High at 81.3. Not 82.9. And the top row is now Gemini 3.8 Live Extended Thinking at 82.6.

The board in full, in order:

• Gemini 3.8 Live Extended Thinking — 82.6

• GPT-Live-1 (Astra backend, medium effort) — 81.5

• Grok Voice Think Fast 2.0 High — 81.3

• GPT-Live-1 (Sol backend, low effort) — 80.1

• Gemini 3.8 Live — 76.0

• GPT-Realtime-2.1 High — 73.9

• Gemini 3.1 Flash Live High — 71.5

Two lessons sit in that list, and neither is about which model is better. The first is that vendor-reported index scores are not durable: xAI's 82.9 did not survive on the public board, which now shows 81.3. This is not an accusation of dishonesty — composite indices get revised, configurations differ, and a July number measured against a July field is not wrong, it is just old. But it does mean the only figure you should plan against is one you read yourself, dated. The second is that the 0.3 points separating first from third place here are not a meaningful quality difference. Anyone choosing between these two models on the strength of that number is choosing on noise.

Read the components, not the composite

Composites hide arguments. The Speech to Speech Index is a weighted average of speech reasoning, agentic performance, arena preference, and task success rate — and the two models in this matchup are not close on two of those four.

This is where Google's claim becomes genuinely interesting, because Grok's entire pitch was agentic performance:

• Agentic performance (τ-Voice) — Google reports 68.6% for Gemini 3.8 Live Extended Thinking. xAI reports 56.5% for Grok Voice Think Fast 2.0, roughly 11 points ahead of GPT-Realtime-2.1's 45.7% and about 19 ahead of Gemini 3.1 Flash's 37.7% at the time. If both vendor figures hold, Google has taken the category xAI was leading by about 12 points.

• Speech reasoning (Big Bench Audio) — Google reports 97.7% for the Extended Thinking model. xAI reports 97.2% for Grok. Effectively level.

• Task success on banking workflows — Google reports 35.1% on Sierra's τ-Voice-banking benchmark. xAI has not published a directly comparable figure.

• Conversational dynamics (Full Duplex Bench) — Grok reports 95.1%, slightly behind GPT-Realtime-2.1's 95.7% and well ahead of Gemini 3.1 Flash's 74.3% at the time. Google has published no Full Duplex Bench figure for Gemini 3.8 Live Extended Thinking.

The honesty note matters here more than usual. The 68.6%, 35.1%, and 97.7% are Google's own numbers, published on Google's blog, and nobody outside Google has reproduced them. The 56.5%, 97.2%, and 95.1% for Grok come from xAI or from Artificial Analysis and are, in the case of the index placement, publicly checkable. The only figures in this comparison that are independently readable are the two index scores: 82.6 and 81.3. Everything else is claim against claim, and the claims are made by parties with an interest in the outcome.

What can be said safely: Google is claiming a large lead on the specific dimension xAI used to justify its price increase. If that claim survives independent testing, this matchup is over on capability. If it does not, the 0.3-point index gap is the whole difference and the price gap becomes indefensible.

The bill: $1.50 versus $4.80 per hour

Cost is the part of this comparison that does not depend on anyone's benchmark methodology, because it is a published rate.

Grok Voice Think Fast 2.0 is priced at $0.08 per minute of audio, roughly $4.80 per hour. That was a 60% increase over Think Fast 1.0's $3.00 per hour, and users on the grok-voice-latest alias were migrated automatically on August 5, 2026 — developers who wanted to stay on the old model had to pin the versioned ID before that date or absorb the change.

Gemini 3.8 Live was announced at $0.005 per minute of audio input and $0.018 per minute of audio output. Artificial Analysis's cost-per-hour-of-input-audio chart puts it at $1.50 per hour — which is exactly 3.2 times cheaper than Grok's $4.80. For context, the same chart shows GPT-Realtime-2 High at $4.14, Qwen Audio 3.0 Realtime Plus at $4.42, and GPT-Realtime-2.1 High at $10.75.

The shape of the market in that list is worth a sentence. The cheapest entry on the board is a Gemini — Gemini 3.1 Flash Live at $1.42 per hour — and the second cheapest is Gemini 3.8 Live at $1.50. Google has positioned its newest voice model as the second-cheapest option in the category, below models it competes with on quality.

Notice what that means for anyone who migrated to Grok in August. Between the price increase and Google's launch, the same workload went from $3.00 an hour to $4.80 an hour on one side of this matchup and became available at $1.50 an hour on the other. Nothing about the models changed to cause it. Routing decisions made in July are now costing more than twice what the alternative costs, which is precisely the situation a router exists to avoid.

Where Grok still leads

Two things, and one of them is not close.

Latency. Grok Voice Think Fast 2.0 cut time to first audio from 1.25 seconds to 0.70 seconds, which is the fastest published figure in the category. Google has published no latency figure for Gemini 3.8 Live at all — only partner testimonials describing "impressive latency." On a phone call, 0.70 seconds is the difference between a conversation and a delay, and it is the one dimension where xAI's claim is both specific and unchallenged. If your product is telephony, this may outweigh everything above.

Transcription robustness. xAI evaluated word error rate across thousands of short phrases in 24 languages and reports 1.5–2.0× lower error than dedicated STT models including Deepgram Nova 3 and ElevenLabs Scribe v2, 1.4× lower than Think Fast 1.0, and roughly 10× lower error in noisy settings involving background noise and telephony compression. These are vendor-reported and not independently verified, and the noisy-environment claim is the one to test first, because it is the most specific and the easiest to falsify.

Grok also has practical developer affordances worth noting: an OpenAI Realtime-compatible WebSocket API, PCM16 at 24 kHz with μ-law for telephony, server-side voice activity detection, ephemeral tokens for browser apps, custom voice cloning from about a minute of speech, and reasoning-token volume roughly 0.4× that of the previous version.

Where Gemini 3.8 Live Extended Thinking leads

Beyond the index placement and the claimed agentic gap, the differentiators are structural:

• Vision — Gemini 3.8 Live processes near real-time visual input. Grok Voice Think Fast 2.0 has no comparable visual grounding.

• Language coverage — 97 supported languages with automatic mid-conversation switching, against Grok's 25+.

• Narrated reasoning — the Extended Thinking variant speaks while it works, using cues like "Let me check that…" and live progress narration, which is a more legible behaviour for a user waiting on a multi-step task.

• Watermarking — all generated audio carries SynthID.

• Cost — 3.2× cheaper per hour of input audio.

The counterweights are real. Google has published no latency figure, no context window, and no per-component numbers that anyone outside the company has reproduced. Its availability is a preview, not general availability: developers get it in the Gemini API and Google AI Studio, enterprises in private preview in Gemini Enterprise, and consumers through Gemini Live, Docs Live, Gmail Live, and Keep Live. Grok's model is generally available today and its API is documented.

The layer both of them sit on

One structural similarity is easy to miss and matters more than either spec sheet. Both models delegate. Grok's parallel reasoning nearly always terminates in a tool call or a lookup, and Google's background tool execution is explicitly built so the conversation continues while API calls finish. Underneath both voice front-ends is an ordinary text-model workload — and that is where the majority of your all-in cost and nearly all of your switching flexibility live.

OrcaRouter runs 190 models behind one key at provider list price with zero markup, which means a vendor price cut is live on our side the same day rather than at the next renewal. That is not a small thing in a category where one major model raised its per-minute rate by 60% within weeks. The other property that matters for voice is automatic failover: a call that dies because a backend timed out is a dropped customer, and the ability to swap the delegation target on live traffic without rewriting the voice integration is what makes the front-end choice reversible. To be precise about scope — the Grok Voice Think Fast 2.0 and Gemini 3.8 Live endpoints are not on our router; those come from the vendors' own APIs. The text models they delegate to generally are.

What would settle it

This matchup is unusually easy to resolve, because the deciding evidence is specific and pending. Three things would do it:

• Independent τ-Voice runs on Gemini 3.8 Live Extended Thinking. Google claims 68.6% against Grok's 56.5%. That is a 12-point claim in a category where Google has historically not led, and it is the single number that most changes the verdict.

• A published latency figure. Until Google prints one, Grok's 0.70 seconds stands unopposed and telephony buyers have only one defensible choice.

• General availability and stable pricing. Preview pricing moves. A $1.50 hourly rate that doubles at GA would rewrite this comparison completely.

Until those land, the defensible position is narrower than either announcement suggests. Grok Voice Think Fast 2.0 remains the right pick where latency is the constraint and the workload is telephony-shaped. Gemini 3.8 Live Extended Thinking is the right pick where reasoning depth, vision, or language breadth matter and you can tolerate preview status and unpublished latency. And the 0.3 points between them should not decide anything — the 3.2× should.