Eleven Realtime v1 Max vs GPT-Live-1: The Benchmark Gap Runs Against ElevenLabs

The press release for Eleven Realtime v1 Max leads with 98.8% on Big Bench Audio and a 0.57-second response. Both numbers are true, both are the best in their class, and both are beside the point when the comparison is G​PT-Live-1 — because the benchmark that decides voice agents in production is not speech reasoning, and on that one O​penAI's model wins by a margin wide enough to swamp the 0.7-second head start E​levenLabs is selling. The documented AI incidents and agent failures behind models in this class are collected in the OrcaRouter AI Incident Archive, where each entry carries its own sources, severity, confidence grade and disputed flag.

Run the same Artificial Analysis harness on both and the split is clean. Eleven Realtime v1 Max leads on speech reasoning, 98.8% to 89%, and on speed, 0.57 seconds to 1.24. G​PT-Live-1 leads on agentic task completion, 59.3% to unmeasured, on conversational dynamics, 97.3% to 94.1%, on arena preference, 1,047 Elo to 997, on task success, 90.9% to 87.1%, and on composite index, 80.1 to nothing at all.

That last figure is not rhetorical. Eleven Realtime v1 Max has no Speech to Speech Index score, because the index requires all four components and the τ-Voice agentic run has not been published for it. G​PT-Live-1 has an index. It is the highest one on the board.

The full scoreboard, side by side

All rows are Artificial Analysis measurements on its own harness unless labelled otherwise. Eleven Realtime figures are from the October 20, 2026 announcement; G​PT-Live-1's API availability dates from September 10, 2026, with the model itself shipping in ChatGPT on July 8.

• Speech to Speech Index — Eleven Realtime v1 Max not scored vs G​PT-Live-1 (Sol, low) 80.1

• Speech reasoning (Big Bench Audio) — Max 98.8% vs G​PT-Live-1 89%

• Agentic performance (τ-Voice) — Max not measured vs G​PT-Live-1 59.3%

• Conversational dynamics — Max 94.1% vs G​PT-Live-1 97.3%

• Task success rate — Max 87.1% vs G​PT-Live-1 90.9%

• Arena preference (Elo) — Max 997 vs G​PT-Live-1 1,047

• Time to first audio — Max 0.57s vs G​PT-Live-1 1.24s

• Cost per hour of input audio — Max $4.80 vs G​PT-Live-1 $4.47

• Transport — Max WebSocket only at launch vs G​PT-Live-1 WebRTC only

One row is closer than the headline suggests: cost. $4.80 against $4.47 per hour of audio puts these two models in the same price band, which is the opposite of the story you get by comparing Eleven Realtime to G​emini 3.8 Live's $0.84. On the O​penAI comparison, the E​levenLabs premium is 7%.

Why the reasoning score points the wrong way

Big Bench Audio is 1,000 audio questions adapted from Big Bench Hard across four categories — formal fallacies, navigation, object counting and a boolean word problem set. The model hears a question and answers out loud, and the answer is graded right or wrong. It measures how well a model reasons about audio content, which is not the same as how well it runs a conversation or completes a task.

Max's 98.8% puts it third on that board, behind StepAudio 3 Realtime (100%) and Q​wen Audio 3.0 Realtime Plus (99.2%), and ahead of Qwen3.5 Omni Plus Realtime (98.7%) and G​rok Voice 3.0 (98.5%). G​PT-Live-1's 89% is twelfth. If your agent's job is to listen to a long spoken description and answer a question about it, that nine-point gap is meaningful and E​levenLabs is the better buy.

But E​levenLabs does not sell Max for that. Its own page positions Max for "complex tasks" with "multi-step tool use" — customer support, lead qualification, assisted scheduling. That is τ-Voice territory, and τ-Voice is where Max has no number and G​PT-Live-1 has 59.3%.

The O​penAI number also needs its own caveat, and O​penAI's own documentation supplies it: the published deployment evidence is thin. The banking voice-support benchmark O​penAI reported at API launch put G​PT-Live-1's pass rate at 32%, against 12.4% for G​PT-Realtime-2.1. A 32% pass rate is a large improvement and still means the system fails two-thirds of a regulated workflow. Artificial Analysis' τ-Voice figure is the more useful one because it is run on a common harness, but it is a mid-field score, not a coronation.

Where G​PT-Live-1's advantage is structural, not marginal

The 3.2-point conversational-dynamics gap and the 50-point arena Elo gap come from an architectural choice that E​levenLabs did not make. G​PT-Live-1 is full duplex in the strict sense: it listens and speaks on the same timeline, and when a caller asks something that needs real work, it hands the task across an asynchronous boundary to a separate backend model that keeps thinking while the spoken conversation continues. The voice layer is a front end. The reasoning happens elsewhere and folds back in when it is ready.

That design is why G​PT-Live-1 can run tool calls without the line going quiet, and it is also why its reasoning score is lower: the model that answers the Big Bench question is not the one doing the deep thinking in a real session. Eleven Realtime Max takes the opposite bet. One network does recognition, turn-taking, reasoning and voice, which is what buys 0.57 seconds — there is no handoff to wait on — and what caps how much reasoning can happen before the caller notices the pause.

Which bet is right depends on whether your latency budget or your task-completion budget is tighter. For a scheduling agent that has to check a calendar, confirm a slot and write a record, the handoff model wins on completion. For a tutor answering a spoken question in one breath, the merged model wins on feel.

The billing and transport details that decide the integration

G​PT-Live-1 bills at a flat $0.05 per minute of voice, charged by the second, with session initialisation billing 15 seconds up front and crediting it against later duration. The backend reasoning model is billed separately at its own normal rates, so a cheap voice layer can still produce an expensive call. OrcaRouter cannot route the voice layer — the Live Sessions endpoint at v1/live/sessions belongs to O​penAI alone — but the backend half is just another model call, and on OrcaRouter that is 209 models from 16 providers on one key at provider list price with no markup, with failover if your chosen backend errors mid-call.

Eleven Realtime bills the other way around: standard ElevenAgents per-minute plus input, cached-input and output tokens at the Realtime model's rate, with Max costing more per token than Mini. Neither model's per-minute rate is published on E​levenLabs' pricing page; the dashboard shows an estimate, and Artificial Analysis' $4.80 is a normalised figure rather than a list price.

Transport is the sharper asymmetry. Eleven Realtime's API is O​penAI-Realtime-compatible over WebSocket — the docs' own example imports OpenAIRealtimeWebSocket from @openai/agents-realtime and points it at wss://api.elevenlabs.io/v1/convai/realtime. G​PT-Live-1 is WebRTC only, and probing the older Realtime path with a G​PT-Live model ID returns an error saying sessions require WebRTC. So the compatibility runs one way: an O​penAI Realtime client can move to E​levenLabs with an endpoint change, but an O​penAI Realtime client cannot move to G​PT-Live-1 without a transport rewrite. Teams that migrate to G​PT-Live-1 are writing an RTCPeerConnection handshake, an oai-events data channel and a server-side offer exchange; teams that migrate to Eleven Realtime are editing two strings.

The rollout caveats on the E​levenLabs side

E​levenLabs' documentation adds constraints that no launch coverage mentions:

• Availability — Realtime models are "rolling out gradually"; a workspace without the Voice mode setting gets Realtime models are not available for this workspace back from the API

• Data residency — Max is available in the US and EU but not in India or Singapore, where only Mini is offered

• Compliance — Zero Retention Mode and HIPAA support Max in the US and EU; Mini supports neither

• Managed settings — the LLM, reasoning effort, thinking budget, parallel tool calls, T​TS model, expressive mode and ASR provider are all controlled by E​levenLabs while a Realtime model is set, and workflow nodes cannot override them

• Custom LLMs — Realtime mode is unavailable while an agent uses a custom LLM or Amazon Bedrock

G​PT-Live-1's own constraints are tighter in one respect and looser in another: it has no Chat Completions, no Responses, no Realtime compatibility layer, no fine-tuning and no image or video input, but its tier limits are denominated in concurrent sessions — 25 on Tier 1 rising to 500 on Tier 5 — rather than requests per minute.

Who should choose which

Choose G​PT-Live-1 when the agent has to complete work: bookings, changes, refunds, anything ending in a database write. It has the higher task-success rate, the higher arena preference, the higher composite index and the lower per-hour cost, and it is the only one of the two with a published agentic score. Budget for the WebRTC rewrite and for the backend model's separate bill.

Choose Eleven Realtime v1 Max when the conversation is the product and the reasoning is shallow: a tutor, a character, a support line where the complaint is being talked over. You get the fastest strong reasoner on the board, the simplest migration from an existing O​penAI Realtime integration, and a 7% price premium over G​PT-Live-1 rather than the 5.7× premium over G​emini. You accept that its agentic behaviour is unmeasured and that the launch is still partial.

The result that would flip this piece is a τ-Voice score for Max. E​levenLabs is asking to be trusted as a task-completing model on the strength of a reasoning benchmark; the task benchmark exists, it has been run on ten other voice models, and it has not been run on this one. That is a gap in the evidence, not in the model — but it is the gap that determines which of these two you should put in front of a paying customer.