Eleven Realtime v1 Max Leads the Board on Speed and Speech Reasoning — and Has No Index Score

Eleven Realtime v1 Max answers in 0.57 seconds and scores 98.8% on Big Bench Audio, and Eleven Realtime v1 Mini answers in 0.62 seconds and scores 76% on the same test. Those two figures, read together, are the most useful thing in E​levenLabs' October 20, 2026 announcement — not because either is impressive on its own, but because of the 23-point gap between them and the fact that the weaker one is what you get by default. 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.

There is a second, stranger fact on the same leaderboard. Artificial Analysis lists both models with a time to first audio, a Big Bench Audio score, a conversational-dynamics score, an arena Elo and a task success rate — and a dash where the Speech to Speech Index should be. Max has no index score. Not a low one. None.

The reason is mechanical, and it explains most of what follows. The index is an equal-weighted average of four components: speech reasoning (Big Bench Audio), agentic performance on the τ-Voice customer-service benchmark, arena preference, and task success rate. A model needs all four to be scored. Max has three. E​levenLabs sells Max specifically for "complex tasks" with "multi-step tool use" — and the tool-use benchmark is the one nobody has run on it yet.

What E​levenLabs actually shipped

The Eleven Realtime family is a different kind of product from the text-to-speech models E​levenLabs is known for. Each Realtime model is a single network that handles recognition, turn-taking, the reasoning step and voice generation together, rather than a pipeline of speech-to-text, a language model and a text-to-speech voice stitched together by an orchestrator. E​levenLabs' pitch is that tuning those four stages as one system removes the seams — the half-second of dead air after a caller stops talking, the agent that talks over an interruption, the mismatch between what a transcription heard and what the voice says back.

Both models are reachable two ways: inside ElevenAgents, where you pick a model from a dropdown and the platform configures the rest, or through a Realtime API that is deliberately wire-compatible with O​penAI's, at wss://api.elevenlabs.io/v1/convai/realtime. Change the endpoint and the model name, and an existing O​penAI Realtime client keeps its event handlers.

What shipped on October 20 is the announcement, not the switch-on. E​levenLabs' own documentation still describes the models as "rolling out gradually," and says a workspace without the Voice mode setting will have the API reject realtime_model outright with "Realtime models are not available for this workspace." The model identifiers are older than the press cycle, too: the August 3 changelog entry for ElevenAgents records that realtime agent configuration had already begun defaulting to eleven_realtime_v1_mini. Treat the launch date as the day the numbers became public, not the day the endpoints opened.

The two models are not two sizes of the same thing

E​levenLabs frames Mini and Max as a cost-quality dial: Mini for assistants, tutors and characters; Max for support, lead qualification and tool-using workflows. The published measurements say the dial moves further than that framing suggests. These are Artificial Analysis figures, run independently on the same harness rather than quoted from E​levenLabs' page:

• Speech reasoning (Big Bench Audio) — Eleven Realtime v1 Max 98.8% vs Eleven Realtime v1 Mini 76%

• Task success rate — Max 87.1% vs Mini 60.3%

• Conversational dynamics (Full-Duplex-Bench) — Max 94.1% vs Mini 95.8%

• Arena preference (Elo) — Max 997 vs Mini 988

• Time to first audio — Max 0.57s vs Mini 0.62s

• Speech to Speech Index — Max not scored vs Mini not scored

Read the last three rows before the first two. On the thing a caller consciously feels — does this sound like a conversation — Mini is the better model and the faster one is Max by only 50 milliseconds. The 23-point reasoning gap and the 27-point task-success gap are real and large, but they show up in whether the agent gets the job done, not in whether it feels good to talk to. That is exactly the trade a support queue makes, and exactly why the default matters: every new agent that never touches the setting runs on the model that resolves 60.3% of tasks instead of 87.1%.

The gap also explains the dash in the index column. Mini is missing an agentic score too, which is unsurprising for a model positioned as the cheap one. Max is missing it despite being positioned as the capable one.

What the missing row means in practice

Task success rate and τ-Voice are not the same measurement, and conflating them is the easiest mistake to make with this release. Task success rate asks whether the conversation ended with the correct final tool call. τ-Voice drops the model into a replica customer-service scenario — a flight change, a disputed charge, a telecom issue — with domain tools and a policy document, and scores whether the database ends in the right state. It is the difference between "the agent called the right function" and "the agent resolved the request without breaking policy."

E​levenLabs markets Max for precisely that second job. The claim on the product page is that Max "reasons through complex questions and calls tools to get things done without slowing the conversation down." Nothing published so far tests that claim end to end. The 87.1% task success rate is a strong number; it is also the easier half of the evaluation, and it sits next to a 94.1% conversational-dynamics score that is a point and a half below Mini's.

For a team deciding today, the honest summary is that Max is the best-measured speech-reasoning model in the field — 98.8% trails only StepAudio 3 Realtime's 100% and Q​wen Audio 3.0 Realtime Plus's 99.2% on the same board — and that its agentic behaviour is unmeasured while its competitors' is not. G​emini 3.8 Live Extended Thinking has a τ-Voice score of 68.6%. G​PT-Live-1 (Astra, medium) has 74.5%. Max has a dash.

What a minute of it costs

E​levenLabs does not publish a per-model number. Realtime billing is the standard ElevenAgents per-minute rate plus input, cached-input and output tokens at the Realtime model's own rates, with Max more expensive per token than Mini; the dashboard shows an estimated cost per minute, and the Realtime API bills per minute of connected audio plus tokens. The pricing page has no Realtime line item at all.

Artificial Analysis normalises all of that into one comparable figure — the cost of running a fixed 40-question Big Bench Audio subset, expressed per hour of input audio. Both E​levenLabs models land at $4.80. That identical number is the interesting part: whatever Max's per-token premium is, the per-minute agent charge dominates the bill at conversational volumes, so the expensive model and the cheap model cost the same to talk to. The savings from Mini appear where tokens are heavy, not where audio is long.

For scale, on the same measure: G​emini 3.8 Live is $0.84 an hour, G​emini 3.8 Live Extended Thinking is $3.50, G​PT-Live-1 (Sol, low) is $4.47, and G​rok Voice 3.0 is $4.80. Eleven Realtime is priced in the middle of the field, not at the bottom, and it is roughly 5.7 times the standard G​emini Live rate.

That is the point where a routing layer earns its keep even though it cannot touch the voice model itself. Eleven Realtime is available through E​levenLabs' own API and nowhere else — no third-party host carries eleven_realtime_v1_max, and OrcaRouter does not either, so there is nothing to route for the speech layer. What you can route is everything the agent calls: the lookup services, the CRM writes, the summariser that runs after the call, and the backup model a cascade falls back to when the realtime model errors out. On OrcaRouter those are 209 models from 16 upstream providers behind one key, at provider list price with no markup, so the half of the bill that scales with how hard your callers think is the half you can actually optimise.

The integration detail that decides your week

E​levenLabs made a bet that most realtime vendors did not: its API speaks O​penAI's Realtime event protocol. The seven-line example in the docs imports OpenAIRealtimeWebSocket, RealtimeAgent and RealtimeSession from @openai/agents-realtime, points the transport at E​levenLabs' WebSocket URL, and passes model: "eleven_realtime_v1_mini". Everything else — the client secret, the session config, the audio output modality, the voice — is unchanged.

The cost of that convenience is transport. There is no WebRTC at launch; the docs say so directly, and any WebRTC integration has to move to WebSocket first. For teams already on a WebSocket Realtime integration that is a non-event. For teams that built on WebRTC because it handles jitter and mobile networks better, it is a rewrite of the transport layer, not a model swap.

The other constraint is architectural: Realtime models run in ElevenAgents and in the Realtime API as separate products. If you need phone numbers, WhatsApp, a knowledge base, guardrails, scripted testing or conversation monitoring, you are on ElevenAgents, and your agent's tools and prompt travel with it. If you are on the Realtime API, you get the model and the events, not the platform around them. You cannot have both on one connection.

Where it does not travel

The rollout is uneven in ways that matter more to European and Indian teams than to US ones:

• Data residency — both models are available in the US and EU; India and Singapore get Mini only, so Max is not an option for agents pinned to those environments

• Compliance — Zero Retention Mode and HIPAA support Max in the US and EU, and do not support Mini at all

• Token limit — Max requires a limit of at least 1,024 tokens, or no limit, and workflow nodes cannot override it

• Custom backends — Realtime mode is unavailable while an agent uses a custom LLM or Amazon Bedrock; you must move to a built-in LLM first

• Agent transfers — a Realtime-mode agent can only transfer to and from other Realtime-mode agents, though Mini-to-Max transfers are allowed

Every one of those rules is enforced at publish time, not at save time. A draft that breaks one still saves; publishing fails and names the setting.

What to watch

Two results would settle this release. The first is a τ-Voice score for Max — the one number that would test the claim E​levenLabs is actually selling. The second is WebRTC support, which is the difference between a model you can demo and a model you can put on a mobile call. Until then, the defensible position is narrow and useful: Mini is a good default for conversational agents that do not need to complete tasks, Max is the fastest strong reasoner on the board and plausibly the better support agent, and nobody has yet published the test that would prove the second half.