Phonic 1 vs ElevenLabs: What a Vendor Has to Pay You to Leave Tells You About the Fight

The most informative number in the Phonic 1 versus ElevenLabs matchup is not a benchmark. It is a discount. Phonic's launch offer, as reported at the end of October 2026, includes three months free for teams switching over from ElevenLabs or Bland, plus $250 in credits for anyone who signs up before November 30. Nobody subsidises a migration against a vendor they expect to beat on product. They subsidise it against a vendor whose customers are locked in by something other than the model.

That something is the whole point of ElevenLabs: a voice library of 17,000-plus voices, 90-plus languages, a text-to-speech line that is still the reference standard for rendered audio, and an agent platform wrapped around it. Phonic 1 is one model with one bill. ElevenLabs is a company you have built your product inside. The comparison only makes sense if you keep those two shapes in mind, because on a benchmark board the answer is clean and on a procurement decision it is not.

Eleven Realtime v1 Max is the model, and it is newer than you think

ElevenLabs' speech-to-speech entry is Eleven Realtime v1 Max, announced on October 20, 2026 in a rolling release alongside a smaller v1 Mini. That is ten days before Phonic 1's launch coverage — this is not a mature incumbent being challenged by a newcomer. It is two models that barely knew about each other when they shipped.

• Eleven Realtime v1 Max — speech-to-speech index 77.3, reasoning 99, conversational dynamics 94.1, agentic 47.6, arena preference 990, task success 87.1%, time to first audio 0.57s.

• Phonic 1 (High) — speech-to-speech index 87.1, reasoning 99, conversational dynamics not reported, agentic 79.9, arena preference 1,012, task success 92.3%, time to first audio 0.61s.

• Eleven Realtime v1 Mini — speech-to-speech index 64.4, reasoning 76, dynamics 95.8, agentic 46.8, arena 981, task success 60.3%, time to first audio 0.62s.

All figures above are Artificial Analysis's, not either vendor's. Note what the comparison does and does not say. Phonic wins the composite index by 9.8 points and the agentic component by 32.3 — a gap so large it is the only number in this matchup that should change an architecture. It wins arena preference by 22 points on a scale where a 22-point Elo difference is real but not decisive. It wins task success by 5.2 points. On latency it loses, narrowly, at 0.61s against 0.57s, and both are under the one-second mark that matters for a phone call.

The gap that should actually move your decision

Agentic performance, 79.9% against 47.6%, is the number to sit with. It measures whether the model can complete a task that requires calling tools and acting on the result, not merely whether it can hold a conversation — and it is the one place where a smaller, newer lab has beaten a company with 17,000 voices and a decade of product surface. If your agent's job is to look something up, book something, or change something while the caller waits, this is the figure that predicts whether your completion rate survives contact with real users.

ElevenLabs' 94.1 conversational-dynamics score against Phonic's unreported one is the honest hole in Phonic's scoreboard, and Artificial Analysis's own structure explains why it does not show up in the headline index. The composite is a weighted average of four components — speech reasoning, agentic performance, arena preference and task success rate — and conversational dynamics is reported alongside the index without being one of its inputs. So Phonic's missing dynamics figure costs it nothing on the 87.1 and tells you something the index cannot: nobody has yet measured how well the model carries a long, messy, self-interrupting conversation. Until that column is filled, claims that Phonic also sounds more natural than ElevenLabs are unsupported in both directions.

What the switch actually costs you, in things that are not the model

Price is the argument Phonic leads with, and on the headline rate it is right. Phonic's starter tier is $0.07 per conversation minute with no token pass-through. Eleven Realtime v1 Max is $0.080 per minute of connected audio plus model tokens at $1.80 per million input and $9.00 per million output — and above your concurrency allowance the rate bursts to $0.160. The per-minute fee is the visible half.

• Headline rate — Phonic 1 $0.07/min (Pro tier $0.065 with a $200 monthly minimum) against Eleven Realtime v1 Max $0.080/min.

• What the rate excludes — Phonic 1 states no per-token or LLM charges; Eleven Realtime v1 Max bills model tokens on top of the per-minute fee, and burst traffic at double rate.

• Voice inventory — ElevenLabs ships 17,000-plus voices and 90-plus languages; Phonic 1 states 50-plus languages with mid-conversation switching that preserves the same voice.

• Deployment — Phonic 1 offers cloud, single-tenant, VPC peering and self-host; ElevenLabs is a hosted platform, and its own release notes flagged the realtime models as unavailable in some data-residency environments at launch.

• API shape — Eleven Realtime v1 Max is OpenAI-Realtime-compatible over WebSocket with no WebRTC at launch; Phonic 1 runs a WebSocket endpoint at 24kHz PCM in both directions with inline tool declarations.

Three months free is a real number, and it is also a rounding error against the cost of moving. Voice identity is the item that does not appear on any comparison page. If your production agent speaks in a cloned brand voice, that voice lives in ElevenLabs' library under ElevenLabs' terms; Phonic offers cloning, but a cloned voice is not a portable file and re-cloning produces a different voice, not the same one. Teams that have built a brand around a specific timbre should price that in before they price the minute rate.

There is a second, quieter portability problem. An ElevenLabs agent is a stack: the voice, the prompt, the tools, the conversation flow. Moving the model means rebuilding the flow, and the tool definitions do not come with it. That glue is where a migration usually dies, and it is also the layer worth owning yourself rather than renting from any single vendor. The reasoning and tool-calling leg of an agent is the part that runs on OrcaRouter today — one key across 200+ models at list price with no markup, a routing rule that keeps cheap turns cheap, and automatic failover so a provider blip does not become dead air on a live call. Neither Phonic 1 nor Eleven Realtime v1 Max is on our catalogue; each still comes from its own vendor. The difference is what you keep when you leave.

Where ElevenLabs is still the right answer

ElevenLabs' TTS line remains the best-reasoned default for anything that is not a live agent. Its Eleven v4 voice model sits around 1,157 Elo on Artificial Analysis's controlled-voice arena at roughly $80 per million characters, and Eleven v3 at $100 per million characters — both well above Cartesia Sonic 3.6's normalized $49, and both earning it for narration, audiobooks, dubbing and branded announcements where the audio is the product.

If your use case is "read this text beautifully in our brand voice", the ElevenLabs versus Phonic 1 question is not competitive. Phonic does not sell a general narration engine; it sells a conversational agent. Buying a speech-to-speech model to render scripted copy is paying for reasoning you will never use, and getting voice quality that has not been independently scored above ElevenLabs' on the arena that measures exactly that.

There is also the matter of breadth and maturity. 90-plus languages against 50-plus, 17,000-plus voices against a curated set, and a platform with an agent product, an SDK and a changelog your team already reads. Those are the things a launch-week model has not had time to accumulate.

Who the discount is actually for

The switcher offer makes sense for one specific buyer: a team whose agent does real transactional work, whose voice identity is either generic or theirs to re-clone, and whose current stack is ElevenAgents end to end. For that team the agentic gap is a product problem, three months free is a real hedge against a migration that might fail, and the flat per-minute bill removes a modelling exercise their finance team has been doing monthly.

Everyone else should wait and watch what Phonic publishes next. Two things would settle this: a conversational-dynamics score on the same board, so the index's fourth component stops being an assumption, and an independent measurement of the vendor's tool-calling-while-speaking claim outside a demo. Neither exists yet. A model that is a week old, has one board behind it, and has never run a Black Friday is not something to migrate a revenue line onto — but it is exactly the kind of thing to put behind a failover path and measure against your own traffic for a month.

The honest summary is asymmetric. Phonic 1 is better at the thing that is hardest to build and easiest to measure: finishing an agentic task. ElevenLabs is better at everything that surrounds it and harder to measure: voice identity, language coverage, and a product your team already knows. The three months free is not a verdict on either. It is a vendor's admission that the surrounding things are what keep customers, and a bet that enough of them can be rebuilt in ninety days.