Cartesia Claims #1 on Artificial Analysis. The Leaderboard Says #4

Cartesia's Sonic page still carries a badge reading "Ranked #1 in Speech Arena" from Artificial Analysis. The live board does not match it. On the Provider Voice Arena captured 16 October 2026, Sonic 3.6 is fourth at Elo 1285 (±15), 45 points behind Eleven v4 Turbo at 1330.

The claim was accurate for several weeks after launch. Sonic's own score barely moved. The rank fell because newer models from ElevenLabs and others scored higher.

Does Cartesia Sonic 3.6 still rank #1 on Artificial Analysis?

No. On the Provider Voice Arena as of 16 October 2026, Sonic 3.6 is fourth of 97 listed models, with a rank band of 3 to 5. On the Controlled Voice Arena it is fifth, with a band of 3 to 6.

These are two separate boards with separate Elo anchors, so scores are not comparable across them. Both are Artificial Analysis (independent) data. The table below shows the top of the Provider Voice board, the one Cartesia cites.

Source: Artificial Analysis Provider Voice Arena, English (US and UK accents), all categories, independent, captured 16 October 2026 (UTC morning). Prices as of the same date. Every model in the table is evaluated on 8 arena voices.

What does Cartesia's own site say, and what is it based on?

The Sonic page shows the badge without a date. A separate Cartesia comparison page, retrieved 16 October 2026, says Sonic 3.6 is "#1 of the 94 models on the Artificial Analysis Provider Voice Arena." Its chart is labelled August 2026 and shows Sonic 3.6 at 1282, against Eleven v3 at 1177, Turbo v2.5 at 1103 and Flash v2.5 at 1084.

The same page's meta description says Sonic 3.6 outscores every ElevenLabs model on quality benchmarks. That is a company claim, and the chart behind it omits the two ElevenLabs models that now lead. RuntimeWire reports that ElevenLabs shipped Eleven v4 and Eleven v4 Turbo on 28 September 2026. The live board lists Eleven v4 Turbo at 1330 and Eleven v4 at 1329, against 1285 for Sonic 3.6.

Cartesia also says Sonic wins 92% of blind US English comparisons against Eleven v3. That figure is from Cartesia's own page, not Artificial Analysis, and Eleven v3 is no longer ElevenLabs' newest model.

When was Sonic 3.6 actually #1?

From its August 2026 launch until at least 23 September 2026, according to Artificial Analysis and two competing vendors citing its board. Artificial Analysis said Sonic 3.6 took first place on both arenas, with a Provider Voice Elo of 1,286 (+19/-19) on 1,300 appearances.

Treat the intermediate rows as snapshots of a moving board, not a clean time series. The sources captured at different times, and the Elo scale is relative. The sources also disagree slightly on the same period: MarkTechPost's 18 August Controlled Voice figure (1,123) differs from Artificial Analysis's post (1,144).

Why did Sonic drop to #4 if its score barely changed?

Because Elo is relative, and better models entered. Sonic 3.6's Elo was 1,286 in August and is 1285 on 16 October 2026. The three models above it are all newer: Eleven v4 Turbo and Eleven v4 (Sept 2026) and Qwen-Audio-3.1-TTS-Plus (Sept 2026).

Artificial Analysis's methodology says all eligible votes in a board's history count equally, with no time weighting or recency decay. That makes a stable score for an older model plausible. It also means a lead is only as durable as the next release cycle.

A rough reading of the gap, using the standard Elo scale the methodology says it uses: 45 points implies about a 56% expected preference for the higher-rated model. That is my arithmetic, not an Artificial Analysis statistic. Sonic 3.6's 95% interval (1270 to 1300) does not overlap the top two (1312 and above), but it does overlap Qwen-Audio-3.1-TTS-Plus (1280 to 1312).

How does Artificial Analysis score the Speech Arena?

It uses blind pairwise votes fit with a Bradley-Terry model. Per the methodology page, each comparison presents two clips generated from the same prompt, model names are hidden, order is randomized, and listeners must play part of each clip before voting. Votes are binary with no tie option.

Other design details, all Artificial Analysis self-described:

• Aggregation: Bradley-Terry maximum likelihood, one model per arena anchored at 1000. Ratings are recomputed multiple times daily from the full vote history.
• Uncertainty: 95% confidence intervals and rank ranges are displayed. Models are ordered by unrounded Elo.
• Prompts: four categories (Customer Service, Entertainment, Knowledge Sharing, Assistants), including alphanumeric and numeric sequences. The prompt set is refreshed over time.
• Audio handling: clips are generated once per model, converted to a consistent format and loudness-normalized.
• Matchup sampling: favors models needing comparisons, caps the share any single opponent gets, and down-weights repeated pairings.
• Provider Voice Arena: 8 voices per model (2 each for US female, US male, UK female, UK male).
• Controlled Voice Arena: the same 8 professionally recorded reference voices cloned for every model, to separate model quality from voice choice.

Who runs it, and are there conflicts of interest?

Artificial Analysis runs both arenas. Its stated position: "No compensation is received from any providers for listing or favorable outcomes." That is a self-description, and the page does not describe an outside audit.

Three details deserve a note. In the Provider Voice Arena, model creators may request specific voices where many are available. Most ratings come from a paid panel of native US or UK English speakers, with paid and public pools compared as a consistency check. And every vendor that cites the board in marketing (Cartesia, ElevenLabs, Inworld, Speechify) has a commercial reason to cite the snapshot that flatters it.

None of this suggests Artificial Analysis changed Sonic's rank. It does mean the vendor pages here are summaries with an interest, which is why dated snapshots matter. The same lab-reported versus independent gap shows up in coding benchmarks, as in Claude Opus 5.5 on Terminal-Bench 4.0 and in the harness effect on coding agent benchmarks.

Does #4 mean Sonic 3.6 is a worse model?

Not necessarily, and the Elo gap does not capture everything. A top-five position within overlapping intervals of third place is a defensible claim. "#1" is not, as of 16 October 2026.

What the arena does not measure matters for production use:

• Latency: Cartesia's "first audio in under 90ms" is a company claim on its own page. The Artificial Analysis board above ranks listener preference, not time to first audio.
• Languages: Cartesia says Sonic supports 44 languages. The English Provider Voice board tests US and UK accents only.
• Price: Artificial Analysis lists Sonic 3.6 at $49.0 per 1M characters as of 16 October 2026, against $40.0 for Eleven v4 Turbo and $19.3 for Qwen-Audio-3.1-TTS-Plus. Higher price with a lower rank is a harder sell on this axis alone.
• Controlled Voice: with voices held constant, Sonic 3.6 scores 1149, behind HeyGen Voice (1198), Qwen-Audio-3.1-TTS-Plus (1186), Eleven v4 Turbo (1162) and Eleven v4 (1159) as of 16 October 2026.

What are the weaknesses of an arena ranking like this?

The main limits are sample size per model, voice selection and prompt coverage. Each of the top seven models is represented by 8 voices and roughly 1,700 to 2,800 samples on the Provider board. That is enough for intervals of about ±15 Elo but not enough to separate neighbors.

Rank bands show the problem directly. Sonic 3.6 is fourth by point estimate, but its band is 3 to 5. Positions three to seven span 34 points in the table above (1296 to 1262), with heavily overlapping bands, so small shifts reorder them.

Preference is also not task success. Listener votes on short prompts do not test long-form stability, phone-number readback under noise, or behavior on your own domain text. Artificial Analysis publishes a separate benchmark for pronunciation checks. A team choosing a voice should run its own audio rather than rely on any single Elo number.

Finally, a plausible analogue of contamination is vendors tuning for the arena's published prompt categories and voice-selection rules. That is an inference, not something the methodology documents. The refreshed prompt set would limit it.

What's still unclear?

• When the badge was last updated. Cartesia's pages do not date the "#1" badge. Only the chart on the comparison page is dated (August 2026).
• Snapshot drift. Figures from August to September come from different captures. Elo values for Sonic 3.6 range from 1,273 to 1,286 across sources, and the model count rises from 92 to 94 to 97 across sources.
• Non-English boards. This article covers the English Provider Voice and Controlled Voice boards only. Rankings on language-specific boards may differ.
• Whether the ElevenLabs lead holds. Eleven v4 and v4 Turbo are weeks old, and early ratings can shift as votes accumulate. Both fall inside a 1-2 rank band, so their order is not settled.
• Weakest point of this argument. A board position is a snapshot. If Cartesia releases a new model, the badge could be true again within days.