Cekura Bench Overview
This Cekura Bench review examines the speech-to-speech model benchmark as it stands in October 2026, after its Product Hunt launch on October 8. Cekura Bench measures realtime voice AI models on live phone calls — reliability, task success, latency, and cost — and publishes every call with transcripts, tool calls, and scores.
What is Cekura Bench?
Cekura Bench is an open, reproducible benchmark for speech-to-speech (S2S) models. Every model runs the same voice agent, prompt, tools, connection, caller scenarios, and scoring rules — only the realtime model changes. The leaderboard ranks models on pass³: the share of scenarios where all three retained calls passed every required check. A scenario must work repeatedly to count.
The methodology is unusually transparent. Each model links to its actual calls, with full transcripts and per-call records of build, configuration, tool calls, usage, and timing. Pricing is kept in a versioned public table with the date and source page for every price — because, as the maintainers note, an undated price is not reproducible.
Recent results tell a clear story: GPT Realtime 2.1 is the most reliable at 79.3% pass³, Phonic v1 replies fastest at 1.60s median, and GPT Realtime 2.1 Mini is cheapest at $0.020 per call minute. The benchmark also covers speech-to-text models, ranking 17 STT engines on word error rate and latency.
Pricing for running your own tests: Cekura starts free with 300 credits (about 60 minutes of voice testing), no card required, then bills $0.25 per voice testing minute. The benchmark harness itself is open source on GitHub under cekura-ai/cekura-bench.
How Cekura Bench Works
The harness has two lanes. The service bench measures a provider’s realtime service over its own websocket with a paced caller, calibrated onset detection, latency, barge-in, backchannel, transcription, and task probes, plus seeded audio degradations. The agent bench runs one Pipecat reference agent per call, with native and cascade rows under identical prompts, tools, transport, and hang-up rules.
Measures include interruption score (does the agent let the caller finish?), voice tone and clarity on the actual phone-call path, tool-call accuracy, saved-data accuracy, and cost per minute. Seeded degradations test robustness the way real phone networks behave — not lab-clean audio.
For teams choosing a voice model, the head-to-head comparison tool is the killer feature: pick any two tested models and get reliability, accuracy, latency, failure, and cost results side by side, all measured on the same live calls.
Who Should Use Cekura Bench?
Cekura Bench fits conversational AI teams choosing between realtime voice models, and anyone who needs to justify a model choice with reproducible numbers rather than vendor claims. If you’re building voice agents, this is required reading.
It’s less useful if you’re not building voice AI at all, or if you need benchmarks for text models — this is strictly the speech path. Teams wanting a fully managed QA loop (simulation through production monitoring) should look at Cekura’s main platform rather than the benchmark.
Our Verdict on Cekura Bench
Cekura Bench is the most credible independent voice-AI benchmark we’ve seen in 2026. Public calls, versioned pricing, and a strict pass³ methodology make it genuinely useful instead of marketing.
The bottom line of this Cekura Bench review: check it before choosing a speech-to-speech model — and run your own scenarios before committing your voice stack to any single provider.
For voice generation tools, explore our AI Audio & Voice category.
Key Features
- pass³ reliability scoring across repeated live phone calls
- Head-to-head comparison of any two tested voice models
- Public call records with transcripts, tool calls, and scores
- Speech-to-text leaderboard covering 17 STT engines
- Seeded audio degradations simulating real phone networks
- Versioned public pricing table with dated sources
- Open-source harness (cekura-ai/cekura-bench) on GitHub
Cekura Bench Pricing
| Plan | Price |
|---|---|
| Leaderboard | $0 |
| Testing Free | $0 (300 credits) |
| Testing Paid | $0.25/min |
Pricing checked on November 3, 2026 — always confirm on the official site.
Cekura Bench Pros & Cons
✓ Pros
- Fully reproducible methodology with public call records
- pass³ scoring rewards consistency, not lucky runs
- Head-to-head model comparison is genuinely useful
- Versioned pricing table keeps cost claims honest
- Open-source harness invites community scrutiny
✕ Cons
- Benchmark scores depend on the agent setup — your mileage will vary with different prompts and tools
- Only measures the speech path, not your full production stack (transport, infra)
- New benchmark; methodology is still being stress-tested by the community
- Running your own full benchmark campaigns costs $0.25/minute after free credits
Cekura Bench FAQs
What does Cekura Bench measure?
What is the best speech-to-speech model in 2026?
How much does Cekura Bench cost?
Is Cekura Bench open source?
How is Cekura Bench different from vendor benchmarks?
Best Cekura Bench Alternatives
Foley
★★★☆☆Foley review (2026): a voice to-do list that files tasks from speech and asks before guessing — free desktop app, $8/mo Pro. Features and honest verdict.
Altered
★★★★☆Altered reviewed for 2026 — AI voice cloning, morphing, and TTS from $30/mo. Voice quality, real pros and cons, and who should actually pay for it.
Synthesys
★★★☆☆Synthesys reviewed for 2026 — 1,000+ AI voices and avatar videos in one studio. Worth $20/mo? Features, pricing, and honest verdict.
SpotScribe
★★★★☆SpotScribe reviewed for 2026 — AI podcast show-notes generator with summaries, timestamped topics and notable mentions, from $6.99/mo. Honest pros, cons, and verdict.