ModelsArchitectures & capability
OpenAI case study says Retell runs 100M monthly calls and GPT-6 Astra led its voice-agent benchmark
An OpenAI case study says Retell handles 100M voice-agent calls a month and that GPT-6 Astra led Retell's benchmark. Retell's pricing page, however, recommends cheaper models, and its guide names GPT 4.1 as the production default.
According to arXiv 2609.20152 (Sep 2026) [8]: OpenAI case study says Retell runs 100M monthly calls and GPT-6 Astra led its voice-agent benchmark.
AT Pulse research deskPublished · Updated · 5 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- What happened
- GPT-6 Astra topped Retell's own benchmark but is priced at $0.32 per minute. [1] [2] [5]
- Original source
- OpenAI News article, company source
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- Studies
- arXiv 2609.20152 (Sep 2026) [8] date from arXiv ID
- When
- October 27, 2026 article published
- Covered
- Oct 28 edition
- Evidence
- 12 sources across 8 sites
Reach: not measured yet
GPT-6 Astra topped Retell's own benchmark but is priced at $0.32 per minute. That is five times the GPT 5.6 Terra and Claude 5 Sonnet options Retell marks as recommended, and Retell's guide names GPT 4.1 as its most-used model in production. [1] [2] [5]
In voice, speed and precision decide model choice. Retell's guide says replies should come in under 800 ms and that callers hang up above 1.5 s. MTVA-Bench found models chose tools at similar rates, with scores within 6.4 points, but overall scores differed by 24.4 points. The gaps came from argument values, action order and rule compliance. [5] [8]
Astra costs five times Retell's recommended options, and callers hang up on slow responses. Implication: teams should test per-minute cost, latency and tool-argument accuracy on their own call flows instead of defaulting to a benchmark leader.
Executive brief
OpenAI's customer story on Retell says GPT-6 Astra scored highest on Retell's voice-agent benchmark. Retell's own pricing page, however, lists Astra at $0.32/minute standard, five times the cost of the GPT 5.6 Terra and Claude 5 Sonnet options that Retell marks as "Recommended." The story also says Retell handles 100M monthly calls and that Codex helped it reach a $1.6M annual revenue run rate per employee (OpenAI). Only a feed excerpt of the OpenAI post was available. Retell's benchmark method has not been published, and no outside party has confirmed the call-volume or per-employee figures.
What changed and event timeline
Retell raises a small seed round
Alt Capital led a $4.6M round. Sacra puts total funding at about $5.1M, which is small next to the revenue estimates that came later (;).
Retell names GPT 4.1 as its production default
Retell's guide calls GPT 4.1 the most popular LLM across "100M+ calls per month." It argues that latency matters more in voice than leaderboard scores ().
Third-party Medicare test
Cekura ran 414 calls across six platforms. Retell had the highest workflow accuracy at 95.7% ().
GPT-6 Astra launches
The first rollout went through OpenAI's Daybreak program, priced at $10/$50 per million input/output tokens ().
Research paper on the LLM inside voice agents
MTVA-Bench tests the language model separately within cascaded voice pipelines. It covers 490 scenarios in 7 languages ().
OpenAI publishes the Retell case study
The post reports 100M monthly calls, Astra leading Retell's benchmark, and $1.6M ARR per employee helped by Codex ().
Capabilities and access
- Models on Retell: GPT 6 Astra ($0.32/min standard, $0.64/min fast tier), GPT 6 Sol and 6.1 Sol ($0.064/min), GPT 6 Luna ($0.0032/min), GPT 5.6 Terra, GPT 5.5, and Claude 5 and 5.5 Sonnet (Retell pricing).
- GPT-6 Astra: 1M-token context, 128k max output, text/image/file input, no video input, closed weights (Vals AI).
- Codex: used for Retell's internal engineering according to the OpenAI post. How it was deployed is not described.
Read the full section
- Models on Retell: GPT 6 Astra ($0.32/min standard, $0.64/min fast tier), GPT 6 Sol and 6.1 Sol ($0.064/min), GPT 6 Luna ($0.0032/min), GPT 5.6 Terra, GPT 5.5, and Claude 5 and 5.5 Sonnet (Retell pricing).
- GPT-6 Astra: 1M-token context, 128k max output, text/image/file input, no video input, closed weights (Vals AI).
- Codex: used for Retell's internal engineering according to the OpenAI post. How it was deployed is not described.
Technical analysis for researchers and developers
- Pipeline type: Retell runs a cascaded pipeline (speech recognition → LLM → speech synthesis) and supports multiple LLM providers (Sacra).
- Latency: Retell's guide says responses should arrive in under 800 ms and that callers start hanging up above 1.5 s (Retell blog).
- Where models differ: MTVA-Bench found tool-selection accuracy within 6.4 points across six models, but a 24.4-point spread in overall scores.
Read the full section
- Pipeline type: Retell runs a cascaded pipeline (speech recognition → LLM → speech synthesis) and supports multiple LLM providers (Sacra). Swapping in a different model only changes the reasoning layer.
- Latency: Retell's guide says responses should arrive in under 800 ms and that callers start hanging up above 1.5 s (Retell blog). This is why a frontier model can top a benchmark and still not be the right production default.
- Where models differ: MTVA-Bench found tool-selection accuracy within 6.4 points across six models, but a 24.4-point spread in overall scores. The gaps came from argument values, action sequencing and rule compliance (arXiv).
- Reproducibility: none of the reviewed sources documents the tasks, scoring or model settings behind Retell's Astra result.
Claims and evidence
- 100M monthly calls
- Astra tops Retell's benchmark
- $1.6M ARR per employee
Read the full section
| Claim | Evidence |
| 100M monthly calls | Reported by Retell/OpenAI (OpenAI; Retell). Coverage notes that the measurement period and the definition of a "call" are not stated (1023 Jack). |
| Astra tops Retell's benchmark | Vendor-run and unpublished. No independent replication found. |
| $1.6M ARR per employee | Retell has not disclosed headcount. Combined with Sacra's $80M ARR estimate for Aug 2026, it would imply roughly 50 staff (inference only). |
| Platform workflow accuracy | Run by a third party, but published on Retell's blog and not tied to a named LLM (Retell). |
Context and prior work
- Earlier LLM tests for voice: Daily.co benchmarked LLMs on tool calling, instruction following and grounding over long voice-style conversations (Daily).
- Growth estimates: Sacra estimates Retell grew from about $5.4M revenue in 2024 to $45M by end-2025 and $60M annualized by April 2026 (Sacra).
- Astra's general standing: Astra ranks #5 overall on the Vals Index at 65.02%. That is a general benchmark, not a voice one (Vals AI).
Read the full section
- Earlier LLM tests for voice: Daily.co benchmarked LLMs on tool calling, instruction following and grounding over long voice-style conversations (Daily).
- Growth estimates: Sacra estimates Retell grew from about $5.4M revenue in 2024 to $45M by end-2025 and $60M annualized by April 2026 (Sacra).
- Astra's general standing: Astra ranks #5 overall on the Vals Index at 65.02%. That is a general benchmark, not a voice one (Vals AI).
Limitations, safety and contested findings
- Rankings depend on the benchmark: on Cekura's broader leaderboard, Retell is second (75.6%), behind PolyAI (87.8%). The Medicare result covers a single workflow (Retell).
- Production signal vs. benchmark claim: Retell's own guide recommends GPT 4.1, and its pricing page recommends Terra and Sonnet.
- Safety: coverage of Astra's launch gives no model-specific safeguards (Constellation Research).
Read the full section
- Rankings depend on the benchmark: on Cekura's broader leaderboard, Retell is second (75.6%), behind PolyAI (87.8%). The Medicare result covers a single workflow (Retell).
- Production signal vs. benchmark claim: Retell's own guide recommends GPT 4.1, and its pricing page recommends Terra and Sonnet. The Astra result is a benchmark claim, not evidence of production preference.
- Safety: coverage of Astra's launch gives no model-specific safeguards (Constellation Research).
Business and practitioner implications
- Model selection: a benchmark leader is not automatically the right default. Test latency, tool-argument accuracy and per-minute cost on your own call flows.
- Multi-provider leverage: Retell's support for OpenAI, Anthropic and Google keeps switching costs low for its customers (Sacra).
- Revenue per employee: the $1.6M figure suggests small teams using AI coding tools can support large usage-based businesses.
Read the full section
- Model selection: a benchmark leader is not automatically the right default. Test latency, tool-argument accuracy and per-minute cost on your own call flows. Astra's 5× price premium only makes sense for complex, high-value calls.
- Multi-provider leverage: Retell's support for OpenAI, Anthropic and Google keeps switching costs low for its customers (Sacra).
- Revenue per employee: the $1.6M figure suggests small teams using AI coding tools can support large usage-based businesses. Without a headcount, it can't be checked.