Latency, throughput and reliability for Telnyx Inference, Voice AI, Speech and Web Search, measured against the providers you would otherwise be using. Same prompts, same profiles, same day. We publish the runs we lose.
Pooled across four workloads at concurrency 1 and 10. Fresh prompts, streaming, reasoning effort low. Lower is better.
Telnyx
4.92s
Modal Shared
8.75s
240 of 240 requests completed on both providers. Modal Shared had the lower p95 time to first visible answer in seven of eight cells. Covers Modal Shared only, not Modal Dedicated or Stateful Actors. Full report and all eight cells ↗
14
Published benchmark reports
4
Products under measurement
p95
Reported on every latency, not just the median
3 of 14
Reports where a rival wins the headline metric
Results
Benchmarks by product.
Pick a product. Every result set below names its metric, its run date, its workload profile and the number of attempts behind it. Bold marks the best value in a column, whoever it belongs to.
Product
7 result sets · latest run 15 Oct 2026 · identical prompts across providers, plus multi-turn and one Telnyx-only model comparison
InferenceTelnyx run15 Oct 2026
GLM 5.3 Flash on Telnyx Inference against Modal Shared
Metric end to end latency, lower is better, and output speed, higher is better. Model zai-org/GLM-5.3-Flash on both, reasoning effort low, streaming. Profiles four, not six: about 1k, 10k and 100k input with 256 output, and 10k input with 1,024 output, each at concurrency 1 and 10. Attempts 240 per provider, fresh prompts, 30 per cell after three discarded warm ups. All 240 completed on both. Pricing published token rates 70 percent lower on Telnyx: $0.135 input and $0.450 output per 1M, against $0.450 and $1.500. Scope Modal Shared only, not Modal Dedicated or Stateful Actors.
GLM 5.3 Flash on Telnyx Inference against Modal Shared
Provider
End to end p50seconds, lower better
End to end p95seconds, lower better
Output speed p50tokens per second, higher better
Successful requestsof 240
Telnyx
4.92
11.91
138.1
240
Modal Shared
8.75
20.62
66.2
240
InferenceTelnyx run08 Oct 2026
Multi-turn task completion across three Telnyx-hosted models
Comparison Telnyx only. Three models on the same Telnyx Inference API, no other provider in this table. Scenarios 15 four-turn conversations per model across developer, business, planning, editing and structured output workflows. Scoring two blind judges, DeepSeek V4.1 Flash and MiniMax M3, for 30 completion decisions per model, plus 28 deterministic checks on the final answers. Settings reasoning effort low, temperature 0, 4,096 token ceiling. Reliability all three models completed every request with no HTTP errors, timeouts or truncated responses.
Multi-turn task completion across three Telnyx-hosted models
Model
Judge completionof 30, higher better
Objective checksof 28, higher better
Conversation p50seconds, lower better
Conversation p95seconds, lower better
Kimi K3
305.00 of 5 on corrections, both judges
23
9.36
55.71
GLM 5.3
27
23
16.38
53.19
GLM 5.3 Flash
27
22
8.35
31.27
InferenceTelnyx run07 Oct 2026
Multi-turn conversations across three models and four providers
Comparison Telnyx, Baseten, Together AI and Fireworks, all called through one orchestrator with explicit conversation history in every request. Scenarios 15 fixed multi-turn workflows across developer, business and structured output tasks, three repeats each: 45 attempted conversations per provider-model pair, 540 in total. Metric total conversation response time, the sum of model response time across the workflow, excluding simulated user pauses. Settings reasoning effort low, temperature 0, 4,096 token ceiling per turn. Scoring one blind judge, DeepSeek V4.1 Flash, plus deterministic checks. Mean blind quality scores ranged from 4.79 to 5.00 with a different leader per model family. Reliability of 135 attempted conversations per provider: Telnyx and Baseten completed 134, Fireworks 133, Together AI 129. Failed conversations stay excluded from the latency percentiles.
Multi-turn conversations across three models and four providers
Model
Telnyx p50seconds conversation, lower better
Next fastest p50seconds conversation, lower better
Next fastest observed p95seconds conversation, lower better
GLM 5.3 Flash
8.26
11.0025 percent higher
35.62
42.7217 percent higher
GLM 5.3
13.48
15.8515 percent higher
41.33
50.65Baseten. Together AI observed 181.20, Fireworks 118.83
Kimi K3
10.49
15.9534 percent higher
46.40
49.667 percent higher
InferenceTelnyx run18 Sep 2026
GLM 5.3 end to end latency across four providers
Metric end to end, streaming, lower is better. Profile shown 88.7k input / 1,200 output. Prompts 60 distinct, identical across providers. Settings temperature 0, reasoning effort high. Attempts 240, of which 230 completed. Telnyx was called from a stateful actor in us-east-2, the other three over the public internet.
GLM 5.3 end to end latency across four providers
Provider
End to end p50seconds, lower better
End to end p95seconds, lower better
Telnyx
9.87
12.80
Together AI
14.67
69.64
Fireworks AI
20.61
26.84
Baseten
21.81
23.18
InferenceTelnyx run02 Sep 2026
GLM 5.3 Flash throughput and completion rate
Metric output throughput, higher is better, and how many of 60 runs returned at all. Settings streaming, temperature 0. Completion rate is reported because a provider that refuses one run in four is not comparable on speed alone.
GLM 5.3 Flash throughput and completion rate
Provider
Throughput p50tokens per second, higher better
Runs completedof 60
Telnyx
196.4
60
Baseten
190.116 runs returned HTTP 429, rate limited
44
Fireworks AI
127.9
60
Together AI
102.4
60
InferenceTelnyx run25 Jun 2026
GLM 5.2, where first token and finished answer disagree
Metric both, lower is better. Attempts 240 completed. Note Telnyx served FP8, the others served the unquantized weights; this is disclosed because it affects both columns. Reasoning was enabled, so the TTFT column records the first token of any kind, which may be reasoning rather than answer text. TTFA is not in this run.
GLM 5.2, where first token and finished answer disagree
Provider
TTFT p50seconds, lower better, reasoning included
End to end p50seconds, lower better
Telnyx
1.12zai-org/GLM-5.2-FP8
6.00
Baseten
0.76
6.19
Together AI
0.79
7.49
Fireworks AI
1.28
14.98
InferenceTelnyx run28 Jul 2026
The same workload, called from inside the region
Metric both, lower is better. What changed only the caller. Identical model, identical prompts, moved from the public internet to a Telnyx stateful actor colocated with the GPUs. Caveat, stated in the source the comparison uses the earlier provider baselines rather than fresh same day reruns.
The same workload, called from inside the region
Configuration
TTFT p50seconds, lower better, reasoning included
End to end p50seconds, lower better
GLM 5.2, called over the public internet
1.12
6.00
GLM 5.2, called from a stateful actor in region
0.76
4.54
MiniMax M3, called over the public internet
1.87
7.56
MiniMax M3, called from a stateful actor in region
0.61
7.27
What the numbers say
Why the numbers look like this.
Four claims. Each one is only here because a table above pays for it, and each links to the run it came from.
01
We own the infrastructure.
Our GPUs sit on our own network, not rented capacity resold to whoever bid. You cannot see that in a median. You see it in the tail, when everybody else's tenants show up at once.
12.80s observed p95 against 69.64s
GLM 5.3, 88.7k input tokens. Telnyx recorded a 12.80s observed p95, compared with 69.64s for Together AI in this sample.
Compute in region is not a data residency checkbox. It is distance, and distance is most of what a first token costs. Move the caller next to the GPU and nothing else has to change.
TTFT down 67.3 percent
MiniMax M3, 1.87s to 0.61s. Same model, same prompts, called from a stateful actor beside the GPUs instead of across the internet.
Telephony, speech to text, the model and text to speech on separate vendors means four networks between a caller and an answer. We measure the stack, because the stack is what the caller hears.
Above the Pareto frontier
ServiceNow Research, 150 live conversations. The first hosted voice API on the leaderboard to beat the Pipecat agents on quality at equal accuracy.
04
We publish the runs we lose.
A vendor benchmark that a vendor always wins tells you the vendor chose the workload. Three of the fourteen reports on this page have somebody else at the top of the headline column, and all three are still here.
Brave 561ms, Telnyx 1,027ms
Web search latency, 20 identical queries. Baseten also shows first activity 360ms sooner than us on GLM 5.2.
The same six rules govern every table on this page. Where a run breaks one of them, the table says so in its own header rather than in a footnote nobody reads.
01
Identical prompts, sent the same day
Every provider in a table gets the same prompt set inside the same session. A number carried over from an earlier run is labeled as a baseline, not presented as a comparison.
02
Six workload profiles, never one
1k, 10k and roughly 100k input tokens, against short and long outputs. A single profile is how a provider picks the shape of the workload it happens to be good at.
03
Distributions, not averages
p50 and p95 together, always. A median tells you what usually happens. p95 tells you what your support queue will look like.
04
Failed calls are counted, not dropped
A rate limited request is a result. Where a provider returned fewer runs than it was asked for, the table shows the completion rate beside the speed, because excluding the failures flatters whoever failed.
05
Asymmetries are disclosed in the table
Quantization, model identifier, and where the request originated from. Our GLM 5.2 run served FP8 against unquantized weights elsewhere, and our traffic came from a stateful actor in region. Both are stated where the numbers are, not at the bottom of the page.
06
The runs we lose stay published
A result is not withdrawn because a competitor came out ahead. Three of the reports above have somebody else at the top of the headline column.
Check every number at its source
Each result we ran links to the full report behind it. Our test harness is internal, so the published reports are the record. To see how your own traffic compares, send your prompts to the same endpoints from a free account.
