AI Model & Voice
Specification Comparison
Published pricing and latency for the models and voice engines used to build AI agents, each figure quoted from the vendor's own page and linked to it.
What this page is
A dated snapshot of what AI model and voice vendors publish about their own products: token pricing and, where the vendor states one, latency. PxlPeak has not run a benchmark and does not present these as measurements. Vendor-published latency is a best-case figure produced on the vendor's own infrastructure.
Key Takeaways
Model pricing
Published list price per 1M tokens. Capability scores are deliberately absent PxlPeak runs no evaluation, and any “logic” or “coding” score shown without a named benchmark behind it is an invented number.
| Model | Input / 1M | Output / 1M |
|---|---|---|
Claude Fable 5.1 Anthropic Replaces Fable 5, which stays available as a legacy model at the same price. platform.claude.com— source for Claude Fable 5.1 pricing, opens in a new tab | $10.00 | $50.00 |
Claude Opus 5.5 Anthropic Replaces Opus 5, which stays available as a legacy model at $5.00 / $25.00. platform.claude.com— source for Claude Opus 5.5 pricing, opens in a new tab | $4.00 | $20.00 |
Claude Sonnet 5.5 Anthropic Legacy Sonnet 5 is also $2.00 / $10.00: Anthropic made its introductory rate permanent and cancelled the rise to $3.00 / $15.00 it had scheduled for 2026-09-01. platform.claude.com— source for Claude Sonnet 5.5 pricing, opens in a new tab | $2.00 | $10.00 |
Claude Haiku 5.5 Anthropic Prompts up to 100,000 tokens; $0.50 / $2.50 above that. Replaces Haiku 4.5, which stays available as a legacy model at $1.00 / $5.00. platform.claude.com— source for Claude Haiku 5.5 pricing, opens in a new tab | $0.10 | $0.50 |
gpt-6-astra OpenAI Standard tier, short context, as are the other GPT-6 rows. developers.openai.com— source for gpt-6-astra pricing, opens in a new tab | $10.00 | $50.00 |
gpt-6.1-sol OpenAI developers.openai.com— source for gpt-6.1-sol pricing, opens in a new tab | $2.00 | $10.00 |
gpt-6-luna OpenAI developers.openai.com— source for gpt-6-luna pricing, opens in a new tab | $0.10 | $0.50 |
GPT-5.6-sol OpenAI Standard tier, short context; long-context prompts are billed higher. Promotional pricing the vendor says runs at least through November 21, 2026. Was $5.00 / $30.00 when read on 2026-08-03. developers.openai.com— source for GPT-5.6-sol pricing, opens in a new tab | $4.00 | $20.00 |
GPT-5.6-terra OpenAI developers.openai.com— source for GPT-5.6-terra pricing, opens in a new tab | $2.00 | $12.00 |
GPT-5.6-luna OpenAI developers.openai.com— source for GPT-5.6-luna pricing, opens in a new tab | $0.20 | $1.20 |
Gemini 3.1 Pro Preview Google Prompts at or under 200k tokens; $4.00 / $18.00 above that. ai.google.dev— source for Gemini 3.1 Pro Preview pricing, opens in a new tab | $2.00 | $12.00 |
Gemini 3.8 Flash Google Price runs through 2026-12-31; $1.50 / $7.50 from 2027-01-01. ai.google.dev— source for Gemini 3.8 Flash pricing, opens in a new tab | $0.75 | $3.75 |
Gemini 3.5 Flash-Lite Google ai.google.dev— source for Gemini 3.5 Flash-Lite pricing, opens in a new tab | $0.30 | $2.50 |
DeepSeek-V4.1-Flash DeepSeek API name deepseek-flash. Off-peak cache-miss input / off-peak output; peak doubles both ($0.30 / $1.20). Peak is 01:00-04:00 and 06:00-10:00 UTC, Monday to Friday. Cache-hit input is $0.003 off-peak. Replaces DeepSeek-V4-flash, which was $0.22 / $0.66 off-peak on 2026-09-02; the old name is still accepted and served by V4.1-Flash. api-docs.deepseek.com— source for DeepSeek-V4.1-Flash pricing, opens in a new tab | $0.15 | $0.60 |
DeepSeek-V4-pro DeepSeek Off-peak cache-miss input / off-peak output. Peak doubles both ($1.32 / $3.96). Cache-hit input is $0.022 off-peak. Was a flat $0.435 / $0.87 on 2026-08-03; the peak-hour pricing this row previously flagged as announced is now live. api-docs.deepseek.com— source for DeepSeek-V4-pro pricing, opens in a new tab | $0.66 | $1.98 |
Published voice latency
The figures measure different things. A text-to-speech model latency covers one component; a platform's end-to-end response time covers speech recognition, model inference, speech synthesis and network round trips together. A component number will always look faster than an end-to-end number, and ranking them against each other would be meaningless.
| Product | What the figure measures | Vendor states |
|---|---|---|
Flash / Turbo TTS ElevenLabs elevenlabs.io— source for Flash / Turbo TTS latency, opens in a new tab | Text-to-speech model latency only | ~75ms |
Multilingual v2/v3 TTS ElevenLabs elevenlabs.io— source for Multilingual v2/v3 TTS latency, opens in a new tab | Text-to-speech model latency only | ~250–300ms |
Scribe v2 Realtime STT ElevenLabs elevenlabs.io— source for Scribe v2 Realtime STT latency, opens in a new tab | Speech-to-text streaming latency only | ~150ms |
Aura-2 TTS Deepgram deepgram.com— source for Aura-2 TTS latency, opens in a new tab | Streaming text-to-speech latency only | sub-200ms |
Voice agent platform Vapi docs.vapi.ai— source for Voice agent platform latency, opens in a new tab | Full end-to-end response time, one conversational turn | sub-600ms |
Retell AI:Publishes per-minute pricing and a cost calculator, but no end-to-end latency figure was found on its pricing page.retellai.com— source for Retell AI pricing, opens in a new tab
How this page was compiled
- What it is
- Latency figures each vendor publishes about its own product, quoted as published and dated.
- What it is not
- Not a benchmark. PxlPeak did not run a controlled test, did not measure these products, and has no independent timing data. Vendor-published latency is a best-case marketing figure measured on the vendor's own infrastructure.
- Why nothing is ranked
- Ranking requires a common measurement. These are vendor claims about different quantities, so there is no honest ordering to publish.
Need real numbers for your stack?
Vendor specs are best-case. We can run head-to-head tests on your own data and call volume to find the most cost-efficient combination.
Ready?
Put AI to Work for You.
Call now and talk to Aria, our AI strategist, or book a free 30-minute assessment.
Aria picks up instantly · 24/7 · Free assessment · 30-day guarantee