AI models and credits
No model is built in. Each step of a call picks a model from one catalog, usage converts into credits, and plans sell a monthly credit pool. Rendered from MODEL_CATALOG in @callwise/pricing.
Callwise does not hard-code an AI model. @callwise/pricing has one typed catalog, MODEL_CATALOG, and every call is a pipeline of steps that each pick a model from it by modality. Whatever models a workspace picks, usage is converted into credits, and plans, industry packs and add-ons sell a monthly pool of credits. The tables on this page are rendered from that catalog.
The steps of a call
| Step | Modality | What it does |
|---|---|---|
listen | transcription | Streaming speech to text for the live call. |
reply | text | The language model that decides what the agent says. |
speak | voice | Text to speech: the agent's voice. |
summarize | text | The summary and classification after the call. |
The catalog's modalities also include image and video, but Callwise has no image or video steps: no call ever uses one.
Models are chosen per agent in Scripts → Models, one per step. The defaults are DEFAULT_STACK; validateStack() refuses an unknown model or one of the wrong modality for its step.
The catalog
Unit prices are estimates, not verified yet
Each entry carries sourceUrl, checkedOn and verified. All 15 of the 15 entries are verified: false with checkedOn: null: the prices below are estimates that nobody has checked against the provider's page yet. Follow the source link before relying on one. The marketing site never shows them; it talks in credits only.
| Step (modality) | Provider | Model | Tier | Unit price | Credits, rest of the stack default |
|---|---|---|---|---|---|
| Listen (speech to text)transcription · default | Deepgram | Deepgram Nova (streaming)deepgram-nova · Streaming speech to text for the live call. | standard | $0.0043 / minuteestimate, not verified · never checked · source | 3.2 per call min |
| Listen (speech to text)transcription | OpenAI | Whisperwhisper-1 · OpenAI speech to text for a recording. The live call still defaults to Deepgram. | standard | $0.006 / minuteestimate, not verified · never checked · source | 3.4 per call min |
| Reply (language model)text · default | Anthropic | Claude Haiku 5.5claude-haiku-5-5 · Fast replies on live calls; summaries and intent. | standard | $1 / 1M input tokens$5 / 1M output tokensestimate, not verified · never checked · source | 3.2 per call min |
| Reply (language model)text | Anthropic | Claude Sonnet 5.5claude-sonnet-5-5 · Harder conversations (insurance, treatment plans); slower and dearer. | premium | $3 / 1M input tokens$15 / 1M output tokensestimate, not verified · never checked · source | 3.9 per call min |
| Reply (language model)text | OpenAI | GPT-5.1gpt-5.1 · OpenAI flagship for harder replies and summaries. | premium | $1.25 / 1M input tokens$10 / 1M output tokensestimate, not verified · never checked · source | 3.4 per call min |
| Reply (language model)text | OpenAI | GPT-5 minigpt-5-mini · OpenAI, faster and cheaper for routine replies. | standard | $0.25 / 1M input tokens$2 / 1M output tokensestimate, not verified · never checked · source | 3 per call min |
| Reply (language model)text | OpenAI | GPT-5 nanogpt-5-nano · OpenAI, cheapest text for short classification. | standard | $0.05 / 1M input tokens$0.4 / 1M output tokensestimate, not verified · never checked · source | 2.9 per call min |
| Reply (language model)text | xAI | Grok 4.7grok-4.7 · xAI flagship. List rate is for prompts under 200K tokens. | premium | $2 / 1M input tokens$6 / 1M output tokensestimate, not verified · never checked · source | 3.4 per call min |
| Reply (language model)text | xAI | Grok 4.3grok-4.3 · xAI, fast replies. List rate is for prompts under 200K tokens. | standard | $1.25 / 1M input tokens$2.5 / 1M output tokensestimate, not verified · never checked · source | 3.2 per call min |
| Reply (language model)text | Z.ai | GLM 5.2 (open weights)glm-5.2 · Open weights. Frontier open model for replies and summaries. | standard | $1.4 / 1M input tokens$4.4 / 1M output tokensestimate, not verified · never checked · source | 3.3 per call min |
| Reply (language model)text | NVIDIA | Nemotron 3 Ultra (open weights)nemotron-3-ultra · Open weights. Large NVIDIA model for replies and summaries. | standard | $0.6 / 1M input tokens$2.4 / 1M output tokensestimate, not verified · never checked · source | 3.1 per call min |
| Reply (language model)text | NVIDIA | Nemotron 3 Nano (open weights)nemotron-3-nano · Open weights. Small, cheap NVIDIA model. | standard | $0.06 / 1M input tokens$0.24 / 1M output tokensestimate, not verified · never checked · source | 2.9 per call min |
| Reply (language model)text | NVIDIA | Nemotron 3.5 Super VL (open weights)nemotron-3.5-super-vl · Open weights. NVIDIA vision-language model, priced per token. | standard | $0.15 / 1M input tokens$0.5 / 1M output tokensestimate, not verified · never checked · source | 2.9 per call min |
| Speak (voice)voice · default | ElevenLabs | ElevenLabs Flashelevenlabs-flash · Most natural voices (Luna, Noah, Ava, Ethan). | premium | $0.015 / minuteestimate, not verified · never checked · source | 3.2 per call min |
| Speak (voice)voice | Cartesia | Cartesia Soniccartesia-sonic · Lower latency and cost; slightly less expressive. | standard | $0.01 / minuteestimate, not verified · never checked · source | 2.7 per call min |
| Summarize and classifytext · default | Anthropic | Claude Haiku 5.5claude-haiku-5-5 · Fast replies on live calls; summaries and intent. | standard | $1 / 1M input tokens$5 / 1M output tokensestimate, not verified · never checked · source | 0.4 per summary |
| Summarize and classifytext | Anthropic | Claude Sonnet 5.5claude-sonnet-5-5 · Harder conversations (insurance, treatment plans); slower and dearer. | premium | $3 / 1M input tokens$15 / 1M output tokensestimate, not verified · never checked · source | 1.2 per summary |
| Summarize and classifytext | OpenAI | GPT-5.1gpt-5.1 · OpenAI flagship for harder replies and summaries. | premium | $1.25 / 1M input tokens$10 / 1M output tokensestimate, not verified · never checked · source | 0.6 per summary |
| Summarize and classifytext | OpenAI | GPT-5 minigpt-5-mini · OpenAI, faster and cheaper for routine replies. | standard | $0.25 / 1M input tokens$2 / 1M output tokensestimate, not verified · never checked · source | 0.2 per summary |
| Summarize and classifytext | OpenAI | GPT-5 nanogpt-5-nano · OpenAI, cheapest text for short classification. | standard | $0.05 / 1M input tokens$0.4 / 1M output tokensestimate, not verified · never checked · source | 0.1 per summary |
| Summarize and classifytext | xAI | Grok 4.7grok-4.7 · xAI flagship. List rate is for prompts under 200K tokens. | premium | $2 / 1M input tokens$6 / 1M output tokensestimate, not verified · never checked · source | 0.7 per summary |
| Summarize and classifytext | xAI | Grok 4.3grok-4.3 · xAI, fast replies. List rate is for prompts under 200K tokens. | standard | $1.25 / 1M input tokens$2.5 / 1M output tokensestimate, not verified · never checked · source | 0.5 per summary |
| Summarize and classifytext | Z.ai | GLM 5.2 (open weights)glm-5.2 · Open weights. Frontier open model for replies and summaries. | standard | $1.4 / 1M input tokens$4.4 / 1M output tokensestimate, not verified · never checked · source | 0.5 per summary |
| Summarize and classifytext | NVIDIA | Nemotron 3 Ultra (open weights)nemotron-3-ultra · Open weights. Large NVIDIA model for replies and summaries. | standard | $0.6 / 1M input tokens$2.4 / 1M output tokensestimate, not verified · never checked · source | 0.3 per summary |
| Summarize and classifytext | NVIDIA | Nemotron 3 Nano (open weights)nemotron-3-nano · Open weights. Small, cheap NVIDIA model. | standard | $0.06 / 1M input tokens$0.24 / 1M output tokensestimate, not verified · never checked · source | 0.1 per summary |
| Summarize and classifytext | NVIDIA | Nemotron 3.5 Super VL (open weights)nemotron-3.5-super-vl · Open weights. NVIDIA vision-language model, priced per token. | standard | $0.15 / 1M input tokens$0.5 / 1M output tokensestimate, not verified · never checked · source | 0.1 per summary |
The last column is what that model would cost in credits if it replaced the default in its step and every other step stayed on the default.
Credits
- One credit is $0.01 of provider cost (
CREDIT_VALUE_USD). Carrier and recording storage per minute are converted into credits too. - A call minute costs
creditsPerCallMinute(stack)credits: listen + reply + speak + carrier for one minute, rounded up to 0.1 credit. With the default models that is 3.2 credits. - The summary after a call costs
creditsPerSummary(stack). - A finished call costs
creditsForCall(billedMinutes, stack): billed minutes (duration in seconds divided by 60, rounded up) × the per-minute rate, plus one summary. minutesForCredits(credits, stack)turns a pool into call minutes; this is what the site's pricing cards show for the default stack, with the range fromstackRange().
Usage per minute (tokens in and out) is an assumption in USAGE_ASSUMPTIONS, to be tuned with real traffic.
Stacks
| Stack | Listen (speech to text) | Reply (language model) | Speak (voice) | Summarize and classify | Credits per call minute | Credits per summary | Minutes per 1,000 credits |
|---|---|---|---|---|---|---|---|
| Default | Deepgram Nova (streaming) | Claude Haiku 5.5 | ElevenLabs Flash | Claude Haiku 5.5 | 3.2 | 0.4 | 300 |
| Cheapest | Deepgram Nova (streaming) | Nemotron 3 Nano (open weights) | Cartesia Sonic | Nemotron 3 Nano (open weights) | 2.4 | 0.1 | 410 |
| Dearest | Whisper | Claude Sonnet 5.5 | ElevenLabs Flash | Claude Sonnet 5.5 | 4 | 1.2 | 227 |
The monthly pool
- Every plan and pack has
creditsper month, shared by every module and every model (Pricing). Credit add-ons (extraCredits) add to it. - Credits do not roll over. Each month starts with the full pool; unused credits are gone at the end of the month.
- At 80% and 100% of the pool, owners get an alert.
- When the pool is used up, the agent stops picking up and calls forward to the fallback numbers, so a person answers. No call is left unanswered. The agent only answers while at least one call minute of its stack is left (
decideInboundCall()). - Owners can opt in to overage instead, at $0.03 per credit (
OVERAGE_PER_CREDIT), and the agent keeps answering. - Text tests of draft scripts are free and use no credits.