Skip to main content
Resources

Changelog

Latest updates, new features, and improvements to the CallMissed API.

November 2026

CRM integration API (2026-11-17)

  • Create-or-update. POST /api/v1/contacts/upsert, POST /api/v1/companies/upsert and POST /api/v1/crm/leads/upsert create a record or update the one that already carries your id, so a sync can be re-run safely. Contacts match on your external id, phone or email; companies on your external id, domain or name; leads on external_ref. The response says which happened (created, matched_on). See Contacts, Companies and Leads.
  • Look up by your own id. GET /api/v1/contacts?external_id=hubspot:123 and the same on companies; GET /api/v1/crm/leads?external_ref=… for leads.
  • Changed since. updated_after on the contact, company, lead, deal and task lists returns the records changed after a time, oldest change first.
  • CRM change webhooks. Eighteen new events, contact|company|lead|deal|task|note × created|updated|deleted, sent for every change whoever made it, with a source field and the full record. See Webhooks.
  • Call transcript by call id. GET /api/v1/telephony/calls/{call_id}/transcript (format=json|txt|srt, scope telephony:read). See Telephony.
  • New guide: Connect your CRM.

Legacy voice WebSocket removed (2026-11-11)

  • wss://api.callmissed.com/ws/voice-agent is removed. For a voice agent over a plain WebSocket, use the Managed Voice Agent at wss://api.callmissed.com/v2/voice/agent: it sends a ConversationText event for each finished caller turn. For WebRTC, use voice sessions. The Real-time STT page, which documented the removed socket, now redirects to the Managed Voice Agent. File transcription on Speech to Text is unchanged.

ElevenLabs voices and Scribe transcription (2026-11-07)

  • Nine ElevenLabs text-to-speech models under their own IDs on POST /v1/audio/speech and in voice agents and voice sessions (tts_model): eleven_v4, eleven_v4_turbo, eleven_v3, eleven_v3_conversational, eleven_multilingual_v2, eleven_flash_v2_5, eleven_turbo_v2_5, eleven_flash_v2 and eleven_turbo_v2. v4 speaks 91 languages and v3 85, both including Hindi, Bengali, Tamil, Telugu, Marathi, Gujarati, Kannada, Malayalam, Punjabi, Odia and Urdu; Flash v2.5 speaks 32 including Hindi and Tamil. eleven_v4, eleven_v3 and eleven_multilingual_v2 are $0.8333 per 10K characters, the rest $0.4167. See Text to Speech.
  • One shared voice library. Pass a premade name (sarah default, george, brian, …) or any ElevenLabs voice ID, including your cloned voices, as voice. language takes a base code (hi) and enforces the language on every model except eleven_multilingual_v2; speed clamps to 0.7–1.2; mp3, wav, pcm, opus, mulaw and alaw output. v3 and v4 perform audio tags such as [laughs]. See Voices.
  • ElevenLabs Scribe speech-to-text. scribe_v2 and scribe_v2_medical ($0.2292 per hour) transcribe files in 100 languages with automatic language detection on POST /v1/audio/transcriptions; scribe_v2_realtime ($0.4063 per hour) is the streaming model for voice agents (stt_model). See Speech to Text.
  • Paid plans only. The catalogue is now 159 models: 50 speech-to-text and 25 text-to-speech.

Kimi K3.1 Pro (2026-11-06)

  • kimi-k3.1-pro (Moonshot AI) is Moonshot's most capable open-weight model, natively multimodal and built for agentic work: coding, 3D creation, financial research and tool use. 1,048,576-token context window, 131,072 tokens of default max output, image input, tool calling, JSON mode (response_format: {"type": "json_object"}) and streaming. Output is text.
  • It thinks by default. reasoning_effort accepts "none" (thinking off), "low", "high" and "max"; "minimal" runs as "low", "medium" as "high" and "xhigh" as "max". See the reasoning_effort matrix.
  • Priced at $2.00 input / $10.00 output per 1M tokens, with cached input at $0.20. Paid plans only; it is not in the free tier.
  • Images go in as base64 data URIs; we fetch and inline https:// image URLs for you.
  • The catalogue is now 147 models: 73 LLMs.

Voice calls keep going through rate limits (2026-11-02)

  • When a voice agent's model cannot answer a turn, for example because of a rate limit, sarvam-105b answers it, and if that fails too, gpt-4o does. Before, a call whose model and sarvam-105b both failed could end. A turn sarvam-105b answered is billed at its own rate; a turn gpt-4o answered is billed at your agent's model rate, and usage shows gpt-4o as the model that answered. On an agent with hosting: "india", gpt-4o is used only while it is processed in India. See Voice agent.
  • When many calls start at the same moment, their speech connections are opened a moment apart, so fewer are refused for opening too fast. A refused one moves to the backup speech provider: listening at once, and a reply after one quick retry.

Numbers nothing answers play a short message (2026-11-02)

  • An inbound call to a number with no agent, or whose agent is switched off, and no call menu, call flow, call queue or agent squad is now answered with "Sorry, this number isn't taking calls right now. Please try again later." and then ends. Before, a number with no agent was answered by a generic assistant, and one whose agent was switched off was not picked up. The message plays in the number's own language setting when it has one, else English.
  • The call is billed like any other short inbound call: carriage for its length, plus the ₹1 a minute platform fee at its 30-second minimum. See Telephony and Bots.

GLM 5.3 Flash and DeepSeek V4.1 Flash 1017 (2026-11-01)

  • glm-5.3-flash (Z.ai) is a fast, low-cost GLM for coding agents and tool use, with a 1,048,576-token context window, image input, tool calling, structured output and streaming. It always reasons: reasoning_effort takes "low", "high" or "max", with "high" as the default; "none" and "minimal" map to "low", "medium" to "high" and "xhigh" to "max". In our tests with "low" the first token arrived in about 0.4 to 0.5s. $0.229 per 1M input tokens, $0.760 per 1M output tokens ($0.052 cached input). Paid plans only.
  • deepseek-v4.1-flash-1017 (DeepSeek) is the October 17 snapshot of DeepSeek V4.1 Flash, with a 1,048,576-token context window, image input, tool calling (including tool_choice: "required") and streaming. It reasons at "high" by default and reasoning_effort: "none" switches thinking off (first token in about 0.25 to 0.6s in our tests). $0.427 per 1M input tokens, $1.771 per 1M output tokens ($0.0104 cached input), the same as deepseek-v4.1-flash. Paid plans only.
  • Images go in as base64 data URIs; we fetch and inline https:// image URLs for you. When you omit max_tokens we send 16,384, since reasoning tokens count toward the limit.
  • The catalogue is now 146 models: 72 LLMs.

October 2026

Deepgram Think-Fast and phoneLLM mini chat models (2026-10-31)

  • deepgram-think-fast (Deepgram) and deepgram-phonellm-mini (Pipecat) are early-access chat models for voice agents, on POST /v1/chat/completions and as the llm_model of a voice agent. 32,768-token context, tool calling and streaming, text only. deepgram-think-fast supports reasoning_effort (none default, minimal, low, medium, high); deepgram-phonellm-mini does not reason. Both are $0.3646 per 1M input tokens and $0.9375 per 1M output tokens, paid plans only, and not a default anywhere.
  • Training on your data. Conversations on deepgram-think-fast are used by the provider to improve its models; the opt-out does not apply to it. deepgram-phonellm-mini is excluded from the provider's model training.
  • Processed outside India. India-only keys and accounts get region_unavailable, and voice agents with hosting: "india" cannot select them. tool_choice takes only "auto" or "none", and response_format returns 400 unsupported_parameter. See Models.
  • Languages. In our tests deepgram-think-fast was reliable in English, Hindi, Marathi, Telugu and Kannada, and deepgram-phonellm-mini only in English and Hindi. Validating a voice agent's config now flags other languages (an error for deepgram-phonellm-mini, a warning for deepgram-think-fast).
  • The catalogue is now 144 models: 70 LLMs.

Gnani Prisma v3, and Prisma v2.5 retired (2026-10-30)

  • gnani-prisma-v3.0 is Gnani's newest India-first telephony speech-to-text model: English plus 9 Indian languages (hi-IN, bn-IN, ta-IN, te-IN, kn-IN, ml-IN, mr-IN, gu-IN, pa-IN, en-IN), automatic language detection with language: "indic-auto", mixed-language speech and keyterm boosting. It runs real-time and batch. Price is $0.2813 / hour, the same as v2.5. Paid plans only.
  • POST /v1/audio/transcriptions. language takes one of the 10 codes, a comma-separated list of up to 4 of them, or indic-auto; a base code such as hi is read as hi-IN. Audio is limited to 30 seconds per request (v2.5 allowed 60); longer audio returns 400 audio_too_long.
  • Voice agents and sessions. On gnani-prisma-v3.0, language: "multi" follows a multilingual caller (it detects the language, so the voice follows), and per-agent stt_keyterms are applied. It is processed in India, so it stays available in India mode.
  • gnani-prisma-v2.5 is retired. The ID keeps working on the API, in voice agents and in saved agents. It is answered by gnani-prisma-v3.0, the response keeps the ID you sent, it keeps its price ($0.2813 / hour) and it is no longer listed by GET /api/v1/models. The same 30-second per-request limit now applies. Agents that set the old code-mixed codes en-hi-in-cm or en-hi-IN-latn keep working. See Retired and deprecated models.
  • The catalogue size is unchanged: one model added, one retired.

Deepgram updates (2026-10-30)

  • Managed voice-agent models. deepgram-voice-anthropic-claude-sonnet-5-5 (Advanced, $0.1698/min), deepgram-voice-anthropic-claude-haiku-5-5 (Standard, $0.07813/min), and the early-access deepgram-voice-deepgram-think-fast and deepgram-voice-deepgram-phonellm-mini-alpha-2-fp8 (Standard). Conversations on deepgram-voice-deepgram-think-fast are used by the provider to improve its models.
  • Speech to text. New deepgram-nova-3-pharma (English only) for drug names. deepgram-nova-3-medical now accepts language=multi (10 languages). POST /v1/audio/transcriptions takes a repeatable language_hint field for deepgram-nova-3 with language=multi.
  • Pricing by mode. deepgram-nova-3, -medical and -pharma are $0.325 / hour for files and $0.4792 / hour on live voice sessions, or $0.5729 / hour when the session's language is multi (-pharma is English only, so always $0.4792) (previously $0.3021 / hour for both). deepgram-flux-general-en is $0.4792 / hour (previously $0.4063).
  • Flux TTS on /v1/audio/speech. deepgram-flux-tts ($0.4688 / 10K characters) now has 266 voices in 42 languages, including Hindi, Tamil, Telugu, Kannada, Marathi, Gujarati, Bengali and Urdu (32 languages are early access), and supports stream: true. Input is capped at 2000 characters, and at 400 without streaming.
  • The catalogue is now 142 models: 47 speech-to-text and 16 text-to-speech.

Preview: Microsoft AI speech models (2026-10-30)

  • Text to speech. mai-voice-2.1 ($0.2292 / 10K characters) is Microsoft AI's most expressive voice model, for narration and long-form audio. mai-voice-2.1-flash ($0.1563 / 10K characters) is built for low latency and real-time voice agents. Both have 97 voices across 23 languages, including Hindi (hi-in-arjun, hi-in-dhruv, hi-in-grant, hi-in-harper, hi-in-kavya, hi-in-priya) and Indian English (en-in-dhruv, en-in-priya). Plain text only, speed 0.5 to 2.0. They work on POST /v1/audio/speech and in voice agents and voice sessions (tts_model). See Voices.
  • Speech to text. mai-transcribe-2 ($0.375 / hour) transcribes files in 60 languages with automatic language detection and Hinglish code-switching, up to 2 hours per file, on POST /v1/audio/transcriptions. It is not available in voice sessions or translation. See Speech to Text.
  • All three are preview. Microsoft ships them with no SLA and may change or retire them at short notice. If one is retired, its ID keeps working and is answered by a named successor, as in Retired and deprecated models, so an agent that uses it keeps working but may sound different. Use generally available models for production-critical agents. They are paid-plan only, are never a default, are not processed in India, and the voices cannot be used in tts_fallbacks. GET /api/v1/models marks them with preview: true.
  • The catalogue is now 137 models: 46 speech-to-text and 16 text-to-speech.

API behaviour fixes

  • Ending a voice session made by another key now needs telephony:read. DELETE /v1/voice/sessions/{id} follows the same rule as reading a session: a key can always end the sessions it created, including calls it placed with POST /api/v1/telephony/calls, and needs telephony:read for sessions the console or another key created. Without it the call returns 403.
  • Key budget refusals say so. When an API key's own spending budget, not the account balance, cannot cover a request, the response is now budget_exceeded instead of insufficient_credits.
  • n > 1 near the time limit. A non-streaming chat completion that reaches the request time limit after some of its n choices are finished now returns those choices, billed as delivered, instead of an error.
  • Audio of unknown length. Speech-to-text uploads whose duration cannot be read from the file are no longer refused up front for a worst-case estimate; they are billed on the duration actually processed.

Hosted in India voice agents

  • New hosting setting. Set hosting: "india" in a voice agent's config, or Where calls run → Hosted in India in the console's Voice card. Speech recognition, the model, the voice, their backups, knowledge-base and memory search, and post-call analysis, notes, scoring and contact memory then all run in India.
  • Models you can choose in India mode. Speech to text: saaras:v4, gnani-prisma-v2.5, microsoft-speech-stt and the Deepgram Nova and Flux models (Whisper excluded). Models: gemma-4-31b, sarvam-105b, sarvam-105b-conversations, glm-5.3 and deepseek-v4.1-flash. Voices: bulbul:v3, bulbul:v4-flash, gnani-timbre-v3.0, microsoft-neural-tts and deepgram-aura-2 / deepgram-aura-1. Of the voice plans, only t1 qualifies. Speech-to-speech models are not available.
  • Saving is checked. Saving an India-hosted agent with a model outside this list returns 422 naming the model. GET /api/v1/bots/config-schema lists the allowed models under india_hosting, and each plan carries hosted_in_india.
  • hosting: "global" runs gemma-4-31b on a global host first, with India as its fallback; nothing else changes. Without hosting, agents behave as before. Prices are the same in every mode.
  • hosted_in_india in the model catalogue now means processed in India on every request. Models served from a global pool are no longer flagged, even when their endpoint is in India.

New voice-agent models: Microsoft Speech and Microsoft Neural voices

  • Speech to text. microsoft-speech-stt ($1.0417 / hour) is Microsoft's real-time streaming recognizer, processed in India. It hears 81 languages (148 locales), including Hindi, Bengali, Tamil, Telugu, Marathi, Gujarati, Kannada, Malayalam, Punjabi, Odia, Assamese, Urdu, Bhojpuri and Indian English, one language per call.
  • Text to speech. microsoft-neural-tts ($0.1563 / 10K characters), processed in India, has 57 voices across 13 languages, with plain-text input and a speaking-speed control (tts_speed, 0.5 to 2.0).
  • Voice sessions only. Both run inside voice agents and voice sessions. POST /v1/audio/transcriptions and POST /v1/audio/speech return 400 for them. See Voices and Speech to Text.
  • The catalogue is now 134 models: 45 speech-to-text and 14 text-to-speech.

New voice-agent models: AssemblyAI, Inworld, Rime, Fish Audio, Gradium and Deepgram Flux TTS

  • Speech to text. assemblyai-universal-3.6-pro ($0.4688 / hour) is AssemblyAI's newest streaming model with built-in turn detection and Hindi, Marathi, Urdu and Indian English among its languages. inworld-stt-1 ($0.1563 / hour) is a low-cost English-only model.
  • Text to speech. deepgram-flux-tts ($0.4688 / 10K characters, 16 English voices), rime-coda ($0.5208, 8 languages including Hindi), inworld-tts-2 ($0.2604) and inworld-tts-2-flash ($0.1563) with 94 languages including 11 Indian languages, fish-audio-drama-3 ($0.5208, 13 languages) and gradium-tts-beta ($0.50, 5 languages).
  • Voice sessions only. All eight run inside voice agents and voice sessions. POST /v1/audio/transcriptions and POST /v1/audio/speech return 400 for them. See Voices and Speech to Text.
  • The catalogue is now 132 models: 44 speech-to-text and 13 text-to-speech.

Models moved, retired and deprecated

  • Same IDs, new hosting. kimi-k2.5, kimi-k2.6, kimi-k2.7-code, gpt-oss-120b, DeepSeek-V4-Pro and DeepSeek-V4-Flash keep their IDs and prices. kimi-k2.5 now has a 262,144-token context window and gpt-oss-120b 131,072. DeepSeek-V4-Pro and DeepSeek-V4-Flash take 1,000,000 input tokens.
  • Price updated 2026-10-17. DeepSeek-V4-Pro input is now $1.813 per 1M tokens (was $1.375) and cached input $0.151 (was $0.04583); output is unchanged at $4.125. DeepSeek-V4-Flash cached input is now $0.02917 per 1M (was $0.01458); input and output are unchanged.
  • Retired IDs keep working. glm-4.7-flash, glm-5.2, nemotron-3-super, gemma-4-26b-a4b-it, mistral-small-3.1, flux-2-klein-9b, lucid-origin, phoenix-1.0, sdxl-lightning, dreamshaper-8-lcm, aura-2-en, aura-2-es, melotts, nova-3 and whisper-large-v3-turbo are no longer listed. Requests that name them are answered by a successor model and the response keeps the ID you sent. Most keep their price; gemma-4-26b-a4b-it, mistral-small-3.1, sdxl-lightning, dreamshaper-8-lcm, melotts and whisper-large-v3-turbo now bill at their successor's price. See Retired and deprecated models for each successor and price.
  • Free plan. The retired LLM, speech-to-text and text-to-speech IDs stay free. The free plan no longer includes an image model: a free key that requests an image gets 403 model_not_available. The default image model for POST /v1/images/generations and social posts is now flux-2-pro. The free tier is now 11 listed models and the catalogue is 124.
  • Deprecated, sunset 2026-11-16: gpt-4o, gpt-4.1, gpt-5.5, gpt-5.6-sol and gpt-5.6-terra (successor gpt-6-sol); gpt-5-mini and gpt-5.6-luna (gpt-6-luna); gemini-3.5-flash, gemini-3.6-flash and gemini-3.7-flash (gemini-3.8-flash); gemini-3.1-flash-lite (gemini-3.5-flash-lite); gpt-realtime, gpt-realtime-1.5 and gpt-realtime-2 (gpt-realtime-2.1); gpt-realtime-mini (gpt-realtime-2.1-mini); gpt-image-1.5 and gpt-image-2 (gpt-image-2.5-flare); flux-1.1-pro (flux-2-pro); and the older Deepgram speech-to-text IDs (deepgram-nova, deepgram-nova-2, deepgram-enhanced, deepgram-base to deepgram-nova-3; their medical variants to deepgram-nova-3-medical; their phonecall, meeting, finance, conversationalai, voicemail and video variants to the matching deepgram-nova-2-* model). The full table is in Retired and deprecated models. Until then they are served as themselves; responses carry Deprecation, Sunset and Link headers, and the catalogue marks them deprecated: true with sunset_date and successor. From 2026-11-16 a request that names one is answered by and billed as its successor.
  • aura-2-en and aura-2-es keep their voices but now accept up to 2,000 characters per request, the limit of the model that answers them.

Gnani Timbre v3, and older speech models retired

  • gnani-timbre-v3.0 replaces gnani-timbre-v2.0. It has 12 voices (default Shivani), every voice speaks all 10 Indian languages plus Hinglish, it streams, and the text accepts expressive audio tags such as <laugh> and <pauses>. Price is unchanged at $0.27 per 10K characters. See Text to Speech. Requests naming gnani-timbre-v2.0 now return the unknown-model error.
  • bulbul:v2 and saaras:v3 are retired. Use bulbul:v3 or bulbul:v4-flash for speech, and saaras:v4 for transcription (same languages, modes and price). saaras:v4 is now the default when model is omitted. Voice sessions that still name saaras:v3 or saarika:v2.5 are served as saaras:v4.
  • The catalogue is now 139 models and the free tier is 26.

New models: DeepSeek V4.1 Flash and Bulbul V4 Flash

  • deepseek-v4.1-flash — a 1M-token reasoning model with tool calling and text-only input. It always reasons: reasoning_effort takes "low", "high" or "max" (the default). Pricing per 1M tokens is $0.427 in / $1.771 out ($0.0104 cached input). Paid plans only.
  • bulbul:v4-flash — low-latency Indic text to speech with 222 persona voices (<name>_<language>_<style>, default shubh_enhi_ads) across 9 Indian languages and Hinglish, at $0.2083 per 10K characters. It is on the free tier. Input is capped at 2,500 characters. ml-IN and od-IN are not available on it; use bulbul:v3.
  • The catalogue is now 140 models and the free tier is 27.

flux-2-dev removed

  • flux-2-dev has been removed: every request to it was failing upstream. POST /v1/images/generations with "model": "flux-2-dev" now returns 404 with code model_not_found. Use flux-2-klein-9b, which is also free. nano-banana-2 and flux-1.1-pro now fall back to flux-2-klein-9b. The free tier is 26 models (5 image). Past images and usage records are unchanged.

Nova Sonic models removed

  • nova-sonic-2 and nova-sonic have been removed from the catalog. GET /api/v1/models no longer lists them, and a voice session that names either returns 422. Use gpt-realtime-2.1-mini for speech-to-speech. An agent still configured with either model keeps answering calls on the standard STT, LLM and TTS pipeline; past usage and session cost records are unchanged.

Upstream failures now return 503 with Retry-After

  • Failures of a model or channel behind the API, and gateway timeouts, now return 503 (with a Retry-After header) instead of 502 or 504. Our CDN replaces an origin 502 or 504 with its own HTML error page, so your client never saw the JSON error body or its reason code. Reason codes and retry advice are unchanged: only the status moved. A timeout that may have applied the action (for example a social action or a Facebook publish) is still ambiguous, so check before retrying. See Errors.

New models: Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5

  • claude-opus-5-5, claude-sonnet-5-5 and claude-haiku-5-5 — Anthropic's Claude 5.5 family, available on the Pro plan and higher (free and Starter keys get 403 model_not_available). 1M context, 128K max output, text + image input, reasoning and tool calling.
  • Pricing per 1M tokens — Opus $4.167 in / $20.83 out ($0.2083 cached input); Sonnet $2.083 in / $10.42 out ($0.1042 cached); Haiku $0.1042 in / $0.5208 out ($0.01042 cached), rising to $0.5208 / $2.604 ($0.05208 cached) for prompts over 100K tokens.
  • reasoning_effort — accepts low, medium, high, xhigh and max; none and minimal run as low. temperature, top_p and top_k are ignored. See API speed and function calling for the anthropic_thinking field to send back with tool calls.

Nova Sonic models under maintenance

  • nova-sonic-2 and nova-sonic are under maintenance: POST /v1/voice/sessions returns 503 with a message naming the alternative, and GET /api/v1/models shows status: "maintenance". Use gpt-realtime-2.1-mini for speech-to-speech. An agent still configured with either model keeps answering calls on the standard STT, LLM and TTS pipeline.

Usage API — costs in US$, prompt-cache totals

  • cost_usd on /v1/usage/summary, /v1/usage/logs and /v1/usage/logs.csv (and the get_usage_summary / list_usage_logs MCP tools) is now the US$ you were billed at the published rate (1 credit = ₹1, US$1 = ₹96). It was previously reported in internal billing units (0.01 per credit), about 4% below the US$ amount; credits = US$ × 96. Past rows are converted the same way, so a window spanning the change stays consistent.
  • /v1/usage/summary reports prompt-cache usage: totals.total_cache_read_tokens, totals.total_cache_creation_tokens, and cache_read_tokens / cache_creation_tokens per service. input_tokens keeps its meaning (the uncached part of the prompt).

Per-model parameters follow each model's documentation

  • Output limits — GET /api/v1/models now lists max_output_tokens where a model documents one, and max_tokens above it returns 400 max_tokens_too_large on /v1/chat/completions, /v1/responses and /v1/messages. The gemini-* models now allow up to 65,536 output tokens (previously capped at 8,192), and an omitted max_tokens leaves the model's own default.
  • reasoning_effort — xhigh and max now reach every model instead of being cut to high first; each model maps them to its own top level. DeepSeek-V4-Pro/-Flash and glm-5.2 now honour their levels, nemotron-3-super switches to its low-effort mode for low/none, sarvam-105b switches thinking fully off for none, and gpt-5-mini uses minimal. See API speed.
  • Sarvam — when you omit max_tokens, we send 4,096 (16,384 for glm-5.3) so thinking no longer cuts answers short at 2,048. A glm-5.3 or gemma-4-31b request over 10 MB returns 413.
  • Models — gpt-4.1 serves a 300,000-token context window. The realtime gpt-realtime-2* models take 32,000 input and 4,096 output tokens. nova-sonic was retired upstream on 2026-09-14 and returns 503; use nova-sonic-2.
  • Images and speech — image sizes outside a model's range, and negative_prompt on lucid-origin, return 400. gpt-4o-mini-tts has 13 voices and rejects an unknown voice with 400. gpt-4o-transcribe-diarize returns speaker segments with response_format verbose_json or diarized_json; it is no longer accepted as a voice-session stt_model (422), because live transcription cannot label speakers. grok-4.3 reasoning tokens now count as output tokens, and its images must have at least 512 pixels.

Email domains — neutral 503 reason

  • The 503 returned by the email domain routes when provisioning or verification is temporarily unavailable now carries reason: "provider_unavailable". Deployments still being upgraded may return the older acs_unavailable for the same condition; status, body shape and retry advice are identical, so match both strings. See Email limits.

Voice session webhooks

  • voice_session.ended and voice_session.failed now fire exactly once per session, on every way a session can end (the agent hanging up, DELETE, the session timing out), and only after the end has been saved. If two end signals arrive together, the first one decides end_reason and duration_seconds.
  • A session that ends because the agent hit an error (end_reason: "agent_error") now has status failed and fires voice_session.failed on every path; it previously fired voice_session.ended for some sessions. See Webhooks.

Speech-to-text error codes

  • /v1/audio/transcriptions and /v1/audio/translations now return 400 invalid_request for a problem with the request (unsupported audio, a streaming-only or unknown model), 429 rate_limit_exceeded when rate limited, and 503 service_unavailable when transcription is temporarily unavailable, the same code /v1/audio/speech uses. See Speech-to-Text.

Batch API, zero data retention, per-end-user budgets and prompt caching

  • Batch API — OpenAI-compatible POST /v1/files and POST /v1/batches: upload a JSONL file of requests, run it asynchronously, and download the results. Gated by the key's llm permission. See Batch API.
  • Zero data retention — send "provider": {"zdr": true} to route a request only to models that keep no data, or enforce it for a whole key or account. See Gateway controls.
  • Per-end-user budgets — monthly credit caps per end user of your app, managed under /api/v1/gateway/end-user-budgets (scopes end_user_budgets:read / end_user_budgets:write). See Gateway controls.
  • Prompt caching — prompt_cache_key on /v1/chat/completions and /v1/responses passes through as a cache-routing hint to the models that support it (and is dropped elsewhere); cache_control blocks on /v1/messages are accepted so Anthropic SDK code runs unchanged, but explicit breakpoints are not applied yet — caching works from the prompt prefix. usage.prompt_tokens_details.cached_tokens is now always present, and the Usage API reports cache reads and writes.

Payments, evals, contacts and telephony

  • Customer payment links — agents can send payment links on calls and WhatsApp that are paid into your own connected Razorpay account. Track them under /api/v1/payment-requests and subscribe to the payment_request.* webhook events.
  • Evals — an llm_judge assertion, a pass-rate gate for CI, and creating an eval case from a real call. See Voice evals.
  • WhatsApp identities on contacts — contacts carry whatsapp_bsuid, whatsapp_parent_bsuid and whatsapp_username, for WhatsApp users who hide their phone number. A BSUID is unique per account like a phone number or email. See Contacts.
  • TRAI readiness for outbound campaigns — a readiness check for AI outbound calling under the TCCCPR amendment. See Voice campaigns.

Newly documented

  • Contacts (/api/v1/contacts, plus each contact's memory), the handoff queue (/api/v1/handoffs), integrations and sheet automations (/api/v1/integrations, /api/v1/sheet-automations), payment requests, and POST /api/v1/crm/csv/inspect were already callable with an API key and now have reference pages.
  • Webhooks — the delivery body, the X-CallMissed-Event / X-CallMissed-Delivery headers, the retry schedule, and the call.amd_detected event.
  • Idempotency and Errors — rewritten to match exactly what the API does.

September 2026

Account MCP server — run your account from an AI assistant

  • Account MCP server — connect Claude, ChatGPT or any MCP client to https://api.callmissed.com/api/v1/mcp, sign in with your CallMissed account or an API key, and act on your account through 331 tools: voice agents, calls and campaigns, the inbox, CRM, support desk, WhatsApp, email and more. Every tool is gated by the key's scopes, spending tools are marked, and destructive ones ask first. See Account MCP.
  • Multilingual voice agents — voice tiers and custom agents accept an auto-detect language, so one agent can answer in the caller's language. See Voice tiers.

New model — GPT-6.1 Sol

  • gpt-6.1-sol — OpenAI's GPT-6.1 Sol: near-Astra performance for complex coding, computer use and professional work at a lower cost. 1.05M context, 128K output, text + image input, reasoning and tool calling. $2.083 in / $10.42 out per 1M tokens ($0.1042 cached input).
  • Long prompts — requests with more than 272K input tokens are billed at 2× input and cached-input rates and 1.5× output for the whole request ($4.167 in / $0.2083 cached / $15.63 out per 1M).
  • reasoning_effort — accepts low, medium (default), high and xhigh. The model always reasons, so none and minimal are sent as low, and max is sent as xhigh. Only the default temperature is supported, and max_tokens must be at least 3. See API speed.
  • Paid plans only. See Models.

New models — GPT-6 Sol and GPT-6 Luna

  • gpt-6-sol — OpenAI's GPT-6 frontier reasoning model for enterprise agents, coding and complex knowledge work; succeeds gpt-5.6-sol. 1.05M context, 128K output, text + image input, reasoning and tool calling. $2.083 in / $10.42 out per 1M tokens ($0.2083 cached input).
  • gpt-6-luna — the efficient GPT-6 model for high-volume, cost-sensitive workloads; succeeds gpt-5.6-luna. Same 1.05M context, vision, reasoning and tools. $0.1042 in / $0.5208 out per 1M tokens ($0.01042 cached input).
  • Long prompts — requests with more than 272K input tokens are billed at 2× input and cached-input rates and 1.5× output for the whole request.
  • reasoning_effort — both accept none, low, medium, high and xhigh (minimal is sent as low). When a request includes tools, reasoning is set to none automatically. max_tokens must be at least 3. See API speed.
  • Paid plans only. See Models.

August 2026

Managed Voice Agent — speech-to-speech over one WebSocket

  • Managed Voice Agent — a full speech-to-speech pipeline behind a single WebSocket. Stream microphone audio in, get synthesized speech and conversation events back; speech recognition, the language model, text-to-speech, turn-taking and interruption handling are all run and tuned for you. No WebRTC and no client SDK. Two protocols on the same host and the same engine: wss://api.callmissed.com/v2/voice/agent (CallMissed-native) and wss://api.callmissed.com/v1/agent/converse (Deepgram Voice Agent compatible — an existing Deepgram integration can repoint its URL and work unchanged). Supports in-call tool calling, live model/prompt/voice updates, and sessions up to 2 hours. See Managed Voice Agent.
  • Voice model catalogue — GET /api/v1/voice/models lists every selectable speech-to-text, language and text-to-speech model with a measured latency verdict (eligible, too_slow, unsupported, unmeasured) plus its p50, sample count and budget. Any eligible combination is valid. Models too slow to hold a conversation are not offered — they are listed with the reason, rather than silently disappearing or being offered with a caveat. Verdicts come from real production traffic, not vendor claims, and a model with too few samples reads unmeasured rather than being assumed fast.

Embeddings, usage API, CRM, support desk and voice-agent operations

New models — Cartesia Ink STT

  • ink-whisper — Cartesia's fastest and most affordable STT at $0.1875 / hour, across 100 languages including Hindi, Urdu and Tamil. Better accuracy than baseline Whisper, and dynamic chunking that cuts hallucination during pauses and silence. Works for both file transcription and voice sessions. See Speech to Text.
  • ink-2 — Cartesia's top-ranked STT for voice agents at $0.5625 / hour: 8% WER on AppTek's 14-accent call-centre benchmark, against 10% for Deepgram Flux and 12% for ElevenLabs. Self-detects turns, so no separate turn detector is needed. Two limits: it is English only, and it is voice-session only — the file transcription endpoint returns 400 and points you to ink-whisper. See Speech to Text.

New models — conversational Indic LLM, Saaras V4 STT, Flux TTS

  • sarvam-105b-conversations — 105B MoE tuned for conversation and voice. 128K context, tool calling, streaming, hybrid thinking. Free-tier, same $0.1563 in / $0.625 out per 1M as sarvam-105b. See Indic Models.
  • saaras:v4 — Sarvam STT with five output modes (transcribe, translate, verbatim, transliterate, code-mix) across 24 languages. Free-tier at $0.3125 / hour. See Speech to Text.
  • Deepgram Flux TTS — streaming-first TTS built for voice agents: turn-based synthesis with prosody carried across turns. 36 English voices including priya (Indian-accented English, the default). English only, no expressive controls. Available only through the managed Voice Agent (tts_engine: "flux"), billed inside the per-minute voice rate. See Voices.
  • Free tier — now 27 models (11 LLM, 4 STT, 4 TTS, 6 image, 2 embedding).

Model catalog update — retired models

  • Retired LLM IDs — the following model IDs are no longer served: openai/gpt-5.4-pro, openai/gpt-5.4, openai/gpt-5.4-mini, openai/gpt-5.4-nano, anthropic/claude-opus-4.6, anthropic/claude-sonnet-4.6, anthropic/claude-haiku-4.5, x-ai/grok-4.20, qwen/qwen3.5-plus, qwen/qwen3.5-flash, mistralai/mistral-small-2603, and the auto auto-router.
  • Migration — use the first-party flagships (gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna, gpt-5.5, grok-4.3) or the direct-routed free tier (kimi-k2.6, kimi-k2.7-code, glm-5.2, gpt-oss-120b, mistral-small-3.1). See Models.
  • Free tier — now 24 models (11 LLM). The auto free auto-router is retired; pick a free model explicitly.
  • Endpoints unchanged — POST /v1/chat/completions and the Anthropic-compatible POST /v1/messages both continue to work and accept every current catalog ID.

June 2026

  • WebRTC voice sessions — POST /v1/voice/sessions returns a connection token + URL; CallMissed handles the STT→LLM→TTS pipeline. List, fetch, fetch transcript (json | txt | srt), and end sessions under /v1/voice/sessions. The legacy /ws/voice-agent WebSocket still works for backward compatibility. See Voice Session API.
  • Image Generation API — POST /v1/images/generations (OpenAI-compatible). Free models include flux-2-klein-9b, flux-2-dev, lucid-origin, phoenix-1.0, sdxl-lightning, dreamshaper-8-lcm; paid models include flux-2-pro, flux-1.1-pro, nano-banana-2, and nano-banana-pro. See Image Generation.
  • Web Search API — POST /v1/search defaults to Serper web search (Exa, Firecrawl also available); flat 1 credit per query. See Web Search.
  • Knowledge RAG — vector knowledge sources at /api/v1/knowledge/sources (ingest text, URL, or PDF; chunked + embedded) with semantic search at POST /api/v1/knowledge/search.

May 2026

v1.5.0 — First-Party Models, Account Security & WhatsApp Platform

  • First-party models — deployments callable by bare ID: gpt-4o, gpt-4.1, gpt-5-mini, grok-4.3, DeepSeek-V4-Pro, DeepSeek-V4-Flash, plus first-party STT (whisper, gpt-4o-transcribe, gpt-4o-mini-transcribe, gpt-4o-transcribe-diarize) and TTS (gpt-4o-mini-tts).
  • More STT/TTS — whisper-large-v3-turbo (99 langs), nova-3 (diarization), aura-2-en / aura-2-es, and melotts — all free-tier.
  • Kimi K2.6 — kimi-k2.6 added to the direct-routed free tier alongside kimi-k2.5.
  • TOTP 2FA & passkeys — two-factor auth (authenticator apps + backup codes) and passkeys for dashboard sign-in. Active sessions can be reviewed and revoked from the dashboard.
  • WhatsApp platform — Embedded Signup onboarding, message templates, broadcast campaigns, and delivery analytics under /api/v1/whatsapp/*. See WhatsApp API.
  • Billing surfaces — coupon redemption, downloadable PDF invoices, and a credit ledger broken down by transaction type, all in the dashboard.
  • Audit log — a sensitive-action audit feed in the dashboard.

April 2026

v1.4.0 — Anthropic API Compatibility & Audio Translation

  • Anthropic Messages API — New POST /v1/messages endpoint. Use the Anthropic SDK with CallMissed by changing only the base_url. Full streaming support with Anthropic SSE lifecycle (message_start, content_block_delta, message_stop).
  • Dual auth headers — Anthropic endpoint accepts both x-api-key and Authorization: Bearer headers
  • Model aliasing — A bare model name on the Anthropic endpoint resolves against the CallMissed catalog
  • Audio Translation — New POST /v1/audio/translations endpoint. Translate audio in 24 languages to English text. OpenAI SDK compatible (client.audio.translations.create())
  • Token counting — POST /v1/messages/count_tokens for input token estimation
  • Anthropic rate limit headers — anthropic-ratelimit-requests-limit, anthropic-ratelimit-requests-remaining, etc.

v1.3.0 — Voice Agent & Ultra-Low-Latency Pipeline

  • Voice Agent WebSocket — Real-time STT→LLM→TTS pipeline over /ws/voice-agent. PCM audio in, streaming MP3 out. LLM and TTS run concurrently for minimum latency.
  • PCM AudioWorklet capture — Raw PCM s16le at 16kHz, no container overhead
  • Streaming MP3 playback — MediaSource API appends and plays chunks as they arrive
  • Profile management — Save and update user profile from the dashboard

v1.2.0 — Security, Google OAuth & Plan Enforcement

  • Sign in with Google — Google sign-in for the dashboard. Auto-creates the organisation and user, and links to an existing account by email.
  • OTP Authentication — Email-based OTP for passwordless login and password reset
  • Plan limit enforcement — Server-side usage caps per plan tier (free/starter/pro/enterprise). API returns 429 quota_exceeded when limits reached. Usage headers (X-RateLimit-*, X-Usage-Warning) on every response.
  • Per-API-key rate limiting — a requests-per-minute limit on every key (today set by plan; see Rate limits)
  • Model catalog update — OpenAI gpt-5.4 family, Anthropic Claude 4.6, Google Gemini 3.1, xAI Grok 4.20, Qwen 3.5, Mistral Small
  • Knowledge Base file upload — Upload PDF, DOCX, TXT files (max 20 MB) with auto text extraction
  • Bot deployment verification — Verify WhatsApp/Twilio channel connectivity from the dashboard
  • Settings verification — Verify WhatsApp, Twilio, and Indic LLM API connectivity

v1.1.0 — Platform Playground

  • Playground rebuild — LLM (streaming + non-streaming), STT (file upload + mic), TTS (37 voices across 11 Indian languages), Voice Agent demo

v1.0.0 — Initial Release

  • Chat Completion API — OpenAI-compatible endpoint with streaming, tool calls, and function calling
  • Speech to Text — saaras:v3 with 22 Indic language support
  • Text to Speech — bulbul:v3 (37 voices across 11 Indian languages)
  • WhatsApp Bot — Full WhatsApp Business API integration
  • Voice Calling — Twilio-based inbound voice with WebSocket streaming
  • Multi-tenant — Complete tenant isolation with role-based access
  • API Keys — Scoped API keys with usage tracking
  • Webhook Delivery — Outbound webhooks with retry and HMAC signing
  • Analytics Dashboard — Real-time conversation and usage analytics
  • Model catalog — LLM, STT, TTS and image models from one OpenAI-compatible endpoint

On this page

November 2026CRM integration API (2026-11-17)Legacy voice WebSocket removed (2026-11-11)ElevenLabs voices and Scribe transcription (2026-11-07)Kimi K3.1 Pro (2026-11-06)Voice calls keep going through rate limits (2026-11-02)Numbers nothing answers play a short message (2026-11-02)GLM 5.3 Flash and DeepSeek V4.1 Flash 1017 (2026-11-01)October 2026Deepgram Think-Fast and phoneLLM mini chat models (2026-10-31)Gnani Prisma v3, and Prisma v2.5 retired (2026-10-30)Deepgram updates (2026-10-30)Preview: Microsoft AI speech models (2026-10-30)API behaviour fixesHosted in India voice agentsNew voice-agent models: Microsoft Speech and Microsoft Neural voicesNew voice-agent models: AssemblyAI, Inworld, Rime, Fish Audio, Gradium and Deepgram Flux TTSModels moved, retired and deprecatedGnani Timbre v3, and older speech models retiredNew models: DeepSeek V4.1 Flash and Bulbul V4 Flashflux-2-dev removedNova Sonic models removedUpstream failures now return 503 with Retry-AfterNew models: Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5Nova Sonic models under maintenanceUsage API — costs in US$, prompt-cache totalsPer-model parameters follow each model's documentationEmail domains — neutral 503 reasonVoice session webhooksSpeech-to-text error codesBatch API, zero data retention, per-end-user budgets and prompt cachingPayments, evals, contacts and telephonyNewly documentedSeptember 2026Account MCP server — run your account from an AI assistantNew model — GPT-6.1 SolNew models — GPT-6 Sol and GPT-6 LunaAugust 2026Managed Voice Agent — speech-to-speech over one WebSocketEmbeddings, usage API, CRM, support desk and voice-agent operationsNew models — Cartesia Ink STTNew models — conversational Indic LLM, Saaras V4 STT, Flux TTSModel catalog update — retired modelsJune 2026v1.6.0 — WebRTC Voice, Image Generation & Web SearchMay 2026v1.5.0 — First-Party Models, Account Security & WhatsApp PlatformApril 2026v1.4.0 — Anthropic API Compatibility & Audio Translationv1.3.0 — Voice Agent & Ultra-Low-Latency Pipelinev1.2.0 — Security, Google OAuth & Plan Enforcementv1.1.0 — Platform Playgroundv1.0.0 — Initial Release