Changelog
This page lists all notable changes to the CompactifAI API. Deprecations and breaking changes are announced by email at least 15 days ahead: see Deprecations.
2026-11-18
Section titled “2026-11-18”- An invalid key returns
401even when the model does not exist – a key that is not recognised, revoked or expired now returns401before themodelis checked onPOST /v1/chat/completions,POST /v1/completions,POST /v1/embeddings,POST /v1/messages,POST /v1/messages/count_tokens,POST /v1/responsesandPOST /v1/audio/transcriptions. Before, a well-formed key that does not exist got400(model_not_found) for an unknown model. A valid key on an unknown model still returns400(model_not_found). See Error handling.
2026-11-05
Section titled “2026-11-05”- Guardrails on
POST /v1/chat/completions– a deployment can enable guardrails that check every chat completion before it reaches the model. A blocked request returns403(content_policy_violation). On a deployment that requires the check, a request it cannot run on is not sent and returns503(policy_check_unavailable, retryable). Deployments without guardrails are unchanged. See Error handling.
2026-11-03
Section titled “2026-11-03”Removed
Section titled “Removed”quasar-438b– Quasar 438B is decommissioned and removed from the model catalog. Migrate toquasar-1-2-358b. See Quasar 1.2 358B.
2026-10-24
Section titled “2026-10-24”- Prepaid wallet balance never goes below zero – a request that costs more than the remaining balance is charged only up to the balance. The balance stops at $0.00 instead of going negative, and the part of that request’s cost that the balance did not cover is not charged.
2026-10-23
Section titled “2026-10-23”Changed
Section titled “Changed”qwen-3-8-27boutput price – the output price is now $0.44 per 1M tokens (was $0.04). Input ($0.09) and cached input ($0.0171) are unchanged, with reasoning on or off. See Pricing.
2026-10-22
Section titled “2026-10-22”Changed
Section titled “Changed”- New API prices – input, cached input and output prices are lowered for
glm-5-3,glm-5-2,qwen-3-8-27b,quasar-438b,quasar-1-2-358bandcarina-60b. Prices per 1M tokens, input / cached input / output:glm-5-3$0.20 / $0.018 / $0.64,glm-5-2$0.17 / $0.0153 / $0.54,qwen-3-8-27b$0.09 / $0.0171 / $0.04 (the same price with reasoning on or off),quasar-438bandquasar-1-2-358b$0.11 / $0.0099 / $0.33,carina-60b$0.022 / - / $0.085. Cached input is billed at 9% of the input price for GLM 5.3, GLM 5.2, Quasar 438B and Quasar 1.2 358B, and at 19% for Qwen 3.8 27B. See Pricing.
2026-10-21
Section titled “2026-10-21”Changed
Section titled “Changed”- A
systemmessage in the middle of a conversation is honoured onPOST /v1/messages– a message with rolesystemthat comes after a user or assistant turn is now treated as a developer-level instruction: it is binding from where it appears and overrides an earlier user request that conflicts with it, for example “calling any tool is forbidden, reply only with: Unable to query” after a user turn that asks for a tool call. Before, the model could read such a message as untrusted content and ignore it, so the tool was still called, as in Tencent’s onboarding report. This matches how the original model vendor behaves and howPOST /v1/chat/completionsandPOST /v1/responsesalready treat the caller’ssystemmessages. Asystemmessage at the start of the conversation, and requests without one, are unchanged. POST /v1/responsesno longer returns empty per-turn usage arrays –usage.input_tokens_details.input_tokens_per_turn,usage.input_tokens_details.cached_tokens_per_turn,usage.output_tokens_details.output_tokens_per_turnandusage.output_tokens_details.tool_output_tokens_per_turnare removed from the response body, from the response objects inside the stream events, and from the stored copy thatGET /v1/responses/{id}returns. They were always[], which is not a valid token count, and the original model vendor does not return them. The token totals (input_tokens,output_tokens,total_tokens) and the other detail fields (cached_tokens,reasoning_tokens) are unchanged.
POST /v1/responsesnames output items in one format, streamed or not – reasoning items arers_…, function callsfc_…and messagesmsg_…, and thecall_idof a function call starts withcall_. A streamed response used ids without these prefixes, and a response that was not streamed usedchatcmpl-tool-…ascall_id. The same applies to theitem_idof the stream events, toresponse.completedand to the stored response.POST /v1/responsesechoesprevious_response_id– a response to a request that carriesprevious_response_idreturns it inprevious_response_id(and in theresponseof theresponse.created,response.in_progressandresponse.completedevents); it used to benull.- A streamed tool call can be answered with
previous_response_idonPOST /v1/responses– in a streamed response,response.completedcould name an output item (id, andcall_idfor afunction_call) differently from theresponse.output_item.doneevent of the same item, and the stored response keeps the names from the item events. A follow-up turn that sent only afunction_call_outputwith thecall_idfromresponse.completedreturned400function_call_not_found.response.completednow carries the same items as the item events, so the stream and the stored response name each item the same way.
2026-10-20
Section titled “2026-10-20”POST /v1/detectchecks whether a text was watermarked – send amodeland atextand get backwatermarkedand a coarseconfidence(none,low,medium,high). The raw score is not returned. A model that is not available for detection returns400(model_not_found). See Watermark Detection.
Changed
Section titled “Changed”- Non-streamed
POST /v1/responsesreturns400for an input it cannot serve – an input of only ignored items (compaction_trigger,item_reference,computer_call, …), or one with custom tool items and no user message, used to return200with afailedresponse. Without streaming it now returns400(invalid_request, e.g.Input cannot be empty[<request id>]), as the origin provider does. Streamed requests are unchanged. - More input item types are ignored on
POST /v1/responses–file_search_call,code_interpreter_call,image_generation_call,local_shell_call,shell_call,apply_patch_call(and their_outputitems),tool_search_call/tool_search_output,compaction,program/program_outputandconfiguration_updateused to return400(Prefill failed). They are now dropped, as the origin provider does. An unknown item type is dropped too, instead of returning the engine’s400validation dump. - Custom tool items without a user message fail like the origin on
POST /v1/responses– an input that holds acustom_tool_callorcustom_tool_call_outputbut no user message now fails witherror.codeinvalid_request(400without streaming, aresponse.failedand anerrorevent when streaming), as the origin provider does. With a user message anywhere in the input, before or after them, a custom tool call and its output are sent to the model as afunction_call/function_call_outputpair and answered. includeas a string no longer fails onPOST /v1/responses– a non-arrayinclude(for example"include": "web_search_call.results") used to return a400without streaming and a502upstream_unavailablewithstream: true. It is now ignored, as the original model vendor does.- The response of
POST /v1/responsesno longer carries backend-only fields –input_messages,output_messages,kv_transfer_paramsandec_transfer_paramsare removed from the response body and from the response objects inside the stream events (response.created,response.in_progress,response.completed,response.incomplete,response.failed), and from the stored copy thatGET /v1/responses/{id}returns. - The model decides when to use
web_searchon/v1/responses– Theweb_searchtool is now offered to the model as a tool. The model decides whether to search; when it calls the tool, the API runs the search, gives the results back to it and the model answers, up toWEB_SEARCH_MAX_USESsearches per request (default 5). Before,/v1/responsessearched on every request that carried the tool andincludeasked for its output, using the last user message as the query, even for prompts like “Reply with OK.” A request where the model does not search has noweb_search_callitem and no extra input tokens, andincludeno longer gates the search: it only chooses whetheraction.sourcesandresultsare returned. The output keeps the model’sreasoningandmessageitems between the searches, a failedweb_search_callcarries anerrorstring, and the model can read a page (web_search_callwithaction.type: "open_page").usageadds up every model call, reasoning tokens included. A streamed request relays the model’s events live and pauses while a search runs. See Web search.POST /v1/chat/completionskeeps searching on every request with an enabledweb_searchtool; its stream now carries theweb_searchsources on the closing chunk and norequest_id.POST /v1/messagesis unchanged: it accepts theweb_search_*server tool and does not run it. container_uploadcontent blocks are accepted onPOST /v1/messages– they used to fail with a 400 (Unsupported content block type 'container_upload'); they are now dropped before the request reaches the model, as the original model vendor does. There is no container or Files API behind the block, so the file is not read. A message that holds nothing butcontainer_uploadblocks is still rejected with a 400. See File uploads to a container.
2026-10-19
Section titled “2026-10-19”Changed
Section titled “Changed”-
Images sent to a text-only model on
POST /v1/responsesno longer fail – For a model without image support (capabilities.supports_imageisfalse), eachinput_imagepart is now replaced by a short text notice and the request is answered normally, so the model replies that it cannot see the image. Before, the request failed with400and the provider’s raw<model> is not a multimodal modelerror. This matches how the original model vendor handles the same request.POST /v1/chat/completionsandPOST /v1/messagesare unchanged and still reject images for such models. See Images sent to a model without image support. -
thinking.display: "omitted"no longer hides the thinking text onPOST /v1/messages– Withthinkingset toenabledoradaptiveanddisplayset to"omitted", the response used to carry athinkingblock with an emptythinkingstring (and nothinking_deltaevents when streaming). It now returns the thinking text, as the original model vendor does, sodisplayis accepted but has no effect. See Thinking.
-
containeris accepted onPOST /v1/messages– the field used to fail with a400(`container` is not supported by this API). It is now accepted in any shape (a string, an object withidandskills, …) and ignored, streaming and non-streaming, as the original model vendor does: there is no code-execution sandbox, so no container is created, reused or returned. See Request shape. -
x-api-keyaccepts aBearerprefix – OnPOST /v1/messagesandPOST /v1/messages/count_tokens, thex-api-keyvalue may be the raw API key orBearer <key>. Both resolve to the same credential, with the same billing and rate limits. Before, the prefixed form was rejected with401. -
input_fileparts on/v1/responsesare converted for the model – The models behindPOST /v1/responsesaccept no file content part and rejectedinput_filewithunsupported_content. The gateway now reads the file itself: PDF, Word (.docx), PowerPoint (.pptx), Excel (.xlsx, first 1,000 rows per sheet) and text files are sent as text, and images (png, jpeg, gif, webp) asinput_image.file_data, adata:file_urlandfile_idare handled, with OpenAI’s limits of 50 MiB per file and per request. A remotehttp(s)file_urlis not supported and returns a 400unsupported_file_url, and PDF page images are not sent, so scanned PDFs without a text layer return a 400file_has_no_text. Other types and files over the limits return a 400 that names the problem. Safeguards are configurable (RESPONSES_FILE_MAX_PDF_PAGES,RESPONSES_FILE_MAX_CHARS,RESPONSES_FILE_MAX_PARTS,RESPONSES_FILE_MAX_CONCURRENT_EXTRACTIONS); when every extraction slot stays busy a request gets a 503file_processing_busy. See OpenAI Compatibility. -
top_p: 0is accepted on chat completions and responses –POST /v1/chat/completionsandPOST /v1/responsesnow accepttop_pvalues in[0, 1]and treat0as a request for greedy sampling: the gateway forwardstemperature=0andtop_p=1so the output is deterministic, taking precedence over any temperature sent in the same request. Before, chat rejected0with a validation error and responses forwarded it to the provider, which refused it.POST /v1/messageskeeps its existing behavior of translating zero intotop_k=1.
-
promptis accepted onPOST /v1/responses– anyprompt(a template reference withid,versionandvariables, or a plain string) used to fail with a 400 from the model provider (prompt template is not supported). It is now dropped before the request reaches the model, streaming and non-streaming, and the request is answered frominputandinstructions. Templates are not resolved. See Responses. -
background: trueis accepted onPOST /v1/responses– it used to fail with a 400 from the model provider; the response is now generated synchronously and returned in the same call. -
Newer fields in
/v1/responsesresponses –POST /v1/responsesno longer returns502witherror.codeupstream_unavailablewhen the model’s response contains values newer than the API’s schema, such asreasoning.effortmaxor tool typescomputer,namespaceandtool_search. Those responses now return200unchanged, andGET /v1/responses/{response_id}returns stored responses that contain them. A real provider outage still returns502. -
compaction_triggerinput items are accepted onPOST /v1/responses– they used to fail with a 400 from the model provider (Unsupported input item type); they are now dropped before the request reaches the model, so the rest of the input is answered normally. The conversation is not compacted. A request whose input holds nothing but ignored items returns200with a response whosestatusisfailedand whoseerror.codeisinvalid_request(Input cannot be empty[<request id>], where<request id>is the response id without itsresp_prefix); withstream: trueit arrives as aresponse.failedevent followed by anerrorevent, as the origin provider does. -
item_referenceinput items are accepted onPOST /v1/responses– they used to fail with a 400 from the model provider (Unsupported input item type); they are now dropped before the request reaches the model. The referenced item is not looked up: send its content inline or chain turns withprevious_response_id. -
Custom tool items are accepted on
POST /v1/responses–custom_tool_callandcustom_tool_call_outputinput items used to fail with a 400 from the model provider (Unsupported input item type); they are now sent to the model as afunction_call/function_call_outputpair, with the free-form input as{"input": "..."}. Acustom_tool_callwithout anamereturns400witherror.codecustom_tool_call_missing_name. -
Replayed
web_search_calland computer-use items are accepted onPOST /v1/responses– they used to fail with a 400 from the model provider (Unsupported input item type). Aweb_search_callnow reaches the model as an assistant message naming its queries and sources;computer_callandcomputer_call_outputitems are dropped.
2026-10-18
Section titled “2026-10-18”- Web search on
/v1/responses–POST /v1/responsesnow runs a web search when the request has aweb_searchtool andincludehasweb_search_call.action.sourcesand/orweb_search_call.results. The response starts with aweb_search_calloutput item carrying the search query and, depending oninclude, the result URLs (action.sources) and the results (results:title,snippet,url); the model answers from those results. Streaming adds theresponse.web_search_call.in_progress,.searchingand.completedevents. Aweb_searchtool without thoseincludevalues does not run a search.
- Unsupported
includevalues on/v1/responses–POST /v1/responsesno longer returns400whenincludecontains a value that is not supported. Unsupported values are ignored and have no effect on the response, withstreamon or off. Supported values (code_interpreter_call.outputs,computer_call_output.output.image_url,file_search_call.results,message.input_image.image_url,message.output_text.logprobs,reasoning.encrypted_content) are unchanged.
2026-10-15
Section titled “2026-10-15”-
x-api-keyauthentication on the Anthropic Messages endpoints –POST /v1/messagesandPOST /v1/messages/count_tokensaccept the API key in thex-api-keyheader, as Anthropic SDKs send it, in addition to the existingAuthorization: Bearerscheme. Same key, same billing and rate-limit identity. Send exactly one of the two headers: a request carrying both is rejected with400, even when both are valid. -
quasar-438bis back – Quasar 438B is served again and re-listed in the models catalog alongside Quasar 1.2 358B. Pricing stays $0.60 per 1M input tokens and $1.80 per 1M output tokens; prompt caching is not enabled for this model.
Changed
Section titled “Changed”quasar-358brenamed toquasar-1-2-358b– The Quasar 1.2 358B model id is nowquasar-1-2-358b(previouslyquasar-358b, andquasar-2-358bbefore that). Prices and capabilities are unchanged. Update themodelfield in your requests.
2026-10-14
Section titled “2026-10-14”Changed
Section titled “Changed”- Unknown
modelreturns400on every endpoint that takes one –POST /v1/chat/completions,POST /v1/completions,POST /v1/embeddings,POST /v1/messages,POST /v1/messages/count_tokens,POST /v1/responsesandPOST /v1/audio/transcriptionsnow return400witherror.codemodel_not_foundanderror.parammodelfor a model that does not exist. Before, chat, messages, embeddings, completions and count_tokens returned404. This matches OpenAI. A request with an invalid key still returns401before the model is checked. The error message no longer lists the available models./v1/completionskeeps its existing error body shape.GET /v1/models/{model}still returns404.
- Disabling reasoning on
/v1/responses–POST /v1/responsesnow refuses a request that disables reasoning on a model whose reasoning is always on, for exampleglm-5-3, with400anderror.codereasoning_not_disableable, asPOST /v1/chat/completionsandPOST /v1/messagesalready do. Before, the request reached the provider and failed with502.
2026-10-11
Section titled “2026-10-11”-
watermarkingrequest field –POST /v1/chat/completions,POST /v1/completionsandPOST /v1/responsesnow acceptwatermarking(boolean). Chat and completions requests previously dropped it before reaching the model, sowatermarking: falsehad no effect. Watermarking is on by default (true) when the model is configured for it.falseopts the request out, and is only accepted in US deployments (us-*); in EU deployments it is refused with400andcode: watermarking_opt_out_not_allowed. See the API reference. -
Total token rate limits – Accounts can now be limited on input and output tokens combined, per minute, per hour, per day and over a rolling 30 days. Each request counts its prompt tokens plus its completion tokens against the total. Input and output limits are unchanged and still apply separately. See What counts against your limits.
2026-10-08
Section titled “2026-10-08”thinking: {"type": "disabled"}on/v1/messages–POST /v1/messagesno longer refusesthinking: {"type": "disabled"}on every model. It now turns reasoning off on models that can, asPOST /v1/chat/completionsdoes, and the response carries nothinkingblocks. Models whose reasoning is always on, for exampleglm-5-3, still refuse it with400anderror.codereasoning_not_disableable. Onlythinkingcontrols reasoning on this endpoint; other reasoning fields are ignored, not rejected.- Empty
toolsarray accepted –tools: []onPOST /v1/chat/completionsnow returns200, the same as omittingtoolsand as OpenAI. It previously returned400.
Changed
Section titled “Changed”-
thinking.clear_thinkingdefaults totrueon/v1/chat/completions– Reasoning that a client sends back on earlier assistant turns is now dropped from the prompt unlessthinking.clear_thinkingisfalse, including when the request has nothinkingobject. Before, such a request kept the earlier reasoning, which added prompt tokens. Setthinking: {"clear_thinking": false}to keep it. -
Image size limit –
POST /v1/chat/completions(image_url) andPOST /v1/responses(input_image) now reject requests whose inline images total more than 256 MB decoded with400andcode: image_too_large, instead of failing with a502/503connection error. See Image size limit. -
thinking.clear_thinkingdefaults totrueon/v1/chat/completions– Reasoning that a client sends back on earlier assistant turns is now dropped from the prompt unlessthinking.clear_thinkingisfalse, including when the request has nothinkingobject. Before, such a request kept the earlier reasoning, which added prompt tokens. Setthinking: {"clear_thinking": false}to keep it.
2026-10-07
Section titled “2026-10-07”quasar-1-2-358b– Quasar 1.2 358B chat model with tool calling, structured output, the Responses API, and the Anthropic Messages API. See Quasar 1.2 358B and the models catalog.- Video input for
qwen-3-8-27b–capabilities.supports_videois nowtrue: send video as a standardfilecontent part in chat completions. See Video understanding. - Prompt caching for
glm-5-2– Input tokens served from the cache are billed at 20% of the input price ($0.22 per 1M tokens, versus $1.10 per 1M for non-cached input). See Pricing. - Anthropic Messages API support (
supports_messages) —carina-60b,glm-5-2,glm-5-3,quasar-1-2-358b, andqwen-3-8-27bcan be called viaPOST /v1/messages. See Anthropic Messages API.
Removed
Section titled “Removed”quasar-438b– Removed from the model catalog. Migrate toquasar-1-2-358b.nemotron-3-nano-omniandmistral-small-3-1– Deprecated models removed from the catalog.
2026-10-05
Section titled “2026-10-05”minimalreasoning effort –reasoning.effortandreasoning_effortnow accept"minimal"alongside the existing levels. Each model still decides which levels it implements.
2026-10-03
Section titled “2026-10-03”- Prompt caching for
glm-5-3– Prompt caching is now enabled for GLM 5.3. Input tokens served from the cache are billed at 20% of the input price ($0.22 per 1M tokens, versus $1.10 per 1M for non-cached input). See Pricing.
2026-09-21
Section titled “2026-09-21”- No more giving up on a model too early – When a model isn’t immediately available, the API now tries an alternative before returning an error. Previously, if availability hadn’t been checked recently, it could skip that alternative and fail the request.
- Stalled requests now report 503 instead of 500 – A request where the model never starts responding now returns the retryable
503first_token_timeoutinstead of a generic500 internal_error. Retry onfirst_token_timeout.
2026-09-16
Section titled “2026-09-16”- No more
404 Model not foundfor a model that exists – A request for an existing model no longer returns404 Model not found. When the model cannot start responding, the client now receives a retryable503witherror.codefirst_token_timeoutorerror.coderequest_deadline_exceededinstead of500or404. Retry on either code.
2026-09-09
Section titled “2026-09-09”Deprecated
Section titled “Deprecated”hypernova-60b– Deprecated. Migrate tocarina-60b, which offers equivalent capabilities. See the models catalog.
2026-08-29
Section titled “2026-08-29”glm-5-3– GLM 5.3 chat model published by zai-org. See GLM 5.3 and the models catalog.
2026-08-25
Section titled “2026-08-25”chat_template_kwargs–POST /v1/chat/completionsnow accepts a top-levelchat_template_kwargsobject that is forwarded as-is to the model’s chat template renderer (vLLM/SGLang convention). The accepted keys depend on the model’s chat template — for example, Qwen models use{"enable_thinking": true}to toggle reasoning. This is model-agnostic: no catalog entry is required per model.
Removed
Section titled “Removed”cai-mistral-small-3-1-slim– Removed from the model catalog.gpt-oss-120b– Removed from the model catalog.
Deprecated
Section titled “Deprecated”reasoning_enabled– The top-levelreasoning_enabledrequest field is deprecated in favor ofchat_template_kwargs. When both are sent, the user-suppliedchat_template_kwargstakes precedence. Thereasoning_enabledfield will be removed in a future release.
2026-08-19
Section titled “2026-08-19”This release contains breaking changes to the error format. If your integration inspects error response bodies, read the migration notes below before upgrading.
Changed
Section titled “Changed”- Standardised error format (breaking) – All errors from
POST /v1/chat/completions,POST /v1/responses,POST /v1/audio/transcriptions,GET /v1/modelsandGET /v1/models/{model}now return a single top-levelerrorobject. Responses that previously used{"detail": ...}— including authentication failures, rate limits, request-validation errors and server errors — now use{"error": {"message", "type", "param", "code"}}. See Error Handling. - Request-validation errors (breaking) – Validation failures no longer return an array of field errors. They now return one
errorobject whoseparamidentifies the first offending field in dotted/indexed notation (for examplemessages[0].role), with a count of any remaining errors appended tomessage. - Streaming errors (breaking) – The
data:line of anevent: errorSSE event is now a JSON error envelope rather than a plain-text message, so it can be parsed with the same JSON decoder as every other event. Clients that string-matched the old text will no longer match. - Rate limit responses (breaking) – The
rate_limit_exceededidentifier has moved from the response body’sdetailfield toerror.code. TheRetry-Afterheader is unchanged. paramandcodeare now populated – Both fields were previously alwaysnull.paramnow names the offending request field andcodecarries a stable machine-readable identifier suitable for programmatic branching. See the error code reference.- Removed the
router_no_llm_providererror type (breaking) – 503 responses now use the standardserver_errortype. The router-specific identifier remains available aserror.code(no_healthy_llm_provider). - Audio transcription status codes (breaking) – A non-boolean
streamform value now returns 400 instead of 422. Model provider failures that are not client errors now return 502 instead of forwarding the provider’s own status code. - File inputs on
/v1/responses– A file resolved fromfile_idis now attached to the request’sinput(as aninput_textpart for text-decodable files, otherwise aninput_filepart) instead of being sent as a separate top-levelfilefield. Text-only models can now read uploaded text files. See OpenAI Compatibility. - Error messages rewritten – Many messages are clearer, and messages for server-side failures are now deliberately generic so that deployment details are not disclosed.
messagehas always been free-form; do not match on it — branch onerror.codeorerror.typeinstead.
Migration
Section titled “Migration”-
Read errors from
error, notdetail. Replaceresponse.json()["detail"]withresponse.json()["error"]["message"]:// Before{"detail": "Unexpected server error"}// After{"error": {"message": "Unexpected server error! Please retry shortly.", "type": "server_error", "param": null, "code": "internal_error"}} -
Branch on
error.code, not on message text or HTTP status alone. The full list is in the error code reference. -
Parse streaming error events as JSON. The
data:payload of anevent: erroris now an object; parse it with the same decoder you use for other events, excluding thedata: [DONE]sentinel. -
Update rate-limit handling to read
error.code == "rate_limit_exceeded"instead of checkingdetail, and continue honouring theRetry-Afterheader. -
If you use
POST /v1/completions, no changes are required — it retains the older{"detail": ...}error format. PreferPOST /v1/chat/completionsfor new integrations.
2026-08-17
Section titled “2026-08-17”qwen-3-8-27b– Qwen 3.8 27B chat model published by Multiverse Computing. See Qwen 3.8 27B and the models catalog.
2026-08-05
Section titled “2026-08-05”quasar-438b– Quasar 438B chat model with tool calling and structured output viaPOST /v1/chat/completions. Currently in beta; capabilities are identical to GLM 5.2. See Quasar 438B and the models catalog.
2026-07-15
Section titled “2026-07-15”glm-5-2– GLM 5.2 chat model with tool calling and structured output viaPOST /v1/chat/completions. Currently available for functional validation only; performance KPIs are not guaranteed during this beta period. See GLM 5.2 and the models catalog.- GitHub Copilot – Integration guide for using CompactifAI models directly inside GitHub Copilot via an OpenAI-compatible custom provider. See GitHub Copilot.
carina-60b– Carina 60B chat model. See Carina 60B and the models catalog.
Removed
Section titled “Removed”llama-4-scout,cai-llama-3-3-70b-slim,llama-3-3-70b,llama-3-1-8b,blackstar-10b,gpt-oss-20b,whisper-large-v3,glm-5-2– These models have been removed. Migrate to an alternative model from the models catalog.
Deprecated
Section titled “Deprecated”2026-05-06
Section titled “2026-05-06”nemotron-3-nano-omni– Nemotron 3 Nano Omni multimodal model (text and image inputs via chat completions). See the models catalog and Multi Modality.- Cursor and OpenCode – Integration guides for using CompactifAI chat completions, custom model IDs, and the OpenAI-compatible base URL: Cursor and OpenCode.
Improvements
Section titled “Improvements”hypernova-60b– Upgraded to version 2605.
2026-04-01
Section titled “2026-04-01”Removed
Section titled “Removed”DeepSeek R1 0528 Slimhas been removed.
2026-03-24
Section titled “2026-03-24”Removed
Section titled “Removed”POST /v1/usage/completions(usage statistics API) – This endpoint is deprecated and removed from the API. Usage and billing visibility is now provided through the CompactifAI Dashboard; the route is no longer served.
Migration
Section titled “Migration”- View usage and billing in the CompactifAI Dashboard instead of calling the API. Use the dashboard to monitor consumption, manage API tokens, and review account settings.
2026-03-10
Section titled “2026-03-10”- Whisper Streaming Support – The Whisper audio transcription endpoint now supports streaming, enabling real-time transcription and lower-latency audio processing.
- GLM-5 (Private Preview) – GLM-5 is now available as a private model in the US region.
- Agentic Tool Calling Enhancements (Beta) – Improved support for agentic workflows and tool-calling capabilities across the following models:
- gpt-oss-20b
- gpt-oss-120b
- hypernova-60b
Improvements
Section titled “Improvements”- Deterministic Tool Usage in Agent Loops – When
tool_choice=requiredortool_choice=<tool_name>is specified, the model consistently returns a tool call, improving reliability in agent-based workflows. - Improved Tool Call Reliability – Reduced tool-call hallucinations and improved adherence to defined tool schemas.
Performance
Section titled “Performance”- Improved inference throughput across several models, delivering ~30% higher throughput and better overall serving efficiency for:
- cai-llama3-1-8b-slim
- hypernova-60b
- gpt-oss-120b
- gpt-oss-20b
- blackstar-10b
Bug Fixes
Section titled “Bug Fixes”- Fixed an issue where streaming with tool calling was not supported for some models. The following models now fully support streaming responses with tool calls:
- gpt-oss-20b
- gpt-oss-120b
- hypernova-60b
2026-02-25
Section titled “2026-02-25”- Added tool calling support for
hypernova-60bmodel
2026-02-16
Section titled “2026-02-16”Bug fixes
Section titled “Bug fixes”- Fixed a bug where Audio Transcriptions endpoint was not working for all the file mime types specified in our API Reference.
- Significantly improved the performance of the Audio Transcriptions endpoint using the
whisper-large-v3model, reducing latency and increasing the speed factor from 15x to 100x on a 10 minutes long audio file (The speed of your network connection might affect the speed factor). - Fixed a bug where the audio transcription endpoint was not working for audio files with a size greater than 1MB. Now, the endpoint can process audio files up to 25MB in size.
2026-01-08
Section titled “2026-01-08”- Added tool calling support for
gpt-oss-20bandgpt-oss-120b
2025-12-23
Section titled “2025-12-23”- Added
hypernova-60b
2025-12-26
Section titled “2025-12-26”- Added
blackstar-10bmodel
2025-10-06
Section titled “2025-10-06”- Speech-to-text transcription endpoint
/v1/audio/transcriptionswith Whisper Large V3 support for multilingual transcription workflows. - Feature and API documentation detailing request parameters, Python examples, and guidance for the new speech-to-text capability.
Models Updates
Section titled “Models Updates”- Removed the
deepseek-r1-0528model from the API.
2025-09-24
Section titled “2025-09-24”- Multi-modality support for chat completions, enabling image-plus-text inputs across the API.
Models Updates
Section titled “Models Updates”- Added
mistral-small-3-1model with full multi-modal understanding and refreshed usage examples.
2025-08-18
Section titled “2025-08-18”- Function tool compatibility has been activated in all models except mistral.
2025-07-01
Section titled “2025-07-01”Models Updates
Section titled “Models Updates”- Added deepseek-ai/DeepSeek-R1-0528 model, accessible via the
deepseek-r1-0528model ID. - Deprecated
deepseek-r1.
2025-06-11
Section titled “2025-06-11”- Initial release of the CompactifAI inference API with the following features:
- Models API endpoint for listing and retrieving available compressed models
- Chat Completions API endpoint for conversational interactions
- Completions API endpoint for text generation
- OpenAI-compatible API design for easy migration and integration
Models Updates
Section titled “Models Updates”- Added the following models:
cai-llama-3-3-70b-slimcai-mistral-small-3-1-slim
Security
Section titled “Security”- HTTPS encryption for all API requests
- Secure authentication using Bearer token scheme