AIVAX
EN
English Português
Copy page
View as Markdown Edit English source

Changelogs

Technical changes that affect AIVAX products, services, or the public API. Dates identify when entries were added or updated, not confirmed production rollout dates. Each item identifies the affected product or service; maintenance with no user-facing effect is omitted.

Monday, October 19th, 2026 #

Breaking changes:

  • Gateways — One time zone per gateway. AI Gateways now have a timeZone parameter (IANA identifier, default America/Sao_Paulo) used for every date the model reads or writes: get_date_time, scheduling tools, and the bash environment. It replaces builtinFunctionsOptions.dateTimeTimeZone (previous default America/Los_Angeles) and bashOptions.timeZone (previous default UTC), which are no longer read. Gateways and Batch workflows that relied on the old defaults now report São Paulo time; set timeZone on the gateway to keep another zone. In /v1/chat/completions, send time_zone instead of builtin_tools.options.dateTimeTimeZone to override the zone for one request. In the dashboard, the setting moved to Date and time in the gateway’s Parameters + Tuning tab and was removed from Batch tool options. See Current Date and Time.

Fixes:

  • Chat clients — Scheduled messages no longer run alongside a reply. In Telegram and WhatsApp integrations, a scheduled message that becomes due while the assistant is still answering the same conversation now waits for that answer to finish instead of running a second reply in parallel. Messages received during a reply also wait and are answered afterwards. Previously, both runs could act on the same conversation, and one reply could be dropped from its history. A delayed scheduled message is still sent if the conversation becomes free within 30 minutes of its scheduled time.

Changes:

  • Chat clients — Scheduling uses the gateway time zone. schedule_assistant_message now reads times without a UTC offset in the gateway time zone, create_recurring_job evaluates cron expressions in that zone, and list_scheduled_jobs returns scheduled times in it. They previously used the server time zone. Times with an explicit UTC offset keep their exact instant. get_date_time is now always available alongside the scheduling tools. Set the gateway timeZone to your users’ time zone so reminders fire at the expected local time. See Current Date and Time.

Sunday, October 18th, 2026 #

Changes:

  • RAG — More details in the collection documents list. The documents table of a collection now shows the creation date, the last update date, and the approximate token count of each document, and scrolls horizontally when the columns do not fit. Rows with more than three tags show the first three and a +N more badge; hover over it to see the remaining tags. GET /api/v1/collections/{collection-id}/documents now also returns createdAt and approximateTokenCount for each document. See Collections.
  • RAG — Formatted documents in the search playground. The collection search playground renders the content of returned and referenced documents as Markdown, in the same layout used by conversation messages, and adds a Read docs button next to View code that opens the semantic search documentation.
  • Account — Custom plan badge. Accounts on a Custom plan now show a plan badge with the same size and style as the Free, Pro, and Max badges.

Saturday, October 17th, 2026 #

Changes:

  • Batch — Cost-optimized routing for workflows that do not use an AI Gateway. Workflows that process items with an integrated model instead of an AI Gateway now always use the Cheapest routing preference, selecting the available provider with the lowest input and output token price for that model. The Balanced default no longer applies in this case, so items may route to a different provider than before. To keep the default preference or choose Fastest or Quality, use an AI Gateway with that routing preference in the workflow. See Batch.

Friday, October 16th, 2026 #

Fixes:

  • Gateways — Blocked read_skill calls keep the active skill. When a gateway hides tools without a skill, a read_skill call blocked by a gateway worker or rejected as invalid no longer removes the tools of the skill activated by the previous successful call. Only successful read_skill calls change the active skills. See Skills.
  • Built-in tools — Current date and time returns JSON. get_date_time now returns the documented JSON object with date, time, day_of_week, time_zone, utc_offset, and date_time fields. It previously returned a text summary with a different layout. Instructions or workers that parse the previous text should read these fields instead. See Current Date and Time.

Changes:

  • Account — Self-service Pro and Max subscriptions. The My account page now shows the subscription state under the current plan and offers Subscribe to plan on Free accounts, Cancel subscription and Change plan on paid accounts, and Resubscribe while a cancellation is pending. Plan selection shows Pro and Max side by side with their monthly price, main resources, and the amount due today or the date a downgrade starts. Plans are purchased for the calendar month: a first purchase or upgrade is charged now for the remaining days of the month, minus the unused part of what was already paid for the current month, and applies as soon as the payment is confirmed; a downgrade or cancellation takes effect at the next renewal on the first day of the month. Renewals are charged automatically on the first day of each month using the saved payment method; while a renewal is unpaid, a reminder is sent every two days until the tenth day, after which the account returns to Free. Custom plans are managed by the AIVAX team and cannot be changed from the page. The same operations are available through GET /api/v1/subscriptions, POST /api/v1/subscriptions/plan, POST /api/v1/subscriptions/cancel, and POST /api/v1/subscriptions/resubscribe. See subscription pricing and Plans and limits.
  • Account — Saved payment methods. Paid accounts can open Payment methods from the My account page to add, remove, or set the default card used for automatic renewals, or list them through GET /api/v1/subscriptions/payment-methods.
  • Account — Credits and subscriptions are paid through Mixtic. The Add credits dialog no longer asks for a payment method: credit purchases and subscription charges use Mixtic payment links or the saved payment method, and the fees shown before payment reflect this provider. POST /api/v1/payments and GET /api/v1/payments now ignore the method field and query parameter; unknown methods are no longer rejected. Unpaid invoices created in the last 24 hours are also checked with Mixtic every few minutes, so a paid invoice is confirmed even when the payment notification does not arrive.
  • Analytics — Invoice type, payment link, and receipt. The Invoices tab now shows whether each invoice is a credit purchase or a subscription payment, a Pay link while an invoice is open and not expired, and a Receipt link once it is paid. GET /api/v1/information/balance returns the same data in the type, paymentLocation, and receiptLocation fields of each invoice. Unpaid subscription invoices are removed when they expire and can no longer be paid.

Thursday, October 15th, 2026 #

Breaking changes:

  • Models — Retired models removed. @anthropic/claude-4-sonnet, @anthropic/claude-3-haiku, @kwaipilot/kat-coder-pro-v2.5, @kwaipilot/kat-coder-air-v2.5, @nex-agi/nex-n2-pro, @nex-agi/nex-n2-mini, @qwen/qwen3-max, @qwen/qwen3-coder-plus, and @unbiased/pareto are no longer available because they were retired or are no longer served upstream. Switch to a current model of the same family, for example @unbiased/pareto-26.9 instead of @unbiased/pareto.
  • Transcription — Qwen3 ASR Flash removed. qwen/qwen3-asr-flash-2026-02-10 is no longer available because it has no active provider. Use qwen/qwen3-asr-0.6b or another transcription model.
  • Models — @openai/gpt-5.4-pro now uses the Pro model. The model previously ran and was billed as GPT-5.4. It now routes to GPT-5.4 Pro and is billed at $30 input and $180 output per 1M tokens. Review budgets for gateways that use it.

Changes:

  • Models — New text models. Adds @meta/muse-spark-open, @nvidia/nemotron-3.5-super-vl, @openai/gpt-6.1-sol:pro, @qwen/qwen3.8-omni-flash, @inclusionai/ling-3.1-flash, @stepfun/step-5-preview (preview), @sakana/fugu-ultra-v2, and @bytedance/seed-2.1-turbo. See the models page of the dashboard for capabilities and pricing.
  • Models — Deprecated models. Gemini 2.5 Pro, Flash and Flash-Lite, Xiaomi MiMo V2.5 and V2.5 Pro, Qwen3.5 9B, Poolside Laguna XS 2.1 and S 2.1, GLM 4.7, Claude 4.5 Sonnet, Claude 4.1 Opus, GPT-5, GPT-5 mini, GPT-5 nano, GPT-5.1, GPT-5.3 Codex, GPT-5.4 nano, Gemini 3 Flash, Gemini 3.5 Flash, and Gemini 3.7 Flash are now marked as deprecated. Some of them are scheduled to be retired by their providers in the coming weeks, so move to a newer model.
  • Models — Provider and release date updates. @minimax/m2.7 and @qwen/qwen3.8-max now route to their current provider versions, @openai/gpt-6-astra:pro uses the dedicated Pro version, and several models received corrected or newly added release dates. Groq no longer serves some Llama, Qwen, and Kimi models; those models stay available through other providers.
  • Image generation — New models. Adds flux-3-image, nanobanana-2.1, seedream-5-flash, and qwen-image-2.1-turbo. mai-image-2.5-flash and nova-canvas are deprecated; mai-image-2.5-flash is now an alias of mai-image-2.6-flash upstream, and nova-canvas is no longer listed by its provider.
  • Transcription — New models. Adds nvidia/nemotron-3.5-asr-streaming-multilingual-0.6b and qwen/qwen3-asr-0.6b. The x-ai/grok-stt-1.0 display name is now “SpaceXAI: Grok STT 1.0”; its identifier and default status are unchanged.
  • Speech synthesis — ElevenLabs Flash v2.5 and Fish Audio Drama 3. Adds ElevenLabs Flash v2.5 (ElevenFlash) and Fish Audio Drama 3 (FishDrama3) as speech synthesis sources. Drama 3 accepts speech direction in square brackets, such as [whispering], and uses the provider’s default voice when none is requested.
  • Re-ranking — Voyage Rerank 3. Adds @voyageai/rerank-3 ($0.05 per 1M tokens) and @voyageai/rerank-3-lite ($0.02 per 1M tokens), both with a 32K context.
  • Semantic decisions — New models and price updates. Adds @nace-ai/drex-v2.0, @upstage/solar-decide-flash, @scaledown/classify-1, and @inception/mercury-decide. @liquid/d1 input price drops from $0.04 to $0.02 per 1M tokens, while @upstage/solar-decide rises from $0.05 to $0.10 per 1M tokens. @cloudflare/clef-flash context is now 24K and @typesafe/jev context is now 262,144 tokens.

Wednesday, October 14th, 2026 #

Fixes:

  • Gateways — Skill tools visible when tools without a skill are hidden. When a gateway hides tools without a skill, the tools listed in a skill’s allowedToolsNames now become available after the model activates that skill with read_skill. Previously, only the always-visible tools and read_skill were offered. Only the skills from the most recent successful read_skill call count as active, so a call blocked by a gateway worker or rejected as invalid does not expose the skill’s tools. See Skills.

Tuesday, October 13th, 2026 #

Changes:

  • Models — Supported reasoning efforts. GET /api/v1/information/models.json now returns modelInformation.reasoningEfforts for each model, listing the reasoning effort levels the model accepts, such as low, medium, high, xhigh, or max. The list is empty when the supported levels are unknown. The model details dialog on the dashboard Models page shows these levels when the model has at least one.

Saturday, October 10th, 2026 #

Changes:

  • RAG — WebSocket JSONL import. A new Index Documents (JSONL WebSocket) endpoint replaces the multipart JSONL upload for new integrations. Clients stream UTF-8 JSONL lines over a WebSocket and finish with an END line. Every 100 lines are checked against balance, storage, and collection access, saved, and queued for indexing, with progress messages reporting committed lines. Each import accepts 10× the plan’s per-request JSONL line limit; the other insertion limits are unchanged. Each message must arrive within 30 seconds, and only one import per collection runs at a time. The multipart Index Documents (JSONL) endpoint remains available but is deprecated. See Batch Import.
  • API — Consumed-Credits response header documented. Responses that charge usage include Consumed-Credits with the request’s total cost in USD, including the plan multiplier. It is omitted when nothing was charged and does not cover streamed responses or usage recorded after the response. See Request cost header.
  • Account — Zero data retention providers only. The account settings page has a new Zero data retention providers only option, disabled by default. When enabled, text inference with integrated models is routed only to providers that do not retain prompts, and other providers are excluded. Requests fail when the model has no zero data retention provider available. The option does not change conversation logging. See Zero data retention providers only.
  • Semantic decisions — Perplexity Decider V1.1, price and limit updates. @perplexity/pplx-decider-v1.1-27b replaces Perplexity Decider V1, which is no longer available upstream; the previous name @perplexity/pplx-decider-v1-27b keeps working and now resolves to V1.1. Its input price drops from $0.04 to $0.02 per 1M tokens. @cloudflare/clef-flash input price drops from $0.09 to $0.038 per 1M tokens, and the @typesafe/jev-1.13 context limit increases from 32,768 to 64,000 tokens. See Decisions.
  • Inference — allow_logging per request. Chat completions accept allow_logging. Set it to false to keep that request out of the account conversation logs; it defaults to true and cannot enable logging when it is disabled for the account. See Response and conversation records.

Friday, October 9th, 2026 #

Breaking changes:

  • Chat clients — Local date is no longer added to web chat conversations. Messages sent from the hosted web chat no longer add a system message with the user’s local date and time zone at the start of the conversation, and the public chat client endpoint ignores the userLocalTime and userTimeZone fields. This keeps the start of the conversation stable between turns, which improves prompt caching. Gateways that need the current date or time should enable the DateTime built-in tool, which returns the current date and time in the configured time zone when the model calls it. See Current Date and Time.

  • Gateways — System instructions follow declaration order. The parts of an AI gateway’s system instructions are now combined in the order they are declared: the gateway’s own instruction first, followed by MCP server instructions, remote system instruction sources in their configured order, and instructions added during the request. Previously, these parts could appear in reverse order, and remote sources were added in the order they finished loading. Identical requests now produce the same system instructions, which improves prompt caching. Gateways whose prompts depend on the previous relative order should review their instructions.

  • Built-in tools — Calendar removed. The Calendar flag and appointment_create, appointment_search, appointment_find, appointment_update, and appointment_delete tools are removed from the API and dashboard. Existing gateways and Batch workflows have the flag removed automatically. Requests containing Calendar in builtin_tools.tools now fail; remove the flag and update instructions that depend on appointment tools. Migrated reminder text is not a replacement calendar or scheduling service.

  • Memory — Tool calls and management changed. Remember now exposes only memory_save and memory_search. Replace memory_update and memory_remove calls with memory_save using an ID and content, or an ID without content to delete. memory_clear is removed. The Memories API (/api/v1/memories, including list, get, delete, and migration prompt generation) and dashboard Memories page are removed; manage records through the Collections and Documents APIs and dashboard pages. Deleting @memories erases all users’ memories in the account.

  • Gateways — Memory retrieval and visibility changed. Memories are no longer added to system instructions automatically: update prompts to call memory_search explicitly. The include_all_memory_context / IncludeAllMemoryContext option is removed and ignored if sent. allowSharedMemory keeps controlling whether a gateway sees the user’s memories from other gateways or only its own; migrated memories have no gateway and are visible only to gateways with shared memory. Expiration is now optional: memory_save accepts expiresInDays (1–365), and memories saved without it never expire, where the old default was 30 days. Review access assumptions and retention needs. See Memory.

Changes:

  • Models — Three new semantic decision models. Adds @microsoft/microsoft-decision-1 (Microsoft-Decision-1), @nace-ai/drex-v1.5 (Nace.AI Drex v1.5), and @cloudflare/clef-omni (Cloudflare Clef Omni) to the semantic decision catalog. All three support noul, choice, and score questions and are billed per input token at $0.042, $0.04, and $0.15 per million tokens respectively, with no output token charge. They are not covered by the daily semantic decision allowance. See Semantic decisions and Pricing.
  • Semantic decisions — Image input. Image-capable decision models (@cloudflare/clef, @cloudflare/clef-flash, @cloudflare/clef-omni, and @openai/gpt-6-luna-decisions) now accept images in state as image_url parts with base64 image/png, image/jpeg, or image/webp data URLs, in the same format as OpenRouter. Clef models accept up to 4 images and GPT-6 Luna Decisions up to 128; other models reject images with 400. GET /api/v1/information/decisions-models.json now returns maxImages for each model. See Send images.
  • Memory — Collection storage and search. Memories are stored as documents in the account’s @memories collection, created on first save, with user and conversation metadata. memory_search accepts exactly one of a semantic query or a document filter, returning up to 10 results restricted to the current user. Semantic searches use no-cost rrf reranking; filter searches return newest matches first without embedding charges. Saves and updates incur document-embedding charges, semantic searches incur query-embedding charges, and both tools use the account’s RAG rate limits and return tool errors when balance or rate limits prevent access. Allow a few seconds for indexing after writes. See Memory and Pricing.
  • Memory — Existing records migrated. Non-expired memories moved to @memories, preserving their IDs, user identifiers, creation dates, and expiration dates, with conversation_token set to null. Non-expired calendar reminders became text memories in the form Reminder at <date> (<n> minutes): <description>. Migrated records are re-indexed and incur document-embedding charges. Review the collection and update integrations that used the removed memory or calendar features.

Thursday, October 8th, 2026 #

Changes:

  • Documentation — New Learn section. The documentation site now includes Learn, a set of 11 modules and 63 short units that explain AI agents from first concepts to production for readers without a computing background: what an agent is, language models, prompts and context, model selection and parameters, tools and integrations, knowledge and RAG, advanced workflows, evaluation and observability, safety and compliance, cost and scale, and practical case studies. Units include interactive examples, step-by-step walkthroughs, comparison tables, charts, and short quizzes, and each one can be marked as complete in the browser. Three suggested learning paths (Beginner, Developer, Business) group the modules by goal. Learn is available in English and Portuguese and is included in site search.
  • Models — Claude Haiku 5.5 added. Adds @anthropic/claude-5.5-haiku (Claude Haiku 5.5), the direct successor to Claude Haiku 4.5, to the text model catalog. It supports image and file input and tool calling, with a 1M-token context window. The @model-router/claude:budget alias now selects Claude Haiku 5.5, replacing Claude Haiku 4.5. Applications using this alias may see changes in response quality, latency, and cost. Existing explicit model identifiers remain unchanged.
  • Inference — Provider selection with routing_options. Chat completions accept routing_options with preset (Balanced, Cheapest, Fastest, or Quality) and allowed_providers, an ordered list of provider tags such as ["azure-us", "azure-eu", "*"]. AIVAX tries each tag in order and moves to the next only when no matching provider is available; "*" matches any provider. allowed_providers defaults to ["*"], and an empty list is rejected. Without "*", the request fails when none of the listed providers is available. routing_preset is deprecated: it still works, but use routing_options.preset instead. AI gateways accept the same ordered list in parameters.allowedProviders (default ["*"]), editable in the gateway editor under Allowed providers for integrated models; request values override it for that request only. See Provider routing.
  • Inference — Serving provider in responses. Chat completion responses and every streamed chunk include a provider field next to model with the tag of the provider that served an integrated model. It is null for gateways that use your own provider credentials.
  • Models — filter on GET /v1/models. The model list accepts an optional filter query parameter with a model name and returns only matching entries: the exact name or its dated snapshots, ordered from the closest match. Without it, the list is unchanged.
  • Models — Provider tags. The provider details on the dashboard Models page show each provider’s tag with a copy button, and GET /v1/models returns it as tag in each provider entry. Provider endpoints in different regions or variants have distinct tags, such as azure-us and azure-eu.

Wednesday, October 7th, 2026 #

Breaking changes:

  • Inference — File and OtherFiles pre-processing now cover every file. Gateways and requests that enable only the OtherFiles multimodal flag now convert PDF files with OCR instead of sending them to the main model unchanged. Those that enable only the File flag now convert non-PDF files with OCR as well, also billed in Processing Units. Gateways and requests with both flags or All are not affected. To leave files unchanged, set multimodal_resolver (or the gateway’s multimodalResolverParameters) without a fileEngine; no file type is then pre-processed.

Changes:

  • Inference — Per-media multimodal resolver engines. Chat completions accept multimodal_resolver, and AI gateways accept multimodalResolverParameters, with imageEngine, audioEngine, videoEngine, and fileEngine fields. Each one selects how that content type is converted to text before inference: InferenceLow (smaller multimodal model; Inference is an alias), InferenceHigh (larger, more accurate multimodal model at a higher cost), Ocr for images and files (billed in Processing Units, like Fetch and OCR), or Stt for audio (speech-to-text, billed per second). OCR accepts base64 data URIs and public URLs. Results of every engine are cached per content and engine, so repeated media is not billed again. The multimodal_preprocess flags and the gateway enabledMultimodalFeatures setting are deprecated but keep working with the equivalent engines; the new settings take precedence when present. The gateway editor now configures one engine per content type. See Multimodal pre-processing.
  • RAG — Document filters in the Collections MCP and the gateway query tool. The search tool of the Collections MCP and the query tool of AI gateways using the QueryFunction query strategy accept an optional filter argument with the same syntax as the semantic search filter field, such as tags has "faq" and updatedAt >= now-30d. Filters are applied before the search terms are embedded; when no document matches, the tool returns no results without embedding or search charges. An invalid filter is returned to the model as a tool error. Automatic gateway RAG still does not apply filters. See Document Filters.

Tuesday, October 6th, 2026 #

Changes:

  • RAG — Faster text segmentation without sanitization. POST /api/v1/generations/segment requests with sanitize omitted or false no longer use a language model. Boundaries are now chosen from the document structure (headings, lists, tables, code blocks, paragraphs, and sentence ends) and the semantic similarity of neighboring passages, targeting segments of about 300 tokens. Segments cover the whole document in source order without overlap and keep the original line breaks; a boundary may fall at a sentence end inside a line. Documents of about 300 tokens or less are returned as a single segment. The response format, quotas, and per-token price are unchanged; usage.processing_units now reports the tokens of the submitted documents. Requests with sanitize: true keep the previous behavior. See Text segmentation.

  • Models — Six new semantic decision models. Adds @upstage/solar-decide (Upstage Solar Decide), @cloudflare/clef (Cloudflare Clef), @cloudflare/clef-flash (Cloudflare Clef Flash), @liquid/d1 (LiquidAI d1), @perplexity/pplx-decider-v1-27b (Perplexity Decider V1 27B), and @openai/gpt-6-luna-decisions (OpenAI GPT-6 Luna Decisions) to the semantic decision catalog. All six support noul, choice, and score questions through POST /api/v1/generations/decisions and are billed per input token at published rates with no output token charge, with existing account and plan adjustments still applying. They are not covered by the daily semantic decision allowance, which remains limited to Julia-1. Existing model identifiers remain unchanged. See Semantic decisions and Pricing.

Monday, October 5th, 2026 #

Breaking changes:

  • Chat clients — Current date and time are no longer added to the instructions. Telegram and WhatsApp integrations, including scheduled messages, no longer add the server date and time to the model instructions. Models must get the current time from a tool. To keep date-aware behavior, such as reminders, enable the date and time built-in function or the shell (date command) on the AI gateway used by the chat client.
  • Account — Reseller plan renamed to Custom. Accounts on the Reseller plan now report the plan as Custom instead of Reseller in API responses, such as the balance endpoint, and in the X-Authenticated-Account-Plan response header. Integrations that compare the plan name with Reseller must accept Custom. Custom accounts can have rate limits defined as a multiple of the Pro plan limits; accounts without a configured multiple keep having no plan rate limits. See Plans and Limits.

Fixes:

  • AI gateways — Shell tool commands without parameters now run. In the gateway shell, calling a tool that has no parameters, such as list_scheduled_jobs, now runs it instead of printing its help. Use --help to view the help.

  • Collections — De-duplication no longer stops before removing documents. A de-duplication task that found duplicate documents could stop with zero documents removed and no backup files, because saving the backup of the removed documents failed. These tasks now save the backup files, remove the duplicates, and include the download links in the completion notification. Tasks that stopped this way did not remove any documents; start a new task to retry.

Changes:

  • Conversations — Provider and routing details. Conversations now record the provider that served the latest response, its data collection policy, the model variant (fast, flex, or priority) when one was used, and the routing option used to select the provider. The View Conversation endpoint returns them in the providerName, providerDataCollection, modelVariant, and routingOption fields, and the List Conversations endpoint returns providerName, modelVariant, and routingOption. These fields are null for gateways that use your own provider credentials and for conversations stored before this change. The conversation list in the dashboard shows Provider, Routing, and Variant columns; Provider shows BYOK when no provider was recorded. The table scrolls horizontally, and the row actions stay visible.
  • AI gateways — Time zone and culture for the shell. The gateway shell options accept timeZone, an IANA time zone such as America/Sao_Paulo, and culture, such as pt-BR. The date command uses the time zone for its output and for dates without an explicit offset, and uses the culture for day and month names. The defaults remain UTC and the invariant culture; date -u always prints UTC.
  • RAG — Document filters for semantic search and answer generation. The semantic search and answer generation endpoints accept an optional filter field that restricts the search to documents matching conditions on name, content, tags, creation and update dates, or metadata, such as tags has "finance" and createdAt >= now-30d. The field accepts a string or an array of strings combined with and. Filters are applied before the search terms are embedded; when no document matches, the request returns an empty result without embedding or search charges. Invalid filters return 400 Bad Request with the error position. Filters are not yet available in AI gateway RAG or the Collections MCP. See Document Filters.
  • RAG — Filters and visual results in the collection playground. The collection playground has a Filters section where each line is a document filter; all lines must match and are sent as the filter array. Results can be viewed as Visual, showing the generated answer and each document with its score, metadata, and referenced documents, or as raw JSON. The playground now uses the rrf reranker by default; you can still choose another reranker.

Saturday, October 3rd, 2026 #

Changes:

  • Documentation — Updated reading and search experience. Documentation adds language-specific search, light and dark themes, mobile navigation, and a page menu for viewing or copying Markdown. AI agents can use the documentation index and full-text files. Existing documentation URLs continue to resolve; some older guides redirect to their current product documentation.

Friday, October 2nd, 2026 #

Changes:

  • Models — Claude Sonnet 5.5 and GPT-6.1 Sol added. Adds @anthropic/claude-5.5-sonnet (Claude Sonnet 5.5), the direct successor to Claude Sonnet 5, and @openai/gpt-6.1-sol (GPT-6.1 Sol), an upgrade to GPT-6 Sol, to the text model catalog. Both support thinking, image and file input, and tool calling; GPT-6.1 Sol also supports structured output. The @model-router/claude:mid and @model-router/openai:mid aliases now select Claude Sonnet 5.5 and GPT-6.1 Sol, replacing Claude Sonnet 5 and GPT-6 Sol, respectively. Applications using these aliases may see changes in response quality, latency, and cost. Existing explicit model identifiers remain unchanged.

Thursday, October 1st, 2026 #

Breaking changes:

  • API — Error status codes reflect the failure source. Failures are no longer all returned as 400 Bad Request by account, RAG, web, and generation endpoints. Invalid requests still return 400, and authentication, balance, and permission failures that previously surfaced as 400 now return 401, 402, or 403. Unexpected AIVAX failures return 500 Internal Server Error with a generic message, and unavailable external services, including model providers, return 503 Service Unavailable with a Retry-After header. On OpenAI-compatible endpoints, the error code is now invalid_request_error, service_unavailable, or server_error instead of always server_error. Clients that treat every non-2xx response as a request error should retry 500 and 503 responses, honoring Retry-After. See Troubleshoot the first request.

Fixes:

  • Chat completions — Malformed tool declarations rejected as invalid requests. Requests whose tools entries have missing or wrongly typed fields now return 400 Bad Request instead of a server error.

  • Gateways — Bash tool reports invalid options to the model. Unknown options or values that do not match a tool’s parameters now return a command error that the model can correct, instead of failing the tool call.

Changes:

  • MCP utilities — Media generation MCP. A new hosted MCP server at https://inference.aivax.net/v1/mcp/media-generation exposes list_models, generate_image, and generate_speech to MCP-compatible clients. list_models takes a type of image or audio; generated images and MP3 speech are returned as public URLs. Use the optional X-Mcp-Enabled-Tools header to choose which tools a client sees. Generations use the same pricing and limits as the Image Generation and Speech Generation APIs and require a positive balance. See Media generation MCP.

Monday, September 28th, 2026 #

Fixes:

  • Gateways — Bash tool help accepts nullable parameters. Requesting help for tools whose parameters accept multiple types, including null, no longer fails while listing their arguments. Help preserves the accepted types and nested parameter paths. No tool schema changes are required.

  • Chat integrations — Failure notifications restored. Streaming and non-streaming conversations again attempt to send “System: something went wrong. Please, try again later.” after an unrecoverable generation failure, exhausted recovery attempts, or a turn that sends no message. A final failure is reported even if an earlier partial reply was delivered. Notification delivery still depends on the messaging service being available.

Changes:

  • Gateways / MCP — Server instructions and remote skills. MCP sources can include server instructions and root skill documents alongside tools, including sources added by workers. Both options default to enabled; set allowClientInstructions or allowRemoteSkills to false to exclude that content. Skill discovery requires the remote server to advertise compatible capabilities; account skills remain available. Remote skill content is checked against its advertised size, digest, and frontmatter. Dynamic skills, supporting files, scripts, and servers requiring the newer discovery protocol are not supported. See MCP functions for compatibility and trust limits.

  • Gateways — Time zone selection. The current date and time setting now offers a time zone dropdown grouped by region, including UTC. Existing saved time zones are preserved when reopening the setting.

  • Models — Seven new text models. Adds @cohere/command-a-plus, @upstage/solar-mini4, @aion-labs/aion-3.5, @aion-labs/aion-3.5-mini, @qwen/qwen3.8-max-prime, @z-ai/glm-5.3-prime, and @fireworks/ember-1 as selectable text models, billed per provider at published token rates with existing account and plan adjustments still applying. Upstage and Fireworks models now display their provider icons instead of the generic fallback. Existing model identifiers remain unchanged.

Sunday, September 27th, 2026 #

Changes:

  • Telegram — Compact tool progress. Streamed replies show only the latest tool preamble in the thinking indicator when tool-call visibility is enabled, instead of accumulating tool-name blocks in the response. The final reply contains no tool-progress blocks. Other messaging channels and non-streamed replies are unchanged.

  • Gateways — Bash tool selection. The Bash tool list now includes a shortcut for get_date_time (Current date and time). Inclusion and exclusion lists accept case-insensitive wildcard patterns: * matches any number of characters, as in something_*, and ? matches one character. Exact tool names remain supported.

  • Semantic decisions and Agentic Tests — Account rate limits. Per-minute account limits apply to semantic decision requests and new Agentic Test runs; manual, scheduled, and direct evaluations share the test-run limit. Requests above the limit return HTTP 429, while scheduled tests wait for a later scheduling check. Clients should pace requests and retry after the rate-limit window clears. Existing inference limits still apply. See Plans and Limits, Semantic decisions, and Agentic Tests.

  • Models — Consistent speech synthesis prices. Text-to-speech catalog prices now derive from the same per-character rates used to calculate synthesis usage, displayed per 1,000 characters. Existing model identifiers and billing rates remain unchanged.

  • Models — Comparable transcription prices. Speech-to-text model prices are displayed consistently in USD per minute, converting hourly and per-second rates for comparison without changing billing rates or duration measurement.

  • Image generation — Fixed output and reference prices. Image generation uses a fixed price per delivered output plus a per-reference price for each output. Models whose providers bill tokens or megapixels now use rounded-up estimates instead of metered token charges. Prompt processing is included in the output estimate; models without a separate reference charge list a zero reference rate. These are fixed tariffs, not receipts for the provider’s actual consumption. Image charges no longer include the AIVAX image-generation markup or account and plan pricing multipliers. See Image generation.

  • Models — Subscription coverage. The Models page now includes a “Subscription Usage” column for rerankers and semantic decision models. “Included” identifies models eligible for daily plan allowances; “-” identifies models without coverage. Eligibility does not indicate an account’s remaining allowance. Both service catalogs expose this eligibility as subscriptionUsage.

  • Subscriptions — Included daily usage allowances. Free, Pro, and Max subscriptions include separate daily allowances for RAG search and insertion embeddings, Reflex input (including cached input), and Julia-1 semantic decisions. Each metered item is either fully included or billed at normal rates; a request with multiple items can combine included and paid usage. This also applies to OCR extraction, replacing partial coverage. Included items appear in subscription consumption without zero-cost billing-history entries; usage indicators may exceed the base allowance within the permitted margin. Uncovered items retain normal billing records. includesSubscriptionModels is false while inference subscriptions are disabled; LLM subscription coverage remains disabled. Reflex’s daily processing-time cap remains separate, and reseller accounts do not receive subscription allowances. See Plans and Limits and Pricing.

Saturday, September 26th, 2026 #

Breaking changes:

  • Image generation — Deprecated models removed. Removes majicMIX-realistic, AbsoluteReality, CyberRealistic, CyberRealistic-Pony, RealCartoon-Realistic, Hassaku-XL, and Meina-Mix from the available models. Direct API requests using these identifiers now fail; select an active model instead. Built-in image generation uses flux-schnell when no valid model is configured, replacing the deprecated AbsoluteReality default. Review saved configurations; output style and pricing differ.

Changes:

  • Image generation — Additional Pollinations models. Adds 16 official raster-image models, including FLUX 1.1 Pro and FLUX 2 variants, MAI Image variants, GPT Image 2.5 Flare and Sunburst, Qwen Image 2.1 and 3, Grok Imagine Image 2.0, Recraft V4.1 Flash, Krea 2 Medium, DreamShaper 8 LCM, and Seedream 5 Pro, with model-generated previews in the image picker. Community and SVG models are excluded. The catalog displays the billing units. See the September 27th entry for the subsequent fixed-price billing change. Failed generations are not counted as delivered images. See Image generation.

  • Models — Service model catalogs. The Models page now includes tables for image generation, speech-to-text, text-to-speech, reranking, and semantic decisions, with backend-provided descriptions, release dates, base USD pricing with billing units, and an Actions dropdown in every service-model table for copying model names and opening integration documentation. Model names show a friendly label when available while copying the identifier accepted by the API. Pricing uses compact input, cached-input, output, image, character, and duration units separated by arrows where applicable, with full rates and units available in the tooltip. Catalogs are ordered newest first. OpenRouter catalog dates and friendly names supplement missing launch metadata; catalog dates are explicitly labeled rather than presented as manufacturer release dates. Models without either date remain last. Failed catalog requests can be retried independently. The public information catalogs expose these details, including the new GET /api/v1/information/speech-models.json catalog. Account and plan pricing adjustments still apply. See Pricing and Semantic decisions.

  • Privacy — Judicial disclosure and retention clarified. The Privacy Policy and Terms of Use specify Brazilian court orders for disclosure, foreign requests to preserve existing logs for up to 1 year, and up to 1 year of technical logs and metadata. They clarify that conversation content is collected only when Conversations is enabled for the request or account, and that available account resources and backups dating back up to 3 months may be disclosed under a Brazilian court order. The content license in the Terms is expressly subject to these limits.

  • Generations — Additional decision models. The Semantic decisions guide explains question types and response interpretation. The public GET /api/v1/information/decisions-models.json endpoint lists canonical names, aliases, supported question types, context lengths, release dates, and base token prices. Adds @respan/span-01, @respan/span-01-lite, @jaredpalmer/kev-4b, and @supersonic-labs/julia-1 as model choices for semantic decisions. Julia-1 supports choice, score, and noul; its input usage includes the state repeated for each question. Existing account and plan pricing adjustments apply, and model identifiers and request formats remain unchanged. See Plans and Limits and Pricing for current limits and rates.

Tuesday, September 22nd, 2026 #

Changes:

  • Models — GPT-6 Sol and Luna added. Adds @openai/gpt-6-sol and @openai/gpt-6-luna, including their :pro reasoning variants. The @model-router/openai:mid and @model-router/openai:budget aliases now select GPT-6 Sol and GPT-6 Luna, respectively. Applications using these aliases may see changes in response quality, latency, and cost. Existing explicit model identifiers remain unchanged.

Monday, September 21st, 2026 #

Breaking changes:

  • Gateways — Off-topic moderation was removed. The dedicated off-topic threshold no longer blocks requests that stray from the conversation’s purpose. If your application relies on this check, review its topic restrictions before adopting this change; the remaining moderation categories are not an equivalent replacement.
  • Built-in tools — Advanced web research is disabled. The AdvancedWebUsage tool returns an unavailable response instead of performing research. Remove reliance on this tool from gateway instructions and workflows. Standard Web Search and URL extraction remain separate alternatives, not equivalent replacements for multi-step research.
  • Models — Mercury 2.5 model identifier changed. Replace @inception/mercury-2.5-preview with @inception/mercury-2.5 in requests and gateway settings. The preview identifier is no longer listed, and the replacement is no longer marked as preview.
  • Gateways / MCP — MCP tool names are source-qualified. Tools from different MCP sources no longer share an unqualified name in the gateway. Review gateway instructions, tool-selection rules, and workers that match exact tool names. The original tool name at the connected MCP server is unchanged. See MCP functions.

Fixes:

  • Collections, Gateways, and Generations — Service model availability. Collection answer generation, gateway routing, chat utilities, and media-processing services avoid selecting temporarily unavailable models. Teach Skill and multimodal preprocessing can try another available model after a retryable failure; success still depends on service availability.
  • Teach Skill — Usage calculation. Teach Skill processing charges use the model pricing associated with the completed request, including when a retry changes the model used. See Teach Skill and Pricing.
  • Fetch and OCR — More reliable page extraction. Cancelled or timed-out page-content requests no longer leave page extraction running indefinitely. This addresses cases where web content extraction could stall.
  • Chat clients — Final reply and message history. completionText now selects the final generated assistant reply rather than combining it with earlier assistant text during tool use. The added createdMessages response field preserves newly generated messages in order, including tool interactions; submitted messages are not repeated. Use this field when you need the complete generated turn.
  • Gateways / MCP — Structured MCP results are retained. Assistants receive structured result content in addition to supported content blocks, avoiding missing information when an MCP tool returns structured output.
  • Fetch and OCR — X post extraction. Improved readable-text extraction from public X post links. Content availability and access restrictions still apply.
  • Fetch and OCR — Plain-text normalization changed. Plain-text conversion no longer collapses internal whitespace or normalizes Unicode characters. Surrounding whitespace may still be trimmed in extraction results. Applications that compare extracted text exactly or require normalized spacing should perform that normalization themselves.

Changes:

  • Fetch and OCR — Optional structured extraction. Supply responseSchema to convert extracted content into JSON. Results add extractedObject and jsonProcessingUnits while retaining extractedText on success. Without a schema, the new fields are null and zero respectively. JSON conversion is billed separately from extraction and is not covered by the daily extraction allowance. See Fetch and OCR.
  • Generations — Semantic decisions. Evaluate multiple named questions against a shared JSON state using noul (true/false criteria), choice, or score. Responses include named answers, token usage, and cost. The service requires a positive balance, and its token usage contributes to account usage totals.
  • Built-in tools — Current date and time. The DateTime option lets assistants request the current date, time, weekday, time zone, and UTC offset. Set dateTimeTimeZone to the desired time zone; the default is America/Los_Angeles, with daylight-saving adjustments, independent of the browser’s time zone. Invalid time zones are rejected.
  • Models — Additional model choices. Adds GLM-5.3-FlashX, Fugu Max, Pareto, Ling 3.0 Flash VL, MiMo-V2.6-Pro, MiMo-V2.6-Flash, MiMo-V2.6-Pro-UltraSpeed, and Grok 4.7 to the inference catalog. The Xiaomi frontier, mid, and budget aliases now select MiMo V2.6 models, while @model-router/grok:latest selects Grok 4.7. Alias users may see different response quality, latency, and cost; availability and supported capabilities depend on the selected model and account plan.
  • Inference — Transient-failure recovery. Inference requests can make additional recovery attempts when a provider is temporarily unavailable. This may avoid some failed requests, but can also increase response time before an error is returned.
  • Avi Assistant — Updated default model. The AIVAX console assistant changes its default model, which may change response style, latency, and usage cost. This does not change the model selected in your own gateways.
  • Documentation — Updated service guidance. The API documentation overview and Avi Assistant guidance cover voice sessions, agentic tests, classification, segmentation, reranking, and web extraction, with current documentation links and service-specific billing guidance. This is a guidance update, not the introduction of those services.
  • Models — DeepSeek V4.1 Flash added. The inference catalog includes @deepseek/deepseek-v4.1-flash with tool-calling support. Check model availability and plan eligibility before selecting it.
  • Models — Router selections updated. @model-router/deepseek:latest and @model-router/deepseek:budget now select DeepSeek V4.1 Flash. Applications using these aliases may see different response quality, latency, and cost without changing the alias. Adds @model-router/claude:frontier-mythos and @model-router/mercury:latest as additional choices.
  • Chat clients — Structured prompt input. Synchronous chat-client prompts accept plain text, a single message, or an ordered array of messages, including assistant tool calls and matching tool results. Existing single-message input remains supported. Optional instructions adds context for that request without replacing saved session context. See Chat clients.
  • Chat clients — Turns without session persistence. Set commit to false to generate a reply without saving the submitted and generated messages to session history. The default remains true. This is not a free preview: inference and tool actions still run. To continue an uncommitted tool interaction, send the assistant tool-call message together with its tool results.
  • Agentic Tests — Optional testing notifications. Account notification preferences can enable weekly test summaries and alerts when a test reaches three consecutive failures. These complement existing failure and recovery notifications. See Agentic Tests.
  • Fetch and OCR — Rendered web content. HTML extraction supports rendered page content, improving coverage of pages whose readable text depends on scripts. Rendering is metered in processing units; this does not guarantee access to every website or restricted page. See Fetch and OCR.

Type to search the documentation.