You can now create, update, and delete your organization’s routing policies programmatically.
POST /v1/manage/policy creates a policy with a name and a config for a fallback, load-balancing, or latency policy; PATCH /v1/manage/policy/{policy_id} changes its name or configuration; DELETE /v1/manage/policy/{policy_id} removes it. Requests require a Management API key with write permission, and changes take effect on the router immediately.Routing policy endpoints →API key list and detail responses now include
expires_at for keys with an expiry date, so you can audit upcoming expirations without checking each key in the dashboard. Non-expiring keys omit the field.List API keys docs →POST /v1/audio/transcriptions now supports Mistral’s Voxtral models hosted on DeepInfra: deepinfra/mistralai/Voxtral-Small-24B-2507 and deepinfra/mistralai/Voxtral-Mini-3B-2507. Responses include language, duration, segments, and words, and usage is billed on audio duration like other transcription providers. List them with /v1/models/transcription.Transcription API →Spending alert webhooks now carry the spend that triggered the alert. JSON payloads include a top-level
id that uniquely identifies each alert event and a spend field on user, group, and API key alerts; Slack and Teams messages show the current spend next to the limit.Spending Alerts docs →Management API responses are now gzip-compressed when your client sends
Accept-Encoding: gzip. Large responses such as log listings and analytics exports download faster; clients that do not send the header keep receiving uncompressed responses.Management API docs →When a request is rejected because your organization’s data-residency policy requires a different region, the error message now points to the correct regional endpoint,
https://router.eu.requesty.ai/v1.EU Routing docs →Your organization’s in-flight request limit now scales with your tier and your current balance. Tiers are based on lifetime totals and only move up. Requests over the limit return a rate limit error that states whether the organization limit or the per-user limit was reached.In-flight rate limits docs →
The provider rankings now include output speed in tokens per second for every provider and model, in both the live leaderboard and the history view.
Guardrails are no longer limited to enterprise plans. Every organization can now enable PII, PCI, and secret-key scanners from the admin panel, with per-scan pricing shown next to each guardrail before you turn it on. Guardrail charges appear in log costs as before.Guardrails docs →
Azure AI Foundry BYOK keys now accept a custom endpoint URL. Set it on the BYOK page when your deployment lives on a non-standard Foundry endpoint, and the router sends requests to that URL with your key. Rotating the key keeps the stored URL.BYOK docs →
The playground can now send requests through the regional router endpoints (EU, US, AP) as well as the global one. If your organization restricts allowed router regions, the playground only offers the regions your policy permits.EU Routing docs →
/v1/responses now accepts explicit cache_control breakpoints on input items and translates them for Anthropic models. Cache read and cache write tokens are reported in the response usage, so caching works the same way as on Chat Completions and Messages.Manual caching docs →Two new organization-level settings decide what group admins can do: add and remove group members, and manage member spending limits. Org admins control both toggles, so you can delegate day-to-day group management without giving up control over budgets.Groups docs →
Organization financials moved to a dedicated admin-only endpoint:
GET /v1/organization/finance returns balance, promotion and gift state, and auto top-up settings. GET /v1/organization keeps the member-visible metadata and configuration, and GET /v1/organization/info has been removed.Also fixed: the /v1/models endpoints now apply your API key’s model allow-list when the key is sent in the X-Api-Key header, matching the behavior of Authorization: Bearer. Enforcement on inference endpoints was never affected; only the model listing was unfiltered.The Compliance page is now a policy dashboard. It pairs each control (Requesty zero data retention, model zero data retention, guardrails) with the traffic observed against it over your selected period, alongside guardrail violations and configuration changes. Require zero-data-retention models and the router rejects requests to models that do not meet the policy. Download the report as PDF or JSON for audit-ready evidence.
For
vertex/... Gemini models, auto caching now skips explicit cache creation when the cacheable prefix includes file attachments with URIs that are not on Google Cloud Storage (gs://). External file URLs (for example https://...) still work for inference; inline/base64 attachments (file_data) and gs:// URIs are supported in cached content.Auto Caching docs →Mistral AI joins the BYOK providers. Add your La Plateforme API key on the BYOK page and use it in routing policies to run Mistral models on your own billing relationship.BYOK docs →
The Model Library now shows discounted model pricing, including DeepSeek’s off-peak windows where token prices drop during set hours. The models endpoint returns a
discount_percentage field on discounted models so you can pick them up programmatically.Send video to models that accept it, such as Google Gemini, through the standard Chat Completions API. Use the
input_file content type with a public video URL (YouTube links work on Gemini) or base64-encoded data, and Requesty translates it to the provider’s native format. Look for input_video pricing in the Model Library to find compatible models.Video Input docs →Connect DeepSeek Harness (
dsh) to Requesty. The harness plugs into the gateway as an OpenAI-compatible provider, so you get 300+ models, cost tracking, and fallback routing for your dsh sessions with one API key. The Requesty CLI can write the configuration for you.Set up DeepSeek Harness →Admins have a new Compliance page that turns the data your gateway already records into an evidence report for auditors, security reviewers, and vendor due diligence. Choose a 7, 30, or 90 day period and see where your traffic ran, whether it used Requesty managed keys or your own BYOK credentials, how many requests were scanned by guardrails and what the scanners found, and how many configuration changes were made by how many people.Download it as a PDF for people, or as JSON for the record. The JSON carries a
payload_sha256 identity hash so an auditor can confirm the copy they hold is the one you generated. Reports are generated on demand and nothing is stored, and the report states its own scope and limitations in every copy: it is evidence, not a certification.The page also hosts Requesty’s pre-signed Data Processing Agreement, ready to countersign and return.Compliance Report docs →Every request now records a guardrail verdict alongside its usual telemetry: passed scanning, flagged, masked, not scanned, or scan error. This makes scan coverage measurable rather than inferred, and it is what the new compliance report counts. Requests recorded before this change carry no verdict and are reported as “not recorded”.The guardrails admin panel now lists the current v2 scanners (PII Fast, PII ML, Secrets Fast, Secrets ML). Legacy scanners you already have enabled stay visible and keep working.Guardrails docs →
Group endpoints now return only the budget fields that apply to your organization’s budget mode:
monthly_limit in Global mode, monthly_budget and monthly_budget_per_user in Group Budget mode. In Group Budget mode, monthly_limit is deprecated and will soon be migrated to monthly_budget. Budget endpoints that only exist in Group Budget mode (group budget, group per-user budget, member budget override) now return 409 Conflict in Global mode instead of silently writing a value that is never enforced, and member updates are scoped to the organization the API key belongs to.Log details now break each request’s cost into the model cost and guardrail charges. Every guardrail plugin that ran is listed with its individual charge and any discount applied, so the total cost of a request is fully accounted for.Guardrails docs →
Guardrail violations are now deduplicated per request. A value that is detected several times shows up as one violation with a
match_count in the logs detail panel and the violations API, instead of one row per occurrence.Guardrails docs →GET /v1/auth/check answers one question: is this API key usable? It returns an empty 200 for a valid key and 401 for anything else, so you can validate a key during onboarding instead of discovering a bad key on the first real request. Expired API keys are now also rejected consistently across all API-key-authenticated endpoints.The Requesty CLI is a terminal app that points the AI coding agents on your machine at Requesty. It detects Claude Code, Codex, OpenCode, Pi, Hermes, and DeepSeek Harness, rewrites each harness’s own configuration so requests go through the gateway, and shows your spend for the last 30 days right in the terminal. Existing config files are backed up before anything is written. Open source under Apache 2.0.Requesty CLI docs →
/v1/models covers chat models; new endpoints list the other modalities. Use /v1/models/embedding, /v1/models/image, /v1/models/transcription, and /v1/models/speech for a single modality, or /v1/models/all for everything at once.The models endpoint now returns
open_weights (whether the model’s weights are openly available), model_lab (the lab that created the model), and model_canonical_name (a canonical name that links the same model across providers and regions).Free model request limits are now 4-5x higher. New organizations get 200 free requests per day (up from 50) and paying organizations get 1,000 per day (up from 200), shared across all free models. Browse the current free lineup in the Model Library.Free Models docs →
Analytics now reports
reasoning_tokens alongside input and output tokens, and the user, group, and api_key filters accept comma-separated lists so you can compare several teams or keys in one view.Usage Analytics docs →The Users tab in the admin panel is now available on every plan, including pay-as-you-go. Admins can see all organization members and set per-user spending and rate limits, which members cannot change. Balance and top-up are now visible to admins only on all plans. Groups and group management remain Enterprise plan features.User Management docs →
The Overview page is the new home of the Requesty Console. It shows spend, requests, tokens, and average latency for your selected period, a usage chart you can group by model, API key, user, group, app, or token type, a totals table with one click drill down, and a yearly spend heatmap. Admins see the whole organization and can filter by user, group, or API key; members see only their own usage across their API keys. Every view is shareable as a URL.Overview docs →
Every API key now has its own detail page. Click a key on the API Keys page to see its spend, requests, tokens, and latency at a glance, a switchable chart over 7 or 30 days, a top-models ranking, and the key’s settings in one place: limits, expiry, labels, logging, and routing policies.
Organization usage can now be broken down by the labels you set on API keys and groups. Pass
group_by=api_key_label.<label_key> or group_by=group_label.<label_key> to the usage endpoints to slice spend by team, environment, cost centre, or any other label. Rows missing the label are grouped under Unknown, so totals still reconcile.The models endpoint now returns more metadata per model:
created and updated timestamps, retires (retirement date as a unix timestamp), quantization, model family, and multi-band pricing for models that charge differently by context size or time window.The API Keys page has been rebuilt. Compose your own filters with “Show keys that …” conditions (dormant, never used, no spending limit, spending over a threshold, and more), and share any filtered view as a link. Bulk-edit expiry, limits, labels, logging, and access lists across selected keys, and see when each key was last used. The admin panel now shares the same view, and expiry is editable by every key owner.
Auto caching now also triggers on multi-turn conversations, and cache breakpoints are anchored at the end of the conversation so long agent sessions keep reusing the cached prefix as they grow. No configuration changes needed.Auto Caching docs →
Requesty-managed routing policies are now available programmatically.
GET /v1/models/managed returns each managed policy with its model chain, so you can discover and reference them without opening the dashboard.Group labels are now available as analytics dimensions. Group and filter your analytics by
Group.<label_key> to see cost per department, cost centre, or any other label you set on your groups, alongside the existing ApiKey.<label_key> dimensions. Keys outside a group resolve to Unknown.Group labels can now be set programmatically. The request body replaces the entire labels map for the group; send an empty object to clear all labels.
Organizations can now set a default group so new members are enrolled automatically when they join. The default group’s access lists, approved models, and budgets apply from day one. Set it from the Gateway tab on the Groups Management page.Groups docs →
Groups can now carry labels, the same key-value pairs already available on API keys. Set them from the group’s page in the dashboard to tag teams by department, cost centre, or environment, then use them for filtering and analytics.Groups docs →
Microsoft’s MAI image models are now fully supported through Azure AI Foundry. Use
azure/mai-image-2.5 for image generation and image edits with the standard OpenAI-compatible request format, including size, with usage tracked and billed like any other image model.Image Generation docs →Router error messages have been rewritten to tell you what went wrong and what to do about it. Errors now distinguish gateway-side problems from provider-side ones, malformed request bodies return the exact JSON parse detail, and common provider errors like content filters, token limits, and capacity issues are translated into consistent, actionable messages.
A group’s total monthly budget and its per-user budget can now be set programmatically through two dedicated endpoints. Set a budget to
0 for unlimited spending. Requires an API key with admin permissions.Live provider rankings are now public at requesty.ai/rankings, powered by real routing traffic: latency, speed, and reliability per provider and region, plus coding-agent and SDK leaderboards.
Requesty now publishes managed routing policies: ready-made policies we maintain for the best experience out of the box. Find them in the new Managed tab on the Routing Policies page, where each policy shows its model chain and a copyable model ID. Use the policy name as-is as the
model in your request, no policy/ prefix needed. If you create a custom policy with the same name, your policy takes precedence.Managed Policies docs →The Mask action now replaces detected values with realistic placeholders before the request reaches the provider and restores the original values in the response, so masked prompts no longer break the model’s answer. The separate reversible mask option is folded into Mask.Guardrails docs →
Session analytics now has its own tab in the analytics navigation. See reconstructed sessions, drill into multi-turn conversations, and track session-level cost and usage without leaving the analytics area.Session Reconstruction docs →
Use Cursor with 300+ models through Requesty. A dedicated Cursor-compatible endpoint handles Cursor’s request format, so you get model routing, cost tracking, and fallback policies for all your Cursor traffic with a single API key.Set up Cursor →
The
reasoning_effort parameter now accepts xhigh, OpenAI’s highest reasoning tier. It is a strictly higher tier than high and is not the same as max, which normalizes to high on OpenAI models. Support is model-specific and passed through to OpenAI unchanged.Reasoning docs →The Routing Policies page has a new editor. Pick your strategy from Fallback, Load balance, or Latency cards, browse the full model catalog with provider, region, and data policy filters, drag models to reorder your chain, and select a BYOK key per model. Every policy gets its own URL and a Live Performance panel showing latency, success rate, and speed across the models in the policy over the last 24 hours or 7 days.Fallback Policies docs →
Latency-based routing now factors reliability into provider selection. Providers returning rate limits or server errors are penalized in the scoring, and routing optimizes for total request latency rather than time to first token alone. Faster and more reliable answers with no configuration changes.Latency Routing docs →
Prompts from the Prompt Library can now be executed by passing the prompt ID as the
model in your request. The prompt’s attached model, parameters, and response format are applied automatically, so one identifier carries your full configuration.Prompt Library docs →Prompts can now carry their own response format. In the Prompt Library, use Response Format to set free-form JSON output or a named JSON Schema with optional description and strict mode. Requests that reference the prompt with
prompt_id inherit the saved response_format automatically, and prompt-level response formats override caller-provided response formats. The editor’s JSON view and version sidebar now surface response format settings alongside model parameters.Prompt Library docs →Requests from the Pi coding agent are now auto-cached on Anthropic models. Repeated system prompts and tool definitions reuse the cached prefix on every turn with no configuration changes.Set up Pi →
/v1/images/edits now accepts source images and masks up to 20 MB per upload.Image Edits API →The Prompt Library now supports model parameters, version diffs, and a redesigned editor. Attach temperature, reasoning effort, max tokens, top_p, and n directly to a prompt so every request inherits them automatically. Compare any two versions with a line-by-line diff viewer that highlights added, removed, and modified messages. The new prompts list includes search, sortable columns, tags, and copyable prompt IDs. Reference any prompt in your API requests with a single
prompt_id and optional prompt_variables for runtime customization.Prompt Library docs →Bring your own Google Cloud credentials for Vertex AI. Paste a service account JSON on the BYOK page and the dashboard detects the key type automatically, so you can run Vertex models on your own Google Cloud project and billing.BYOK docs →
Organizations with a zero balance can now call free models. Paid models still require a positive balance, and a routing policy that lists a paid model before a free one falls through to the free route instead of failing.Free Models docs →
Hermes Agent by Nous Research now integrates with Requesty out of the box. Configure Hermes to route through the gateway using the native Anthropic Messages format, which enables automatic prompt caching on multi-turn conversations. Hermes sends a large system prompt with tool definitions on every turn, and Requesty’s auto-caching means subsequent turns reuse the cached prefix instead of reprocessing it from scratch. One config file gets you 300+ models, fallback routing, and cost tracking across all your Hermes sessions.Set up Hermes →
Send
X-Conversation-Id to group requests into one trace and X-User-Id to attribute them to an end user, from any client. Requests carrying X-Parent-Session-Id, as sent by OpenCode subagents, are merged into their parent session.Session Reconstruction docs →Access lists and approved models now apply to embedding, image, transcription, and speech models as well as chat models, so one list controls everything an API key or group can call.Access Lists docs →
When Anthropic returns a 502 before the first token of a streamed response, Requesty retries the request on the next route in your policy instead of ending the stream with an error.Fallback Policies docs →
Full programmatic control over access lists and group member settings is now available in the Management API. Create, read, update, and delete access lists, set the organization’s default group access list, update a member’s role within a group, and set per-member budget overrides, all via API key with manage permissions.
json_schema response formats on Anthropic models now use Anthropic’s native structured output support instead of a tool-call workaround, and tools and json_schema can be used in the same request.Structured Outputs docs →OpenCode traffic is now auto-cached on Anthropic models, and message-level
cache_control sent by OpenCode is honored instead of being dropped.Set up OpenCode →The organization usage endpoint accepts
group_by=member_id and group_by=member_email, so you can break spend down per person.Two new ways to control which models your teams can reach. Mark an access list as restricted so only org admins can attach it to groups or API keys, while group admins stay free to attach the rest. Set a group default access list that applies to any group without its own explicit list, giving every team a baseline without per-group setup. The resolution order now runs through four layers: API key list, then group list, then the group default, then your organization’s approved models.Access Lists docs →
Spending alert webhooks now include the resource ID alongside the name in every payload. User, group, and API key alerts carry
user_id, group_id, and api_key_id fields, and the Slack, Teams, and JSON formats all show the ID next to the resource name. No more guessing which “Production” key or “Engineering” group an alert refers to when names collide or change.Spending Alerts docs →Organizations that need a hard data residency guarantee can now enforce EU routing server-side. With strict EU enforcement enabled, every request from your organization must go through
https://router.eu.requesty.ai, and requests to non-EU endpoints are rejected. Contact [email protected] to activate it for your organization.EU Routing docs →Four models are now free to use through the gateway:
nvidia/nemotron-3-ultra-550b-a55b, nvidia/nemotron-3-super-120b-a12b, poolside/laguna-xs.2, and poolside/laguna-m.1. New organizations get 50 free requests per day and paying organizations get 200, shared across all free models. Free for now, any pricing change will be announced here first.Free Models docs →The Requesty n8n package (
@requesty/n8n-nodes-requesty v1.1.0) now includes a Requesty Image Generation node. Generate images from text prompts with models like azure/openai/gpt-image-1, with control over size, quality, background (including transparent), and output format. Get binary image data with a preview in the n8n output panel, or URLs for downstream use. The node also works as an AI Agent tool, so an agent can generate images on its own when a user asks.Set up n8n →Automate member offboarding alongside your existing API-based onboarding. Call
DELETE /v1/manage/org/member/{user_id} with a write-manage API key to remove a member from your organization: they lose access to the organization, all API keys they created in it are invalidated, and they are removed from all of its groups. Look up user IDs with the List Organization Members endpoint.Delete Organization Member reference →Models now carry a
retires_at date so you can plan migrations before a model goes away. The List Models response includes the field for any model with a scheduled retirement, and the date is shown on each model card in the Model Library. Once a model’s retires_at date passes, requests to it fail, so switch to a successor or add a fallback policy ahead of time.List Models reference →Enable real-time web search with a single tool definition that works across the Messages, Chat Completions, and Responses APIs. Pass
{ "type": "web_search" } and Requesty translates it to each provider’s native web search format automatically, normalizing citations and results behind one interface. Look for supports_web_search: true in the List Models response to find compatible models.Web Search docs →Get guaranteed, schema-valid JSON from supported models. Set
response_format to { "type": "json_schema" } with your schema and the model output conforms to it exactly. The List Models response now includes a supports_json_schema flag so you can find compatible models programmatically.Structured Outputs docs →/v1/audio/transcriptions supports response_format=verbose_json and timestamp_granularities[]. Responses include word and segment timestamps, detected language, and duration for OpenAI Whisper and Mistral Voxtral models.Transcription API →Latency-based routing now scores providers separately for short and long prompts, accounts for output speed on streaming requests, ignores outliers, and keeps requests that share a
trace_id on the same provider so prompt caches stay warm.Latency Routing docs →The
ttl: "1h" option in Anthropic cache_control blocks now works for Anthropic models served through AWS Bedrock and Google Vertex AI, not only Anthropic direct.Manual caching docs →Perplexity models are now available through the gateway with the standard Chat Completions format, including streaming and citations.
Append
:fast to a supported Anthropic model ID to request Anthropic’s fast service tier. The requesty metadata block is also accepted on /v1/responses, matching Chat Completions.Request Metadata docs →Build AI workflows in n8n with 300+ models through Requesty. Install the Requesty Chat Model community node (
@requesty/n8n-nodes-requesty) and drop it into any AI Agent or Basic LLM Chain. Get strict JSON Schema structured output, native web search, reasoning effort control, and routing policies, all behind one API key.Set up n8n →Logs and Sessions accept
start, end, and time_range so you can look at an exact window, and the log detail view shows the request parameters used for each call: temperature, reasoning effort, response format, and more.Logs & Traces docs →Use GitHub Copilot Chat in VS Code with 300+ models through Requesty. Add Requesty as a Bring Your Own Key (BYOK) Custom Endpoint provider to get model routing, cost tracking, and fallback policies for Copilot Chat, tools, and MCP servers. Requires VS Code 1.122+.Set up GitHub Copilot →
System messages can now appear anywhere in a conversation and are translated correctly for every provider. Anthropic’s
tool_search tool is supported, and /v1/responses accepts namespace tools so grouped MCP tools from clients like Codex work through the gateway.A comprehensive reference for every router error code: what it means, where it originates, and how to fix it. Plus, you can now group by
status_code in Advanced Analytics to track error trends and costs in your dashboards.Error Codes docs →Two new guardrail scanners detect personally identifiable information:
pii/v2/fast for low-latency pattern matching and pii/v2/ml for higher recall on names, phone numbers, and account identifiers. Enable them per policy with the Report or Mask action.Guardrails docs →Pick an exact start and end date for any analytics view. Use relative presets or switch to Date Range mode for full control over the time window.Usage Analytics docs →
Inspect every LLM request in a searchable table with click-to-filter, configurable columns, and pagination. Switch to Traces mode to see multi-step agent runs grouped by
trace_id. Click any row to open a detail panel with the full message timeline, tool call arguments, metadata, guardrail violations, and a model arena for side-by-side comparison.Logs & Traces docs →Download your advanced analytics data as a professional PDF report or CSV spreadsheet. The PDF includes your Requesty logo, date range, and a formatted data table, perfect for sharing cost reports with your team or keeping monthly records.Analytics Exports docs →
The organization usage endpoint accepts
group_by=group_id, group_by=group_name, and group_by=api_key_id so you can break spend down per team or per key.Use OpenAI Codex with 300+ models through Requesty. Get model routing, cost tracking, and fallback policies for your Codex coding agent.Set up Codex →
The Get API Key and List API Keys endpoints now include a
group field showing which group each key belongs to, making it easier to manage keys programmatically.API Keys reference →Route requests through the EU endpoint with the newest Gemini models:
vertex/gemini-3.5-flash@eu and vertex/gemini-3.1-flash-lite@eu. Full data residency when combined with the EU endpoint.EU routing docs →Lock down access with Entra ID (Azure AD), Okta, or any OIDC/SAML provider. Members authenticate through your identity provider and land directly in Requesty.Set up SSO →
Create named model allow-lists and attach them to individual API keys or groups. Control exactly which models each key can call without touching your org-wide settings.Create an access list →
Route OpenAI
/v1/responses calls through the gateway with full analytics, fallback, and cost tracking. Custom tool types are supported.Responses API reference →API responses now include a
usage.cost field with the exact dollar amount. For streaming, set stream_options.include_usage to get cost on the final chunk.Cost tracking docs →/v1/audio/transcriptions returns segments for Mistral Voxtral models and supports speaker diarization, so each segment carries the speaker who said it.Transcription API →A new Management API endpoint returns aggregated spend and token counts across your entire org, with the same time filters available on key-level usage.Org Usage API →
Chat Completions requests that use OpenAI’s
file content part are accepted and handled like input_file, so code written against OpenAI’s file inputs works through the gateway unchanged.PDF support docs →The
/v1/models endpoint now returns geolocation data for each model. The Model Library shows EU/US region chips so you can pick the right model before routing.EU routing docs →Connect the Pi coding agent for model routing, cost tracking, and fallback policies across your coding workflows.Set up Pi →
Context length overflows, unsupported image formats, and other provider errors are now translated into plain, actionable messages instead of generic errors.
Attach your own metadata to any request with
X-Requesty-<Name> headers, for example X-Requesty-Team or X-Requesty-Environment. The values are stored with the request for filtering and grouping in analytics and logs, and are stripped before the request reaches the provider.Analytics headers docs →The models endpoint now returns
data_retention, data_retention_days, data_used_for_training, and privacy_comments per model so you can pick models by how the provider handles your data. Analytics can also be grouped by user agent and request path.Requesty no longer adds Anthropic’s
context-1m beta header on your behalf, ahead of Anthropic retiring it. Beta headers you send yourself are forwarded unchanged.Organization members no longer need an admin to create a routing policy. Any member can create policies and attach them to the API keys they own.Fallback Policies docs →
Two new endpoints:
/v1/audio/speech for text-to-speech and /v1/audio/transcriptions for speech-to-text. Multiple providers including OpenAI and Mistral with automatic fallback.Speech API → · Transcription API →Send image edit requests through
/v1/images/edits with the same multi-provider routing and fallback as generation.Image Edits API →Pass
HTTP-Referer and X-Title headers to label requests by app or site. Filter your analytics dashboard by these values to see cost and latency per integration.Analytics headers docs →Set dollar thresholds on API keys and receive Slack or Microsoft Teams webhooks when spend crosses them. Configurable trigger percentages give you time to act.Configure alerts →
Reasoning requests to Claude Opus 4.7 are translated to Anthropic’s adaptive thinking mode, and web search results on the Responses API now carry the correct output index.Reasoning docs →
Use Requesty as the backend for the Claude Cowork desktop assistant. All traffic gets unified analytics, cost controls, and model policies.Set up Claude Cowork →
Route OpenCode terminal agent traffic through Requesty. A one-line installer adds analytics tracking to your setup.Set up OpenCode →
The admin panel now lets you set each guardrail policy to Disabled, Report, or Mask individually. A new violations column and detail tab in logs shows exactly what fired.Guardrails docs →
JSON Schema and json_object modes are now available on
/v1/responses, matching the Chat Completions feature set.Structured outputs docs →Send PDFs to Mistral models with the
input_file content type and Requesty converts them to Mistral’s native document format. Sending a PDF as an image_url now returns a clear error pointing you to input_file.PDF support docs →Azure deployments across European regions are now automatically recognized under the EU filter in the Model Library and routing engine.EU routing docs →
/v1/embeddings accepts multimodal input for Gemini Embedding 2.0: pass an array of content objects with text, image URLs, or data URLs to embed images and text in one call. Azure OpenAI embedding deployments are also available as a provider.Embeddings API →OpenAI models hosted on AWS Bedrock are now available as a provider, with streaming and tool calling through the standard Chat Completions format.
Provider grouping with expand/collapse, region and capability filters, bulk approve/remove, preset quick-filters, and a “New” tab that surfaces recently released models per provider.Manage approved models →
Pick two models, send the same prompt, and see which responds better. The redesigned chat playground also supports image attachments and markdown rendering.
DeepSeek models are now available through Google Vertex AI as an additional provider, using the same OpenAI-compatible request format as every other model.
Expandable table rows now show each service account’s API keys, monthly spend, and creator at a glance.Service accounts docs →
The gateway extracts and formats PDF content across providers that support document input. Just include the file in your chat completions request.PDF support docs →
If a provider connection drops in the middle of a streamed Chat Completions response, the stream now ends with an explicit error event instead of stopping silently, so your client can retry instead of waiting on a truncated answer.Streaming docs →
logprobs is supported on /v1/responses, including streaming, and Chat Completions requests for Responses-backed models return logprobs as well.Responses API reference →The API key usage endpoint now aggregates chat, embedding, transcription, speech, and image usage, and accepts its filters as query parameters on a
GET request instead of a JSON body.Dedicated model aliases like
coding/ select the right model, provider, and parameters for your workload automatically.Dedicated models docs →/v1/responses supports web search and returns URL citations as output annotations, both streaming and non-streaming, matching what you get from Chat Completions.Web Search docs →Advanced analytics supports a
minute time grouping for ranges up to 96 hours, plus hour-based time range presets, so you can inspect traffic spikes and incidents at fine resolution.Usage Analytics docs →Generate embeddings with Google’s latest Gemini Embedding 2.0 model through Requesty, with automatic provider selection.
Models served through the OpenAI Responses API now return their reasoning in Chat Completions responses.
reasoning.summary is supported, reasoning content is streamed alongside text and tool calls, and reasoning from earlier turns is forwarded back to the model.Reasoning docs →Every automatic top-up now emails the invoice PDF to your billing contact, and invoices are linked to their payment so they appear matched in your billing history.
The analytics dashboard now supports P95 and P99 latency percentiles, pivot tables for multi-dimensional breakdowns, and flexible time ranges including This Week, Month, Quarter, and Year.Usage analytics → · Performance monitoring →
When an Anthropic model performs a web search, the results and URL citations are translated into the Chat Completions response as annotations, for both streaming and non-streaming requests.Web Search docs →
The models list now shows which models support tool calling. Use this to filter models by capability before routing or to build smarter model selection.
Image generation requests can now be served by Azure OpenAI deployments.Image Generation docs →
Route OpenClaw autonomous agent workloads through Requesty for unified analytics and cost controls.Set up OpenClaw →
Structured outputs on Google Gemini models now pass your
json_schema to Google natively, so schema features that were previously simplified on Gemini behave as they do on other providers. Image generation also accepts models that return JPEG output.Structured Outputs docs →Reasoning content and signatures now survive translation between the Messages API and Chat Completions. Thinking blocks are carried through tool calling turns, so multi-step agent loops with extended thinking work whichever format your client speaks.Reasoning docs →
Organizations can now ask Requesty to enable Zero Data Retention (ZDR). ZDR turns logging off for every API key and service account, prevents anyone from turning it back on, and cannot be disabled once enabled, so request and response content is never stored.Data Privacy & ZDR docs →
Groups can now have their own approved model list, independent of the org-wide setting. Regional model approval handles providers with location-specific deployments correctly.Groups docs →
Set dollar thresholds on your organization and receive Slack notifications when spend crosses them.Alerts docs →
The API keys table is easier to scan with cleaner columns, hover tooltips for long values, and one-click copy. The same improved layout appears in both admin and user views.
Select any two requests in the logs table and open a JSON diff viewer. Added, modified, and removed fields are highlighted with one-click filtering.
Filter the traces page by trace ID, API key name, or user email. Cached percentage is now visible per trace.Session reconstruction docs →
Groups now show spend percentage and budget overrides directly in the table. Admins can adjust limits without navigating away.Groups docs →
Org admins can change member roles at both the organization and group level. Safety checks prevent admins from accidentally demoting themselves.Users and roles docs →
A new column shows how many reasoning tokens each request consumed, giving visibility into model “thinking” costs.
Scope cost, latency, and usage breakdowns to specific API key labels for more targeted reporting.Usage analytics →