Skip to main content
Track changes to the Auriko API, SDKs, and supported models.

​
1.0.1 — 2026-06-27

​
API

  • API key identity restructure: GET /v1/me returns full credential introspection — credential_type, status, profile, key_id, key_prefix, key_name, nested workspace, scopes, structured rate_limits, expires_at, created_at. The object, user_id, and tier fields are removed
  • API key list default: GET /v1/workspaces/{workspace_id}/api-keys excludes revoked keys by default; pass include_revoked=true to include them
  • Credit balance fields removed: current_tier, platform_fee_rate, tier_volume_usd, next_tier_threshold_usd from the credit balance API response
  • /api/v1/pricing/tiers endpoint removed
  • Management API: Programmatic workspace management via API key — BYOK key lifecycle (byok:read/byok:write), BYOK discovery, workspace settings (workspace:write), routing defaults (routing:read/routing:write), billing and subscription status (billing:read), member listing (members:read), invite lifecycle (members:read/invites:write), audit events (audit:read), budget management (budgets:write), API key lifecycle (keys:write). All write operations return 202
  • Analytics dashboard API: Summary, performance trend, routing evidence, paginated request log, and CSV export with truncation headers (X-Export-Row-Count, X-Export-Row-Limit, X-Export-Truncated)
  • /v1/models contract: endpoint query parameter and supported_endpoints/context_window response fields
  • Model directory aliases: Each entry in /v1/directory/models includes an aliases list
  • Budget headers on /v1/me: X-Budget-* headers when budgets are configured
  • Response API phase field: output[].phase ("commentary" | "final_answer") typed in both SDKs and the API reference schema
  • Usage dashboard access: All workspace members can view usage analytics

​
Error Handling

  • response_api_only error code: Response-API-only models return this dedicated code with an actionable suggestion field
  • client_disconnected error code: Error code for client-initiated disconnections
  • idempotency_replay_unavailable: Idempotent endpoints return 409 when replay data is unavailable
  • Model-not-found suggestions: model_not_found errors include an optional suggestion field with similar model names
  • X-Rate-Limit-Source header: Rate limit errors distinguish between provider and Auriko limits (provider or routing)
  • Double /v1/ path detection: Requests to /v1/v1/* return 400 naming the duplicate prefix
  • Retry-After minimum: Rate-limited responses return Retry-After: 1 minimum

​
Parameters & Request Handling

  • Response API parameters: frequency_penalty, presence_penalty, max_tool_calls on responses.create()
  • Response API echo-back fields: temperature, top_p, frequency_penalty, presence_penalty, top_logprobs, truncation reflect request values in responses
  • max_tool_calls clarification: Limits built-in tool calls only (e.g., web_search, code_interpreter), not user-defined function tools
  • reasoning.effort on Claude: Enables thinking and produces reasoning output through the Response API. Values exceeding a model’s range are clamped with an unsupported_parameter warning
  • Forced tool_choice with reasoning on Anthropic: tool_choice (any/tool) coexists with reasoning_effort; incompatible parameters are dropped with a warning

​
Routing & Providers

  • Per-model rate limit isolation: Rate limits track each model independently; custom per-key rate_limit_rpm values are enforced
  • Provider error messages: Include the provider name in error.message with status-appropriate messages. Provider 403s are retryable
  • Timeout reporting: Stream first-byte timeouts report reason: "timeout" with provider_timeout warnings
  • tool_choice="required" on GLM models: Targets capable providers; models with no capable provider return tool_choice_required_not_supported

​
Model Compatibility

  • Claude Opus 4.8: Tools, structured output, vision, web search, code execution, prompt caching (1,024-token minimum). Adaptive thinking only
  • Claude Fable 5: Adaptive thinking, 1M context, 128K output. Tools, structured output, vision, web search, code execution, computer use, prompt caching (512-token minimum)
  • Gemini 3 Pro Image and Gemini 3.1 Flash Image: Image generation with text-and-image I/O, tool calling, web search, reasoning
  • MiniMax M3: Tiered pricing, tool calling, reasoning, vision, prompt caching. Additional provider via SiliconFlow
  • Nemotron 3 Ultra 550B: Reasoning, tool calling, JSON mode via Together AI. Additional provider via DeepInfra
  • Nex N2 Pro: 262K context, adaptive thinking, vision via SiliconFlow (free tier)
  • Undated Claude model aliases resolve to dated counterparts
  • 12 responses-only models (o1-pro, o3-pro, gpt-5/5.2/5.4/5.5-pro, codex family, deep-research) discoverable in /v1/directory/models
  • Qwen 3.7 Max via DeepInfra; MiniMax M2.7 via DeepInfra
  • Structured output works with array-typed schemas (e.g. ["string", "null"]) on Gemini
  • SiliconFlow Qwen 3 VL 32B/8B does not support response_format: json_schema
  • reasoning_effort with low max_tokens on Gemini 3.x returns finish_reason: "length"
  • Response API streaming usage includes token counts for all providers

​
SDKs

​
Documentation

  • Response API guides: overview, streaming, tool calling, structured output, reasoning, and routing

​
0.3.0 — 2026-05-24

​
API

  • Response API: POST /v1/responses creates model responses using the OpenAI Response API format, with streaming, tool calling, reasoning summaries, and multi-model routing
  • Gemini image generation: Image-capable Gemini models return generated images inline as images[] in chat completions, with image_tokens in usage
  • Gemini 3.5 Flash: Routing to gemini-3.5-flash on Google AI Studio with function calling, structured output, streaming, vision, reasoning, prompt caching, and code execution

​
Routing & Providers

  • Streaming-only provider routing: Providers that only support streaming are excluded from non-streaming requests. Requests targeting a specific provider return non_streaming_not_supported with guidance to set stream: true
  • kimi-k2.5, kimi-k2.6, gpt-oss, and glm-5.1 support the Response API
  • tool_choice: "required" downgrades to "auto" on reasoning models with an unsupported_parameter warning. Models with no capable provider return tool_choice_required_not_supported
  • tool_choice: "none" is honored on all endpoints

​
Parameters & Request Handling

  • Default timeouts: 120s streaming first-byte, 5-minute non-streaming total, 18-minute non-streaming deadline. User-set timeout_ms / deadline_ms override defaults
  • Unsupported parameter warnings: Human-readable format (e.g., 'metadata' is not supported.)
  • system as top-level parameter: Returns a structured warning directing to the messages array

​
Model Compatibility

  • claude-opus-4-7 sampling parameters: temperature, top_p, and top_k are accepted with an unsupported_parameter warning on all endpoints
  • Anthropic reasoning with low max_tokens: Requests with reasoning_effort and low max_tokens return successful responses with an unsupported_parameter warning
  • Response API returns structured reasoning output for DeepSeek R1 0528, Qwen 3 30B A3B, and Qwen 3 32B

​
SDKs

​
0.2.0 — 2026-05-18

​
Routing & Providers

  • Priority tier routing: gateway.routing.tier: "priority" enables faster inference (2.5x speed, 6x cost). Premium-tier offerings not included by default
  • Requests with tool_choice="required" target providers that honor the constraint. Returns tool_choice_required_not_supported if no capable provider is available
  • SiliconFlow-hosted models are supported, including 13 exclusive models and reasoning variants (DeepSeek-R1, Qwen3 thinking)
  • qwen-3.6-plus available through additional providers
  • 9 Qwen 3 models are available through alternative providers

​
Parameters & Request Handling

  • Default timeouts support reasoning models with longer first-token times. deadline_ms is opt-in only

​
Model Compatibility

  • Anthropic response_format supports type arrays (e.g., type: ["string", "null"]) with nullable enum and object handling
  • Zero-parameter tools validate on Anthropic and Google models
  • reasoning_effort coexists with temperature, top_p, and top_k on Claude models. Incompatible sampling parameters are omitted with a warning
  • reasoning_effort mapping across providers:
    • Claude: "xhigh" preserved on Opus 4.7, clamped to "high" on Opus 4.5, mapped to "max" on other models
    • Gemini 3.x: "xhigh" and "max" map to "high"
    • OpenAI and xAI: "max" maps to each model’s ceiling
    • All mappings include an unsupported_parameter warning
  • reasoning_effort with extreme max_tokens on Gemini clamps to each model’s accepted range
  • DeepInfra Gemma 3, Qwen3 VL, and MiMo return tool call responses, streaming and non-streaming
  • MiniMax models return structured reasoning output
  • output_tokens in Anthropic-format responses includes reasoning tokens
  • llama-3.1-8b-instruct-turbo returns a capability error for unsupported structured output

​
Error Handling

  • Streaming errors return structured messages with message, type, code, and provider fields
  • Multi-provider fallback errors include per-provider failure details

​
SDKs

​
Integrations

​
0.1.1 — 2026-05-12

​
Parameters & Request Handling

  • User-configurable timeouts via gateway.routing.timeout_ms (per-attempt) and gateway.routing.deadline_ms (request deadline)
    • Streaming uses timeout_ms as time-to-first-byte
    • Defaults: 10s streaming first-byte / 120s streaming deadline; non-streaming 120s / 180s
  • reasoning_effort accepts "xhigh" and "max" levels. Anthropic models use adaptive thinking with automatic effort-to-budget translation
  • Timeout errors include actionable guidance naming the configurable parameters
    • provider_timeout warning emitted when fallback succeeds after a timeout
  • timeout_ms and deadline_ms apply to all execution paths including allow_fallbacks: false

​
Error Handling

  • Error codes unknown_parameter, capability_mismatch, routing_constraint_unsatisfiable, tokens_per_min_exceeded, concurrent_requests_exceeded, and routing_failed — capability-specific codes listed below
  • Inference validation returns specific error codes — missing_required_parameter, invalid_parameter_value, unknown_field, operation_not_allowed — for precise programmatic error handling
    • invalid_request is reserved for malformed request bodies. HTTP status codes unchanged
  • Platform operations return state_precondition_failed, duplicate_resource, and operation_not_allowed. 405 responses return method_not_allowed
  • New platform codes: field_immutable, invalid_recovery_code, mfa_verification_failed

​
Model Compatibility

  • Gemini tool calls return finish_reason: tool_calls for client-side tool execution
  • Gemini accepts non-object tool results (numbers, booleans, arrays)
  • reasoning_effort works on Gemma-4, qwen-3-vl-thinking, and gpt-5.4+ (when combined with tools)
  • Explicit thinking.budget_tokens requests (e.g. from Claude Code) work on models that support them

​
SDKs

  • Python SDK auriko 0.1.1
  • TypeScript SDK @auriko/sdk 0.1.1
  • Response models include service_tier, annotations, and prediction token fields

​
Integrations

​
0.1.0 — 2026-05-04

​
API

  • OpenAI-compatible chat completions with multi-provider routing and automatic failover
  • Anthropic Messages API compatibility — POST /v1/messages and token counting
  • Streaming and non-streaming responses across all supported providers
  • Model directory with pricing, capability metadata, and provider health status
  • Workspace billing and credit balance (Preview)
  • Structured reasoning support with cryptographic signatures for multi-turn round-trip
  • Structured error responses with machine-readable code and type fields for programmatic error handling

​
SDKs

​
Frameworks

​
Models

  • GPT-4o, Claude 3.5 Sonnet, Gemini 1.5 Pro, Llama 3.1, Mistral Large, DeepSeek V4, and more — see Models for the complete list