cailos

Enhanced dashboards with Time to First Token metrics and improved routing resilience

Cailos improves system observability and reliability by introducing Time to First Token and Tokens Per Second metrics to team analytics and health dashboards, with sorting and skipped check reporting. Endpoint reliability rankings now penalise failures by 30% and use TTT rather than total latency for more accurate assessment. Additionally, new request routing race executor tests endpoints concurrently and seamlessly hands over tasks if primary stalls, improving request resilience.

Enhanced model selection during endpoint deletion with infrastructure scaling

Cailos improves endpoint deletion experience by adding searchable field for selecting replacement models, with automatic prioritisation of identical models and visual badges for semantically identical options. A new option allows redirecting to nowhere when no replacement is selected. Additionally, web service memory limit has been increased to 3 GB to support growing workloads.

Fix eval status tracking and bulk run behaviour

This release improves how evaluation statuses are tracked and displayed across system. Eval statuses now accurately reflect queued or running states rather than defaulting to fixed timeouts. bulk run evals button now queues specific due endpoints individually instead of marking all active endpoints as queued at once. Health checks and new endpoints properly show queued status rather than immediately stacking duplicate evals.

Enhanced accessibility, responsive design, and user experience on landing page

Cailos improves landing page accessibility with focusable code panels, stable navigation, and screen reader labels. Support for reduced motion preferences has been added, and responsive design now accommodates screens down to 320 pixels. Additionally, idle tour with pause controls was introduced, hero text flicker in Chromium was resolved, and "Send your own prompt" console was removed for cleaner interface.

Enhanced landing page interactivity and improved API documentation

Cailos adds interactive 3D scene and story chapters to landing page to better showcase product capabilities. API documentation has been updated to include new parameters, error codes, and dedicated privacy section. registration page layout has also been optimised for better responsiveness across devices.

Enable prompt caching for Anthropic API with accurate token usage reporting

Cailos now supports automatic prompt caching for Anthropic provider through cache control headers on requests. Token usage reporting has been updated to include cached reads and writes in prompt token counts, giving users accurate visibility into API costs and efficiency gains.

Improve JSON output support and make Claude refusals easier to interpret

Claude-compatible requests now support both structured JSON-schema output and schema-free JSON objects, while structured output compatibility has been restored. Claude refusals are also reported with `content_filter` finish reason, giving clients consistent way to identify filtered responses.

Retire models safely, preserve replacement routing, and make request failures easier to understand

Cailos now retires models automatically when their final endpoint is deleted, redirecting requests to compatible replacements and migrating affected custom model references and routing plans while preserving strategy suffixes. Administrators can compare model similarity and preview retirement outcomes from endpoint and administration pages. Teams and users also gain grouped request-failure panels in live dashboard, while weekly digest emails include privacy-conscious failure summaries without exposing sensitive provider error details.

Enable simultaneous reasoning and tool calls on GPT-5.x models with improved error handling

This release adds support for simultaneous reasoning and tool calls on OpenAI provider, resolving 400/502 errors on GPT-5.x models. Tool loops with reasoning context are now enabled through stateless caching and content replay. Additionally, failed adapter responses are raised as exceptions to improve how routing system treats provider failures.

Reduce unnecessary Jev fallbacks on slower network connections

Jev-backed requests now allow up to three seconds for response, reducing unnecessary fallbacks when network connections are slower and making request handling more resilient.

Unified chat and confirmation dialogs, new team oversight features, and improved input handling

This release brings consistent styling across confirmation dialogs, preserves chat reply state, and fixes form re-rendering on validation errors. A new team overview page displays spend and per-model request metrics with automatic outlier identification. User timezone preferences now apply across feeds, events, charts, and digests. Request logs without associated teams are now restricted to staff-only visibility, and trust-level descriptions are clarified for data training and retention. registry page now handles deleted providers gracefully without malformed ID errors.

Define routing logic with natural language and organise models more intuitively

Cailos now supports natural-language routing instructions in custom models, with preview feature that validates routing plans against model pools before saving. Models listing page has been reorganised into Starred and Other categories with improved filtering and sorting. Request trace views now display active routing instructions to aid debugging and analysis.

Use flexible language inputs with reliable model selection, and see how each request will be routed.

Cailos now accepts recognised language aliases and alternate forms, normalising them to canonical language codes while rejecting unsupported values. Language preferences are applied consistently during automatic model selection without unnecessarily excluding otherwise eligible routes. Routing previews now show selected execution strategy and candidate endpoints, making routing decisions easier to understand.

Attach images in web chat and discover vision capabilities available to your team.

Web chat now supports up to five images per prompt, preparing uploads efficiently by resizing and re-encoding them to reduce transfer size and remove metadata. When images are included, Cailos preferentially routes requests to vision-capable models. API now exposes model vision capabilities in model listings and through authenticated per-model capability endpoint, while ensuring each team sees only models it can use.

Explore how Cailos balances quality, speed and cost while maintaining resilient multi-provider routing.

Cailos landing page has been rebuilt as interactive product experience. Visitors can explore live endpoint data, compare quality and price preferences, and see how routing handles fallback, provider outages and throughput-based racing. refreshed responsive interface also improves accessibility, navigation, code sharing and support for reduced motion and colour themes.

Discover, manage and select models more easily, with consistent dashboard filtering and clearer routing analytics.

This release introduces searchable Models catalogue with coverage, pricing, capabilities and performance details, alongside personal favourites that follow users across teams and model pickers. Superusers can now maintain key model metadata, while provider pages and custom-model selection offer better search, coverage and failover guidance. Team management and dashboard analytics have also been refined through durable tabbed navigation, shared time filtering, more accurate cost reporting and clearer routing . Evaluations now use each model’s normal reasoning and sampling defaults.

Chat completions gain race, judge and wizard routing modes, with resilient streaming and richer routing telemetry.

This release introduces multi-model routing for chat completions, including race, judge and wizard execution modes. Race-mode streaming can now recover from stalled or failed contenders through speculative hand-offs, while new telemetry exposes selected mode, contenders, winners, timings and baton passes. Eligible teams can access routing modifiers according to their plan , and playground provides dedicated views for comparing contenders and following judge and wizard progress. and public API have also been refreshed with consistent responsive experience, improved navigation and interface controls, and clearer .

New UI for pricing, tightened governance thresholds, and corrected provider URLs.

This release introduces a new user interface for endpoint pricing, providing real-time warnings and cost cap. It also includes a reduction in endpoint freshness/governance threshold to 90 days and essential fixes for provider base URL handling to ensure correct request paths for Groq and Cerebras.

Router evaluations now support path-based validation constraints and feature improved data readability.

The evaluation framework has been strengthened with the introduction of path-based validation constraints and revised scoring for validator types, enabling more precise quality checks. The experience is further refined with an expanded question detail page that distinguishes between router-test and endpoint-ranking questions, alongside UI improvements for model naming and numeric alignment to enhance data readability across evaluation pages.

Router evaluation results and strategy scores now surface token counts, per-call cost breakdowns, and cost-efficiency summaries.

Router evaluation has been enriched with two layers of new financial and performance data. Evaluation results now include token counts and a detailed cost breakdown for each individual call, giving users precise visibility into the expense of each routing decision. Strategy scores have also been extended with cost-efficiency metrics and performance summaries — including median latency and throughput rate — so users can compare strategies not just on quality but on their overall value and speed profile. Together, these additions make router evaluation a more complete tool for cost-conscious optimisation of routing strategies.