Opik changelog: what's new each month
Every stable Opik release summarised by month: the highlights, new features, improvements, fixes and anything you need to act on. 3 months covered; the current month updates daily.
November 2026 · so far
6 releases: 2.2.138 → 2.2.143
November brought improvements to evaluation workflows, thread searching, and span management, along with fixes for scoring, judge behavior, and database handling.
Highlights
- Search within threads directly in the thread side panel for easier navigation.
- Open any LLM span in the playground to inspect and test prompts interactively.
- Query experiment items across all dataset experiments for broader analysis.
- Evaluate more than 1000 threads without hitting previous limits.
- Service account token authentication now supports plain-text token replies.
New
- Search within a thread in the thread side panel.
- Open any LLM span in the playground.
- Query experiment items across all dataset experiments.
- Filter experiments by evaluation method.
- Service account token authentication with plain-text token replies.
- Organization context with a default organization.
- Always persist online-evaluation monitoring traces.
- Score online-eval Decision rules with gpt-6-luna through OpenAI's Decisions API.
Improved
- Agentic judge now submits its verdict through a tool.
- Evaluate_threads can now handle more than 1000 threads.
- LLM judges on gpt-6 and Claude 4.7+/5.x no longer send refused sampling parameters.
- Project names are now limited to 150 characters.
- Thread page messages are read from first and last traces only for better performance.
- User-facing log queue is now bounded to prevent heap overflow.
- Python SDK dependencies are capped at releases at least 3 days old.
- Prompt() and ChatPrompt() now use the default project.
Fixed
- Run simulation traces are now logged to the given project_name.
- User fields named start_time / end_time / last_updated_at remain unchanged on replay.
- Span cost breakdowns are now supported.
- Task-span metrics that re-scoring cannot score are now reported.
- A dataset run with several prompts into a new project no longer fails with "Project already exists".
- Non-object LLMJudge verdicts are now treated as parse failures.
October 2026
50 releases: 2.2.87 → 2.2.137
October brought major improvements to the playground, model support across more providers, and performance optimizations throughout the platform. The month focused on enhancing LLM integrations, refining the user experience, and strengthening backend reliability.
Highlights
- Playground now runs experiments on the server with support for datasets and test suites, and includes new prompt actions like duplicate, load, and save.
- Added support for OpenAI Decisions API, Anthropic thinking effort, and parameter panels for Ollama and Bedrock providers.
- Implemented a model catalog with per-provider parameters and automatic parameter validation across OpenAI, Anthropic, Gemini, and other providers.
- Added environment filtering to threads, made dashboard editing discoverable, and introduced Ollie widget mode on dashboards.
- Performance improvements to trace and span reads using indexed week-based filtering across multiple queries.
New
- Native Ollama SDK integration.
- OpenAI Decisions API tracing support.
- Ollie widget mode on dashboards and Ollie chart widget query building by reference.
- AWS EventBridge alert publishing for triggered alerts.
- Online evaluation rules editor sidebar with new card interface.
- Environment filter for threads tab.
- Column hiding in playground dataset and test suite tables.
- Prompt template version history viewing without restoration.
Improved
- Playground now runs and stops experiments on the server for regular datasets.
- Model catalog with profiles provides per-model parameters, replacing hardcoded settings.
- Responses API now supports OpenAI's max reasoning effort and extra body fields.
- Anthropic thinking effort is forwarded as output_config.effort in completions.
- Compare prompts render as messages and store prompt templates for better organization.
- Logs filters persist when returning to the page.
- Prompt library saves only occur once per click to prevent duplicates.
- Thread messages wrap properly instead of cutting off at the left edge.
Fixed
- Playground crashes show a recovery screen with Reset and Copy details options.
- Playground runs with Cmd/Ctrl+Enter only execute the prompt being edited.
- Rate limiter now fails open gracefully when Redis script errors occur.
- Project name conflicts show a clear error message.
- Playground traces are excluded from home page trace count.
- Thread search now picks threads and keeps all their traces together.
September 2026
30 releases: 2.2.57 → 2.2.86
September brought significant performance improvements across data ingestion and processing, along with new features for experiment management and annotation queues. The month focused on optimizing bulk operations, expanding LLM integration support, and enhancing the user interface.
Highlights
- Performance improvements to dataset and experiment bulk uploads through streaming and compression instead of materializing data.
- New ability to name experiments directly from the Playground and export full experiment result sets from the browser.
- Annotation queue automation data model and services introduced for queue management and routing.
- Support for additional LLM providers and models, including Gemini 3.8 Flash thinking levels and Jev via OpenRouter.
- Security fix for SSRF and response-body disclosure vulnerability in alert webhook endpoints.
New
- TypeSafe AI integration added to SDK.
- Annotation queue automation data model and configuration interface for queue-based workflows.
- Dataset item field mapping to trace and span fields for richer dataset enrichment.
- Jev classifier judge support in online evaluation rules via OpenRouter.
- MCP announcement banner and expanded MCP connection metadata capture for OAuth registration.
- Data anonymization documentation page added.
- Extended free-form SQL ClickHouse user provisioned behind feature flag.
- Support for Gemini 3.8 Flash thinking levels in Playground and online scoring.
Improved
- Experiment bulk uploads now stream instead of making four full passes, reducing memory usage.
- Dataset item uploads stream instead of materializing entire datasets in memory.
- Playground output grid virtualized for better performance with large result sets.
- Request body encoding uses orjson where available for faster serialization.
- Feedback score bulk inserts streamed via ClickHouse v2 client.
- Experiment item inserts and dataset item version inserts now streamed via ClickHouse v2 client.
- Trace and thread stats queries bounded to relevant time periods for faster execution.
- Claude sampling parameters now governed by model rather than provider for more accurate configuration.
Fixed
- Python metric scores with no value no longer fail the batch; they are dropped instead.
- Duplicate Cursor traces across edits prevented.
- Default ascore now runs off the event loop to prevent blocking.
- LiteLLM proxy cost correctly forwarded to ADK spans.
- OpenTelemetry requests with nothing to store no longer return 500 errors.
- Failed evaluator responses rejected instead of stored as zero scores; wholly unusable responses now fail gracefully.
Summaries are written automatically from the official release notes (full changelog ↗); check the original notes before relying on a detail. Opik: pricing, features and alternatives · All changelogs