The car wash test is the simplest AI reasoning benchmark that nearly every model fails. We tested 53 models through Opper, first once each, then 10 times. Only 5 passed consistently.
Opper Skills bring the full power of the Opper platform directly into your AI coding agent. Install once, and your agent knows how to use Opper.
A technical deep-dive into how Opper AI context features enabled a small open-weight model to match the quality of a large frontier model, at a fraction of the cost and with faster response times.
Turn your agent behavior into high-quality training data automatically. Learn how Opper's new feedback system creates a self-improving loop for your LLM applications.
Introducing Model Interaction Protocol (MIP) - a task specification that decouples what we want models to do from how they should do it, enabling scalable, reliable AI agents in production without constant maintenance.
Introducing Opper Agent SDKs - powerful Python and TypeScript frameworks for building intelligent, reliable, headless agents with seamless MCP integration, multi model support, comprehensive hooks, and built-in observability.
Opper makes LLM calls more reliable with built-in fallbacks, aliases, and observability—so your apps stay stable without extra code.
Opper AI has acquired FinetuneDB and will integrate its fine-tuning workflow into Opper’s reliability stack to accelerate reliable, task-specific agents.
Opper AI partners with Berget AI to bring EU-hosted serverless inference to one API
We present results of our Benchmark for Context tasks, with insights into what is easy and hard for each of the models
We built TaskBench to measure real-world LLM performance on practical tasks like RAG, SQL generation, and agentic workflows. Here are our findings across accuracy, cost, and model size.
Three reference‑free evaluators to demonstrate how to evaluate RAG systems at runtime without gold references.
In this post, we will build an initial version of a simple but effective GitHub PR review agent using the ReAct pattern.
This guide explores how to automatically index your documentation in Opper Indexes using GitHub Actions for integration with your CI/CD pipelines
Schema based prompting is a technique that allows for instructing models through clear data structures instead of natural language prompts. This post introduces the concept and shows how to use it with Opper.
Opper now provides an OpenAI-compatible API endpoint that works seamlessly with OpenAI SDKs and its ecosystem of popular AI tools and frameworks.
Opperator is a programmable, autonomous web agent for automating tasks on the web built with the Opper SDK.
Using a reasoning model to generate detailed thought traces that improve the quality of AI responses, while keeping costs low.
We build a chatbot to assist users by utilising tools, while maintaining context and handling errors gracefully.
In this blog post we explore how OpenAIs reasoning models o1-mini and o1-preview perform in a RAG pipeline with structured output
Three days at the AI Engineering World Fair in San Francisco, covering the what, how and why of LLMs and how to best use them.
We build a pipeline to shape the output of LLM calls with synthetic examples and few shot retrieval, and see how multiple non-frontier models perform
Exploring how to set up APIM with an OpenAI-compatible API, and how to connect it to multiple OpenAI deployments. We will also cover how to set up failover and load-balancing.
In this blog post we explore how to implement and use RAG metrics to evaluate the quality of answers generated by a model. We will use the answer correctness metric to evaluate the quality of answers generated by a model.
Delvin is an agent fixing issues from the SWE-Bench Lite dataset, achieving state-of-the-art accuracy(23%) with very simple code leveraging the Opper SDK.
In this blog post we explore how to use multimodal models in Opper. We will use the newly released model gpt-4o to generate structured data from images
In this blog post we show how to build a simple RAG feature with citations, using Opper with structured input/output and Mistral-Large.
Getting GPT 3.5 Turbo to 80% accuracy on the GSM8k benchmark by leveraging the teacher student pattern, synthetic examples and few shot retrieval.
Dramatically improving LLMs accuracy with structured generation and chain of thought
At Opper, our mission is to accelerate adoption of Generative AI by making it simpler to build production grade reasoning applications, agents and features.