Topics has a dedicated page that shows all clusters generated from your production logs.
Compare past and present groupings to understand how user behavior evolves over time and gain visibility into how Topics categorizes conversations, complaints, and feature requests.
I’ve seen a lot of confusion about how to best leverage the new family of GPT 5.6 models, and honestly, sameeee! Running this eval helped me pin down which models I should use (or have an agent use) to build effective code without burning extra tokens and dollars.
We evaled the GPT-5.6 family, plus Anthropic's Fable, Opus 4.8, and Sonnet 5, on the key building blocks of agentic workflows.
Then we broke results down by task type and difficulty and turned them into a decision map you can route against.
Here's what we found.
We evaled the GPT-5.6 family, plus Anthropic's Fable, Opus 4.8, and Sonnet 5, on the key building blocks of agentic workflows.
Then we broke results down by task type and difficulty and turned them into a decision map you can route against.
Here's what we found.
The family-level results show data transforms are nearly solved for the OpenAI models, symbolic rules pull them apart, and the Anthropic rows are dragged down by refusals rather than by wrong answers.
If you're building a voice agent, picking the right speech-to-text model is not obvious. Every provider claims to be accurate, fast, and production-ready, but the benchmarks they publish rarely look like actual traffic.
We used Braintrust to build a controlled eval across six
Your agent is in production. Now you need to understand what it is doing, where it is failing, and what to improve next.
Join our live workshop to build a repeatable workflow for turning agent behavior into evals, measuring improvements, and catching regressions.
Then hear
The Braintrust Go SDK enables automatic instrumentation with no code changes using Orchestrion, or the use of manual middleware for explicit control.
Supports OpenAI, Anthropic, Google Gemini, Google ADK, and more.
Phrase search breaks when every word is common but the exact sequence is rare. At 100TB+ of agent traces, queries time out and traditional databases fail.
Brainstore uses shingled bloom filters, indexing trigrams instead of tokens.
On a 290 GB dataset, 98.5% of segments were
Switching to a cheaper model can cut token costs, but if failures increase and retries pile up, the savings disappear.
Learn how to run eval experiments to compare models objectively so you can reduce costs without compromising agent quality.
In January, I joined Braintrust, told my manager @morgane_paloma that I was huge into pickleball, and 6 months later she threw a pickleball event just for me ❤️
Okay maybe not just for me, but the rest is true!
Had an amazing time last week in SF at The Agent Open hosted by
