Jev vs. sonnet with @pydantic AI. cost: sonnet $0.0026 vs. jev $0.001 time: sonnet 2.36s vs. jev 643ms Shown in Pydantic Logfire. code (just 17 lines):

I created an MCP server so my team can draft emails for me. Just one tool: draft_email(to, subject, body_markdown, cc, bcc) I check my drafts and click send. Running on @render free tier. instrumented with @pydantic Logfire, using GH/samuelcolvin/cloudkv for storage. Free to run. (not open source because it's mostly vibe coded)

Jev (and the hybrid derivatives that come soon and support generative types) are going to change the game for structured output. Pydantic AI supports Jev today. We'll carry on supporting the bleeding edge of structured output from LLMs.

For the last two weeks I've been putting Pydantic Logfire into an agent stack, and there's one architectural choice that keeps making my job easier: Every span you send to Logfire becomes a row in a Postgres table. Not a document in a proprietary column store. Not a segment in a custom time-series engine. A row. In a table called records. With JSONB attributes. Which means: you can point psql at it. Or DBeaver. Or Metabase. Or a Jupyter notebook via SQLAlchemy. Or dbt. Or your CI pipeline. For classic web apps, that's a nice-to-have. For LLM and agent workloads, it collapses the observability→BI pipeline from six hops (SDK → collector → vendor backend → nightly export → warehouse → BI) down to two (SDK → Postgres → BI). The question every ops team eventually asks — "which model is costing me the most and failing most often?" — becomes a single GROUP BY attributes->>'gen_ai.request.model' instead of a multi-widget dashboard. Full write-up (real query, honest trade-offs, when NOT to use it): Repo:

AI is likely already being used in your tutoring business. Our free guide shows you how to use it well and safely. From admin and teaching to the tricky conversations, plus AI prompt + policy templates. Get your copy here 👉

Terminal-Bench 2.1 pass@1 run is scored 76.40% with Luna High using Pydantic AI adapter of my Agent SDK. Next one will have same experiment with LangChain adapter and we'll see the difference.

There is currently a nasty bug in @github where diffs display wrong when you switch to the "Files changed" tab. It goes away if you reload the page, but it's confused me about 10times this weekend.

Significant `jiter` release, in particular this includes SIMD support for x86, which should give up to 7x performance improvement for parsing long strings. There are also significant improvements for tape reuse, arm64 string SIMD, float parsing and more. jiter is the most used 3rd party JSON library in Python as a dependency of @pydantic validation. it's also used in Pydantic Monty, and in Pydantic Logfire via datafusion-functions-json. Short story: apply strict constraints, provide clear benchmarks for the thing you want to improve, use @codspeedhq, let fable and astra do their thing.

I go to New York to sell @pydantic Logfire. Easy - it's the best. Pydantic team wonder how long it'll take for enterprise to even think about Pydantic Monty (it's Samuel's crazy side project, brand new, is it even useful?), so I ask: ALL the financial institutions we speak to have teams (whole fucking TEAMS!) building Monty as a service for internal use. Multiple massive financial institutions are spending millions to enable them to adopt one sandboxing technology. We thought we were too early to build a product around Monty, we're late. Full Monty commercial service, is ready for production now. LMK if you want a snapshotable, durable, forkable sandbox with <1ms launch times.

Pydantic AI + Monty + a local 27B on my own machine, doing real arithmetic over my real transactions. The model doesn't compute the number, it writes the code that computes it. All through one ledger() function I wrote and control. @samuelcolvin @pydantic Y'all doing the damn thing 💪🏾 Literally just learned about ClassInstance, trying that next.