DSPy 3.4 stops routing LM calls through LiteLLM by default

DSPy's 3.4.0b1 prerelease changes how the framework talks to language models. Instead of routing calls through LiteLLM by default, it now executes natively first, choosing engine="auto", and only falls back to LiteLLM when a route isn't supported or an input can't be represented faithfully in the native path. The request, response, and streaming types for this live under a new dspy.lm15 module, per the release notes.

The detail that matters most is what doesn't trigger a fallback. Auth failures, timeouts, and provider errors stay put. They error out instead of rerouting to a different backend behind the scenes. That's a real constraint to design around if you've built anything that assumed a failed call might retry through LiteLLM without you knowing.

Custom backends also get simpler. You can now implement a single complete(Request) -> Response method instead of subclassing BaseLM, and this replaces the experimental LM types that shipped in 3.3. If you built against those, expect to update.

DSPy names 3.5 as the deadline for this migration, which puts a real clock on it rather than leaving it open-ended. If you're running DSPy in anything production-shaped, this prerelease is worth pulling now to see what breaks, before the fallback path you may be depending on goes away for good.