LiteLLM 1.100.0 deletes prompt_token_calculator and adds Bing grounding as a search provider

BerriAI published LiteLLM v1.100.0 on 6 September 2026. The release carries two entries marked as breaking changes in the changelog, and the version list runs to 412 linked pull requests since v1.99.0, which shipped five days earlier on 1 September. LiteLLM is an open-source gateway that puts a single OpenAI-shaped interface in front of many model providers, and the repository stands at 58,129 stars.

The function that counted Claude models wrong

The first breaking entry deletes prompt_token_calculator from litellm.utils. The pull request describes it as a second token-counting entry point that duplicated token_counter, and gives the reason for removing it rather than fixing it: the function passed Claude model IDs straight through, so most Claude IDs were counted with OpenAI tokenization instead of Claude's own. The stated migration is to call token_counter instead. The pull request text puts it bluntly: "keeping a second entry point only preserves a trap."

That is worth reading twice if any billing estimate, budget check, or context-fitting logic in your codebase calls it. The failure was quiet. Code that used it did not error, it just returned counts computed with the wrong tokenizer, and for a Claude model the number would have been wrong in every call.

The second breaking entry is narrower. It adds max_retries and extra_headers to the supported OpenAI parameters reported for Cerebras.

Search becomes a provider slot

The release also adds Grounding with Bing Search, registered as bing_grounding, to LiteLLM's search provider system. The pull request says it wires the provider across the unified search runtime, the proxy interception path, pricing metadata, tests, and the dashboard provider picker, so it is reachable through the /v1/search endpoint, through chat websearch interception, through the SDK, and through the admin UI. It supports API keys, Entra tokens, and Azure identity for authentication, and it carries separate per-query pricing for the free and paid modes.

This continues a pattern worth noticing: search is being treated the same way model calls already are, as a slot with interchangeable providers behind one interface, alongside the existing Tavily option.

A gateway earns its place by absorbing provider differences so your application code does not have to know about them. The cost of that bargain is that the gateway's own interface becomes a dependency with its own release cadence, and here that cadence is five days between minor versions with 412 pull requests inside them. A deleted utility function is the cheap version of that bill. The expensive version is finding out months later that a number you trusted was computed with the wrong tokenizer the whole time, which is the specific thing this release fixes by taking the function away rather than repairing it.

Source

LiteLLM v1.100.0 release notes, with the two changes detailed in PR #38132 and PR #38119.