Handle NULL service_tier in ProviderOpenAI

NEWS.md

@@ -5,6 +5,7 @@
       `gen_ai.output.messages`, and `gen_ai.system_instructions`. This is opt-in via the
       `OTEL_INSTRUMENTATION_GENAI_CAPTURE_MESSAGE_CONTENT` environment variable (set to `"true"`), since these payloads may contain user data.
 * The `input` column of `token_usage()` and `parallel_chat()` no longer includes a hidden 1.25x pricing weight for Anthropic cache-creation tokens; the column now reports raw integer token counts. Cost calculations are unchanged.
+* `chat_openai()` now uses the appropriate prices when a non-default service tier (e.g. `"priority"`) is used (@trangdata, #903).
 * `chat_aws_bedrock()` now supports reasoning/thinking content. To enable thinking in Anthropic Claude models, see the `api_args` argument in `?chat_aws_bedrock` for an example (#964).
 * New `chat_lmstudio()` and `models_lmstudio()` provide support for [LM Studio](https://lmstudio.ai), a local model server with an OpenAI-compatible API (#963).
 * Fixed three bugs that caused errors when streaming web search results: Claude's `citations_delta` events were mishandled, `server_tool_use` input wasn't parsed from JSON during streaming, and OpenAI's `web_search_call` failed for non-search action types like `open_page` (#941).

R/provider-openai.R

@@ -333,7 +333,8 @@ method(value_turn, ProviderOpenAI) <- function(
   })
 
   tokens <- value_tokens(provider, result)
-  cost <- get_token_cost(provider, tokens, variant = result$service_tier)
+  variant <- result$service_tier %||% "default"
+  cost <- get_token_cost(provider, tokens, variant = variant)
   AssistantTurn(
     contents = contents,
     json = result,

R/tokens.R

@@ -118,6 +118,8 @@ has_cost <- function(provider, model) {
 }
 
 get_token_cost <- function(provider, tokens, variant = "") {
+  check_string(variant, .internal = TRUE)
+
   needle <- data.frame(
     provider = provider@name,
     model = provider@model,

tests/testthat/_snaps/tokens.md

@@ -13,6 +13,16 @@
             provider model input output cached_input price
       1 testprovider  test     1      1            1 $0.00
 
+# informative internal error if variant is missing
+
+    Code
+      get_token_cost(provider, tokens(), variant = NULL)
+    Condition
+      Error in `get_token_cost()`:
+      ! `variant` must be a single string, not `NULL`.
+      i This is an internal error that was detected in the ellmer package.
+        Please report it at <https://github.com/tidyverse/ellmer/issues> with a reprex (<https://tidyverse.org/help/>) and the full backtrace.
+
 # token_usage() shows price if available
 
     Code

tests/testthat/test-provider-openai.R

@@ -143,6 +143,20 @@ test_that("can extract dummy response from malformed JSON", {
   )
 })
 
+test_that("value_turn handles NULL service_tier gracefully", {
+  provider <- chat_openai_test()$get_provider()
+
+  result <- list(
+    output = list(
+      list(type = "message", content = list(list(text = "Hello")))
+    ),
+    usage = NULL,
+    service_tier = NULL
+  )
+
+  expect_no_error(value_turn(provider, result))
+})
+
 test_that("value_turn() handles web_search_call action types", {
   provider <- chat_openai_test()$get_provider()
 

tests/testthat/test-tokens.R

@@ -55,6 +55,15 @@ test_that("can compute price of tokens with a variant", {
   )
 })
 
+test_that("informative internal error if variant is missing", {
+  provider <- test_provider("OpenAI", "gpt-4o")
+  expect_snapshot(
+    get_token_cost(provider, tokens(), variant = NULL),
+    error = TRUE
+  )
+})
+
+
 test_that("price is NA if we don't have the data for it", {
   provider <- test_provider("ClosedAI", "gpt-4o")
   expect_equal(