Resolve the -fast suffix without a model table

README.md

@@ -101,9 +101,10 @@ Use your `server.auth_token` if configured; otherwise `tinyllm` is a client
 placeholder. Choose models your account can access. The prefix is the provider
 table name; the rest is the upstream model ID, including any further slashes.
 Client reasoning effort and service tier override configured model defaults.
-For a configured OpenAI GPT model, append `-fast` (for example,
+For an OpenAI model ID beginning with `gpt-`, append `-fast` (for example,
 `openai/gpt-5.6-sol-fast`) to request the priority tier while sending the base model
-upstream, overriding client and configured service tiers; subscription requests also
+upstream, overriding client and configured service tiers; the base needs no model
+table, matching how unlisted IDs are forwarded; subscription requests also
 carry Codex's routing hint. The suffix does not stack, and upstream decides what it
 delivers: responses may still report `default`. Fast/priority processing can consume
 more credits or cost more.

src/providers/openai/compatible_tests.rs

@@ -301,7 +301,7 @@ fn replayed(details: &Value) -> Value {
 }
 
 #[test]
-fn fast_suffix_resolves_only_configured_gpt_models() {
+fn fast_suffix_resolves_gpt_models() {
     use super::super::models::{Config, ModelOptions, OpenAiAuth};
     use crate::providers::Provider;
     let server = crate::config::Server::default();
@@ -319,15 +319,17 @@ fn fast_suffix_resolves_only_configured_gpt_models() {
         server,
     )
     .unwrap();
-    assert_eq!(
-        provider.synthetic_fast_base("gpt-6-astra-fast"),
-        Some("gpt-6-astra")
-    );
+    for (native, base) in [
+        ("gpt-6-astra-fast", "gpt-6-astra"),
+        // An unconfigured base still resolves; model tables are not an allowlist.
+        ("gpt-9-fast", "gpt-9"),
+    ] {
+        assert_eq!(provider.synthetic_fast_base(native), Some(base), "{native}");
+    }
     for native in [
         "gpt-5.4-fast",      // configured upstream ID stays literal
         "gpt-5.4-fast-fast", // the suffix never stacks
         "o4-mini-fast",      // not a gpt- model
-        "gpt-9-fast",        // base is not configured
         "gpt-6-astra",
     ] {
         assert_eq!(provider.synthetic_fast_base(native), None, "{native}");

src/providers/openai/mod.rs

@@ -47,10 +47,7 @@ impl OpenAiProvider {
         let base = native
             .strip_suffix("-fast")
             .filter(|b| !b.ends_with("-fast"))?;
-        (!self.config.models.contains_key(native)
-            && base.starts_with("gpt-")
-            && self.config.models.contains_key(base))
-        .then_some(base)
+        (!self.config.models.contains_key(native) && base.starts_with("gpt-")).then_some(base)
     }
 
     fn model(&self, native: &str) -> Model {

src/tests.rs

@@ -3269,7 +3269,7 @@ async fn openai_model_defaults_and_client_overrides_reach_each_endpoint() {
             ("gpt-5.4", "gpt-5.4", Some("low"), "low"),
             ("gpt-5.4", "gpt-5.4", None, "xhigh"),
             ("gpt-5.4-fast", "gpt-5.4", None, "xhigh"),
-            ("gpt-9-fast", "gpt-9-fast", Some("low"), "low"),
+            ("gpt-9-fast", "gpt-9", Some("low"), "low"),
         ] {
             let mut body = if path.ends_with("/responses") {
                 json!({"input":"hello"})
@@ -3313,16 +3313,8 @@ async fn openai_model_defaults_and_client_overrides_reach_each_endpoint() {
                 assert_eq!(response["service_tier"], "default");
             }
             match native_model {
-                "gpt-5.4" => assert_eq!(upstream["service_tier"], "priority"),
-                // An unconfigured base forwards the name literally and keeps the client tier.
-                "gpt-9-fast" => assert_eq!(
-                    upstream["service_tier"],
-                    if path.starts_with("/anthropic") {
-                        "default"
-                    } else {
-                        "flex"
-                    }
-                ),
+                // An unconfigured base still resolves, so both take the fast tier.
+                "gpt-5.4" | "gpt-9" => assert_eq!(upstream["service_tier"], "priority"),
                 _ => assert!(upstream.get("service_tier").is_none()),
             }
         }

tinyllm.example.toml

@@ -126,10 +126,10 @@ type = "Subscription"
 # not an allowlist. Unlisted native IDs are forwarded too. Quote IDs containing
 # dots or slashes. Model availability depends on the upstream account.
 # Clients select the main/small/subagent model; there are no configurable aliases.
-# Each configured native ID beginning with gpt- also exposes a generated -fast
-# name that sends the base model with priority. Exact configured names stay literal,
-# so listing an upstream ID that itself ends in -fast keeps it verbatim; the suffix
-# never stacks.
+# Any native ID beginning with gpt- also accepts a generated -fast name that sends
+# the base model with priority, configured or not; discovery lists it for configured
+# IDs only. Exact configured names stay literal, so listing an upstream ID that itself
+# ends in -fast keeps it verbatim; the suffix never stacks.
 [providers.openai.models."gpt-5.6-sol"]
 # Optional fallback when the client omits effort. Client reasoning controls win.
 # Accepted values: none, minimal, low, medium, high, xhigh, max. Providers map