@@ -71,6 +71,11 @@ const DEFAULT_PROVIDER_PREFIXES = [
7171 'openai/' ,
7272 'azure/' ,
7373 'openrouter/openai/' ,
74+ 'zai/' ,
75+ 'deepseek/' ,
76+ 'dashscope/' ,
77+ 'deepinfra/' ,
78+ 'deepinfra/zai-org/' ,
7479] ;
7580
7681function createLogger ( logger ?: PricingLogger ) : PricingLogger {
@@ -86,6 +91,74 @@ function createLogger(logger?: PricingLogger): PricingLogger {
8691 } ;
8792}
8893
94+ const normalizeModelName = ( modelName : string ) : string => modelName . toLowerCase ( ) ;
95+
96+ function getBaseModelName ( modelName : string ) : string {
97+ const normalized = normalizeModelName ( modelName ) ;
98+ const segments = normalized . split ( '/' ) ;
99+ return segments [ segments . length - 1 ] ?? normalized ;
100+ }
101+
102+ /**
103+ * Heuristic scoring so we can prioritize the “right” pricing row when multiple
104+ * candidate keys partially match (e.g. glm-4.5 vs glm-4.5-air vs zai/glm-4.5).
105+ * LiteLLM exposes dozens of provider/model variations, and a plain substring
106+ * check tends to grab the wrong variant (for example the slower “air” edition).
107+ * This mirrors the better-ccusage matching order while keeping the logic lightweight.
108+ */
109+ function calculateModelMatchScore ( targetModel : string , candidateModel : string ) : number {
110+ const normalizedTarget = normalizeModelName ( targetModel ) ;
111+ const normalizedCandidate = normalizeModelName ( candidateModel ) ;
112+
113+ if ( normalizedCandidate === normalizedTarget ) {
114+ return 200 ;
115+ }
116+
117+ if ( normalizedCandidate . endsWith ( `/${ normalizedTarget } ` ) ) {
118+ return 190 ;
119+ }
120+
121+ const targetBase = getBaseModelName ( targetModel ) ;
122+ const candidateBase = getBaseModelName ( candidateModel ) ;
123+
124+ if ( candidateBase === targetBase ) {
125+ let score = 150 ;
126+
127+ if ( ! normalizedCandidate . includes ( 'air' ) ) {
128+ score += 10 ;
129+ }
130+
131+ if ( normalizedCandidate . startsWith ( 'zai/' ) ) {
132+ score += 5 ;
133+ }
134+
135+ return score ;
136+ }
137+
138+ if ( normalizedCandidate . includes ( targetBase ) ) {
139+ let score = 80 ;
140+
141+ if ( ! normalizedCandidate . includes ( 'air' ) ) {
142+ score += 5 ;
143+ }
144+
145+ return score ;
146+ }
147+
148+ if ( targetBase . includes ( candidateBase ) ) {
149+ return 40 ;
150+ }
151+
152+ if (
153+ normalizedCandidate . includes ( normalizedTarget )
154+ || normalizedTarget . includes ( normalizedCandidate )
155+ ) {
156+ return 20 ;
157+ }
158+
159+ return 0 ;
160+ }
161+
89162export class LiteLLMPricingFetcher implements Disposable {
90163 private cachedPricing : Map < string , LiteLLMModelPricing > | null = null ;
91164 private readonly logger : PricingLogger ;
@@ -214,15 +287,29 @@ export class LiteLLMPricingFetcher implements Disposable {
214287 }
215288 }
216289
217- const lower = modelName . toLowerCase ( ) ;
290+ const normalizedTarget = normalizeModelName ( modelName ) ;
291+
292+ let bestMatch : LiteLLMModelPricing | null = null ;
293+ let bestMatchScore = 0 ;
294+
218295 for ( const [ key , value ] of pricing ) {
219- const comparison = key . toLowerCase ( ) ;
220- if ( comparison . includes ( lower ) || lower . includes ( comparison ) ) {
296+ const normalizedCandidate = normalizeModelName ( key ) ;
297+
298+ if (
299+ normalizedCandidate === normalizedTarget
300+ || normalizedCandidate . endsWith ( `/${ normalizedTarget } ` )
301+ ) {
221302 return value ;
222303 }
304+
305+ const score = calculateModelMatchScore ( modelName , key ) ;
306+ if ( score > bestMatchScore ) {
307+ bestMatch = value ;
308+ bestMatchScore = score ;
309+ }
223310 }
224311
225- return null ;
312+ return bestMatchScore > 0 ? bestMatch : null ;
226313 } ) ,
227314 ) ;
228315 }
@@ -508,5 +595,51 @@ if (import.meta.vitest != null) {
508595 } , 'theoretical-model' ) ) ;
509596 expect ( costBelow ) . toBe ( 0 ) ;
510597 } ) ;
598+
599+ it ( 'matches GLM models with provider prefixes and selects primary variants over air versions' , async ( ) => {
600+ using fetcher = new LiteLLMPricingFetcher ( {
601+ offline : true ,
602+ offlineLoader : async ( ) => ( {
603+ 'deepinfra/zai-org/GLM-4.5' : {
604+ input_cost_per_token : 1e-6 ,
605+ output_cost_per_token : 2e-6 ,
606+ } ,
607+ 'zai/glm-4.5-air' : {
608+ input_cost_per_token : 5e-7 ,
609+ output_cost_per_token : 1e-6 ,
610+ } ,
611+ } ) ,
612+ } ) ;
613+
614+ const basePricing = await Result . unwrap ( fetcher . getModelPricing ( 'glm-4.5' ) ) ;
615+ expect ( basePricing ) . not . toBeNull ( ) ;
616+ expect ( basePricing ?. input_cost_per_token ) . toBeCloseTo ( 1e-6 ) ;
617+
618+ const providerPricing = await Result . unwrap ( fetcher . getModelPricing ( 'zai/glm-4.5' ) ) ;
619+ expect ( providerPricing ) . not . toBeNull ( ) ;
620+ expect ( providerPricing ?. output_cost_per_token ) . toBeCloseTo ( 2e-6 ) ;
621+
622+ const airPricing = await Result . unwrap ( fetcher . getModelPricing ( 'glm-4.5-air' ) ) ;
623+ expect ( airPricing ) . not . toBeNull ( ) ;
624+ expect ( airPricing ?. output_cost_per_token ) . toBeCloseTo ( 1e-6 ) ;
625+ } ) ;
626+
627+ describe ( 'model match scoring' , ( ) => {
628+ it ( 'prefers exact matches over provider-suffixed variants' , ( ) => {
629+ expect ( calculateModelMatchScore ( 'glm-4.5' , 'glm-4.5' ) ) . toBeGreaterThan (
630+ calculateModelMatchScore ( 'glm-4.5' , 'provider/glm-4.5' ) ,
631+ ) ;
632+ } ) ;
633+
634+ it ( 'prefers main variants over air editions' , ( ) => {
635+ expect ( calculateModelMatchScore ( 'glm-4.5' , 'provider/glm-4.5' ) ) . toBeGreaterThan (
636+ calculateModelMatchScore ( 'glm-4.5' , 'provider/glm-4.5-air' ) ,
637+ ) ;
638+ } ) ;
639+
640+ it ( 'assigns zero to unrelated models' , ( ) => {
641+ expect ( calculateModelMatchScore ( 'glm-4.5' , 'gpt-4.1-mini' ) ) . toBe ( 0 ) ;
642+ } ) ;
643+ } ) ;
511644 } ) ;
512645}
0 commit comments