promptfoo 0.120.24 → 0.120.26

package/dist/src/{types-CgrxBFgm.js → types-t52w-XsS.js}

@@ -1,5 +1,5 @@
 #!/usr/bin/env node
-import { i as isJavascriptFile, t as JAVASCRIPT_EXTENSIONS } from "./fileExtensions-CzcPkiKZ.js";
+import { i as isJavascriptFile, t as JAVASCRIPT_EXTENSIONS } from "./fileExtensions-BpuMmaFL.js";
 import { z } from "zod";
 import dedent from "dedent";
 
@@ -220,6 +220,7 @@ const FINANCIAL_PLUGINS = [
 	"financial:hallucination",
 	"financial:impartiality",
 	"financial:misconduct",
+	"financial:sox-compliance",
 	"financial:sycophancy"
 ];
 const PHARMACY_PLUGINS = [
@@ -290,6 +291,7 @@ const ADDITIONAL_PLUGINS = [
 	"imitation",
 	"indirect-prompt-injection",
 	"mcp",
+	"model-identification",
 	"medical:anchoring-bias",
 	"medical:hallucination",
 	"medical:incorrect-knowledge",
@@ -305,6 +307,7 @@ const ADDITIONAL_PLUGINS = [
 	"financial:hallucination",
 	"financial:impartiality",
 	"financial:misconduct",
+	"financial:sox-compliance",
 	"financial:sycophancy",
 	"ecommerce:compliance-bypass",
 	"ecommerce:order-fraud",
@@ -430,6 +433,7 @@ const REMOTE_ONLY_PLUGIN_IDS = [
 	"hijacking",
 	"indirect-prompt-injection",
 	"mcp",
+	"model-identification",
 	"off-topic",
 	"rag-document-exfiltration",
 	"rag-poisoning",
@@ -1467,7 +1471,7 @@ const subCategoryDescriptions = {
 	"ecommerce:order-fraud": "Tests for detection of fraudulent order patterns, unauthorized transactions, or payment manipulation attempts",
 	"ecommerce:price-manipulation": "Tests for price manipulation through discount abuse, inventory exploits, or unauthorized price modifications",
 	intent: "Tests for manipulation of system behavior via specific prompts",
-	jailbreak: "Single-shot optimization of safety bypass techniques",
+	jailbreak: "[DEPRECATED] Use jailbreak:meta instead. Single-shot optimization of safety bypass techniques.",
 	"jailbreak:composite": "Combines multiple jailbreak techniques for enhanced effectiveness",
 	"jailbreak:hydra": "Multi-turn conversational attacks with meta-learning that adapts strategy based on full conversation history",
 	"jailbreak:likert": "Uses Likert scale-based prompts to bypass content filters",
@@ -1494,6 +1498,7 @@ const subCategoryDescriptions = {
 	"financial:hallucination": "Tests for fabricated market data, non-existent financial instruments, or fictional company information",
 	"financial:impartiality": "Tests for unauthorized financial advice or transaction facilitation",
 	"financial:misconduct": "Tests for facilitation of financial crimes or market manipulation",
+	"financial:sox-compliance": "Tests for violations of Sarbanes-Oxley requirements including internal controls, financial reporting, and audit integrity",
 	"financial:sycophancy": "Tests for agreeing with risky investment strategies or validating get-rich-quick schemes",
 	"goal-misalignment": "Tests whether AI systems recognize when optimizing proxy metrics might not align with true goals (Goodhart's Law)",
 	"off-topic": "Tests whether AI systems can be manipulated to go off-topic from their intended purpose",
@@ -1528,6 +1533,7 @@ const subCategoryDescriptions = {
 	"sql-injection": "Tests for SQL injection vulnerabilities",
 	ssrf: "Tests for server-side request forgery vulnerabilities",
 	"system-prompt-override": "Tests for system prompt override vulnerabilities",
+	"model-identification": "Tests whether an AI system can be tricked into revealing its underlying model identity",
 	"tool-discovery": "Tests for enumeration of available tools and function calls",
 	unsafebench: "Tests handling of unsafe image content from the UnsafeBench dataset",
 	"unverifiable-claims": "Tests for claims that cannot be verified or fact-checked",
@@ -1612,6 +1618,7 @@ const riskCategorySeverityMap = {
 	"financial:hallucination": Severity.Low,
 	"financial:impartiality": Severity.Medium,
 	"financial:misconduct": Severity.High,
+	"financial:sox-compliance": Severity.High,
 	"financial:sycophancy": Severity.Low,
 	"goal-misalignment": Severity.Low,
 	competitors: Severity.Low,
@@ -1633,6 +1640,7 @@ const riskCategorySeverityMap = {
 	"off-topic": Severity.Medium,
 	"divergent-repetition": Severity.Medium,
 	"excessive-agency": Severity.Medium,
+	"model-identification": Severity.Medium,
 	"tool-discovery": Severity.Low,
 	foundation: Severity.Medium,
 	"guardrails-eval": Severity.Medium,
@@ -1747,6 +1755,7 @@ const riskCategories = {
 		"bola",
 		"cca",
 		"debug-access",
+		"model-identification",
 		"hijacking",
 		"indirect-prompt-injection",
 		"rbac",
@@ -1842,6 +1851,7 @@ const riskCategories = {
 		"financial:hallucination",
 		"financial:impartiality",
 		"financial:misconduct",
+		"financial:sox-compliance",
 		"financial:sycophancy",
 		"medical:hallucination",
 		"medical:anchoring-bias",
@@ -1932,6 +1942,7 @@ const categoryAliases = {
 	"financial:hallucination": "FinancialHallucination",
 	"financial:impartiality": "FinancialImpartiality",
 	"financial:misconduct": "FinancialMisconduct",
+	"financial:sox-compliance": "FinancialSoxCompliance",
 	"financial:sycophancy": "FinancialSycophancy",
 	"goal-misalignment": "GoalMisalignment",
 	"off-topic": "OffTopic",
@@ -1940,6 +1951,7 @@ const categoryAliases = {
 	"pharmacy:drug-interaction": "PharmacyDrugInteraction",
 	"divergent-repetition": "DivergentRepetition",
 	"excessive-agency": "ExcessiveAgency",
+	"model-identification": "ModelIdentification",
 	"tool-discovery": "ToolDiscovery",
 	foundation: "Foundation",
 	"guardrails-eval": "GuardrailsEvaluation",
@@ -2063,6 +2075,7 @@ const pluginDescriptions = {
 	default: "Executes comprehensive baseline security testing across multiple risk categories",
 	"divergent-repetition": "Tests repetitive patterns that can cause the model to diverge from normal behavior and leak training data",
 	"excessive-agency": "Evaluates system boundary enforcement and unauthorized action prevention",
+	"model-identification": "Tests whether an AI system reveals its underlying model identity, version, or creator when probed",
 	"tool-discovery": "Tests for enumeration of available tools and function calls",
 	ferpa: "Tests for Family Educational Rights and Privacy Act compliance",
 	foundation: "Tests a collection of plugins designed to run against foundation models",
@@ -2153,6 +2166,7 @@ const pluginDescriptions = {
 	"financial:hallucination": "Tests for fabricated market data, non-existent financial instruments, fictional company financials, or invented market events",
 	"financial:impartiality": "Tests for inappropriate provision of financial advice, investment recommendations, broker suggestions, or transaction facilitation without proper authorization or disclaimers",
 	"financial:misconduct": "Tests for facilitation of financial crimes including insider trading, market manipulation, front-running, spoofing, bribery, or circumventing financial regulations",
+	"financial:sox-compliance": "Tests for violations of Sarbanes-Oxley requirements including internal controls bypass, financial statement manipulation, CEO/CFO certification fraud, audit obstruction, document retention violations, and whistleblower retaliation",
 	"financial:sycophancy": "Tests for agreeing with risky trading strategies, unrealistic return expectations, or get-rich-quick schemes",
 	"goal-misalignment": "Tests whether AI systems recognize when optimizing measurable proxy metrics might not align with true underlying objectives (Goodhart's Law: \"When a measure becomes a target, it ceases to be a good measure\")",
 	"off-topic": "Tests whether AI systems can be manipulated to go off-topic by performing tasks outside their domain",
@@ -2539,6 +2553,8 @@ const ProviderEnvOverridesSchema = z.object({
 	OPENAI_BASE_URL: z.string().optional(),
 	OPENAI_ORGANIZATION: z.string().optional(),
 	CODEX_API_KEY: z.string().optional(),
+	OPENCLAW_GATEWAY_TOKEN: z.string().optional(),
+	OPENCLAW_GATEWAY_URL: z.string().optional(),
 	PALM_API_HOST: z.string().optional(),
 	PALM_API_KEY: z.string().optional(),
 	PORTKEY_API_KEY: z.string().optional(),
@@ -3418,4 +3434,4 @@ const EvalResultsFilterMode = z.enum([
 
 //#endregion
 export { HARM_PLUGINS as $, DATASET_PLUGINS as A, Severity as B, PolicyObjectSchema as C, ADDITIONAL_STRATEGIES as D, isUuid as E, STRATEGY_COLLECTION_MAPPINGS as F, ALIASED_PLUGIN_MAPPINGS as G, pluginDescriptions as H, getDefaultNFanout as I, BIAS_PLUGINS as J, ADDITIONAL_PLUGINS as K, isCustomStrategy as L, DEFAULT_STRATEGIES as M, MULTI_TURN_STRATEGIES as N, AGENTIC_STRATEGIES as O, STRATEGY_COLLECTIONS as P, FOUNDATION_PLUGINS as Q, isFanoutStrategy as R, PluginConfigSchema as S, isValidReusablePolicyId as T, riskCategorySeverityMap as U, categoryAliases as V, subCategoryDescriptions as W, DEFAULT_PLUGINS as X, DATASET_EXEMPT_PLUGINS as Y, FINANCIAL_PLUGINS as Z, PromptSchema as _, EvaluateOptionsSchema as a, MULTI_INPUT_VAR as at, ConversationMessageSchema as b, TestSuiteConfigSchema as c, PLUGIN_CATEGORIES as ct, isGradingResult as d, REMOTE_ONLY_PLUGIN_IDS as dt, INSURANCE_PLUGINS as et, isResultFailureReason as f, STRATEGY_EXEMPT_PLUGINS as ft, RedteamGenerateOptionsSchema as g, RedteamConfigSchema as h, EvalResultsFilterMode as i, MULTI_INPUT_EXCLUDED_PLUGINS as it, DEFAULT_MULTI_TURN_MAX_TURNS as j, ALL_STRATEGIES as k, TestSuiteSchema as l, REDTEAM_MODEL as lt, isProviderOptions as m, UNALIGNED_PROVIDER_HARM_PLUGINS as mt, BaseAssertionTypesSchema as n, LLAMA_GUARD_REPLICATE_PROVIDER as nt, OutputFileExtension as o, PHARMACY_PLUGINS as ot, isApiProvider as p, TELECOM_PLUGINS as pt, ALL_PLUGINS as q, CommandLineOptionsSchema as r, MEDICAL_PLUGINS as rt, ResultFailureReason as s, PII_PLUGINS as st, AssertionOrSetSchema as t, LLAMA_GUARD_ENABLED_CATEGORIES as tt, UnifiedConfigSchema as u, REDTEAM_PROVIDER_HARM_PLUGINS as ut, ProviderOptionsSchema as v, StrategyConfigSchema as w, PartialGenerationError as x, ProvidersSchema as y, isMultiTurnStrategy as z };
-//# sourceMappingURL=types-CgrxBFgm.js.map
+//# sourceMappingURL=types-t52w-XsS.js.map

package/dist/src/{util-M2MoNCUR.js → util-BSh4a_Q8.js}

@@ -1,11 +1,11 @@
 #!/usr/bin/env node
-import { O as cliState_default, b as getEnvBool, o as logger_default } from "./logger-Bzi5o47S.js";
-import { O as TERMINAL_MAX_WIDTH, P as VERSION, n as fetchWithProxy } from "./fetch-Mb7SbeM-.js";
+import { O as cliState_default, b as getEnvBool, o as logger_default } from "./logger-BotXmWKW.js";
+import { O as TERMINAL_MAX_WIDTH, P as VERSION, n as fetchWithProxy } from "./fetch-SRsE6Ctl.js";
 import { t as invariant } from "./invariant-DT20jrBd.js";
-import { m as isProviderOptions, o as OutputFileExtension, p as isApiProvider, s as ResultFailureReason } from "./types-CgrxBFgm.js";
-import { i as isJavascriptFile, t as JAVASCRIPT_EXTENSIONS } from "./fileExtensions-CzcPkiKZ.js";
-import { o as safeResolve, r as importModule, t as getDirectory } from "./esm-Lh6czAFT.js";
-import { r as runPython } from "./pythonUtils-B9JA-gsC.js";
+import { m as isProviderOptions, o as OutputFileExtension, p as isApiProvider, s as ResultFailureReason } from "./types-t52w-XsS.js";
+import { i as isJavascriptFile, t as JAVASCRIPT_EXTENSIONS } from "./fileExtensions-BpuMmaFL.js";
+import { o as safeResolve, r as importModule, t as getDirectory } from "./esm-CYhseqj4.js";
+import { r as runPython } from "./pythonUtils-r1uBuA0n.js";
 import dotenv from "dotenv";
 import * as fs$2 from "fs";
 import * as path$1 from "path";
@@ -1379,4 +1379,4 @@ function printBorder() {
 
 //#endregion
 export { filterRuntimeVars as A, extractVariablesFromTemplates as C, setupEnv as D, parseFileUrl as E, isAnthropicProvider as F, isGoogleProvider as I, isOpenAiProvider as L, resultIsForTestCase as M, doesProviderRefMatch as N, deduplicateTestCases as O, getProviderDescription as P, isProviderAllowed as R, extractVariablesFromTemplate as S, loadFunction as T, parsePathOrGlob as _, ComparisonEvalNotFoundError as a, renderEnvOnlyInObject as b, mergeComparisonTables as c, getResolvedRelativePath as d, maybeLoadConfigFromExternalFile as f, maybeLoadToolsFromExternalFile as g, maybeLoadResponseFormatFromExternalFile as h, writeOutput as i, getTestCaseDeduplicationKey as j, extractRuntimeVars as k, fetchCsvFromGoogleSheet as l, maybeLoadFromExternalFileWithVars as m, createOutputMetadata as n, evalTableToJson as o, maybeLoadFromExternalFile as p, writeMultipleOutputs as r, generateEvalCsv as s, printBorder as t, getNunjucksEngineForFilePath as u, readFilters as v, getNunjucksEngine as w, renderVarsInObject as x, readOutput as y };
-//# sourceMappingURL=util-M2MoNCUR.js.map
+//# sourceMappingURL=util-BSh4a_Q8.js.map

package/dist/src/{util-DTJWKLkl.cjs → util-C08Kns6-.cjs}

@@ -1,6 +1,6 @@
-const require_logger = require('./logger-CMp-NS-e.cjs');
-const require_util = require('./util-C7A-PlKK.cjs');
-const require_fetch = require('./fetch-MvWxCS5U.cjs');
+const require_logger = require('./logger-BdZ-IqBc.cjs');
+const require_util = require('./util-DUYOvxAy.cjs');
+const require_fetch = require('./fetch-ZMn_oemb.cjs');
 let openai = require("openai");
 openai = require_logger.__toESM(openai);
 
@@ -329,6 +329,20 @@ const OPENAI_CHAT_MODELS = [
 			output: 14 / 1e6
 		}
 	})),
+	...["gpt-5.3-codex"].map((model) => ({
+		id: model,
+		cost: {
+			input: 1.75 / 1e6,
+			output: 14 / 1e6
+		}
+	})),
+	...["gpt-5.3-codex-spark"].map((model) => ({
+		id: model,
+		cost: {
+			input: .5 / 1e6,
+			output: 4 / 1e6
+		}
+	})),
 	...["gpt-audio", "gpt-audio-2025-08-28"].map((model) => ({
 		id: model,
 		cost: {
@@ -609,4 +623,4 @@ Object.defineProperty(exports, 'validateFunctionCall', {
     return validateFunctionCall;
   }
 });
-//# sourceMappingURL=util-DTJWKLkl.cjs.map
+//# sourceMappingURL=util-C08Kns6-.cjs.map

package/dist/src/{util-DHBpsbZE.js → util-CUEt0Vum.js}

@@ -1,7 +1,7 @@
 #!/usr/bin/env node
-import { g as safeJsonStringify, p as getAjv } from "./logger-Bzi5o47S.js";
-import { f as calculateCost } from "./fetch-Mb7SbeM-.js";
-import { m as maybeLoadFromExternalFileWithVars } from "./util-M2MoNCUR.js";
+import { g as safeJsonStringify, p as getAjv } from "./logger-BotXmWKW.js";
+import { f as calculateCost } from "./fetch-SRsE6Ctl.js";
+import { m as maybeLoadFromExternalFileWithVars } from "./util-BSh4a_Q8.js";
 import OpenAI from "openai";
 
 //#region src/providers/openai/util.ts
@@ -329,6 +329,20 @@ const OPENAI_CHAT_MODELS = [
 			output: 14 / 1e6
 		}
 	})),
+	...["gpt-5.3-codex"].map((model) => ({
+		id: model,
+		cost: {
+			input: 1.75 / 1e6,
+			output: 14 / 1e6
+		}
+	})),
+	...["gpt-5.3-codex-spark"].map((model) => ({
+		id: model,
+		cost: {
+			input: .5 / 1e6,
+			output: 4 / 1e6
+		}
+	})),
 	...["gpt-audio", "gpt-audio-2025-08-28"].map((model) => ({
 		id: model,
 		cost: {
@@ -556,4 +570,4 @@ function formatOpenAiError(data) {
 
 //#endregion
 export { calculateOpenAICost as a, getTokenUsage as c, OPENAI_TRANSCRIPTION_MODELS as i, validateFunctionCall as l, OPENAI_COMPLETION_MODELS as n, failApiCall as o, OPENAI_REALTIME_MODELS as r, formatOpenAiError as s, OPENAI_CHAT_MODELS as t };
-//# sourceMappingURL=util-DHBpsbZE.js.map
+//# sourceMappingURL=util-CUEt0Vum.js.map

package/dist/src/{util-Bnw6EyZN.js → util-Cl0zfT3V.js}

@@ -1,6 +1,6 @@
-import { d as getAjv, m as safeJsonStringify } from "./logger-q5I8CByj.js";
-import { l as maybeLoadFromExternalFileWithVars } from "./util-WYC3rB_p.js";
-import { g as calculateCost } from "./fetch-BIfcW-Fd.js";
+import { d as getAjv, m as safeJsonStringify } from "./logger-w8Ozp0Td.js";
+import { l as maybeLoadFromExternalFileWithVars } from "./util-DiCePfDu.js";
+import { g as calculateCost } from "./fetch-18MuNu9i.js";
 import OpenAI from "openai";
 
 //#region src/providers/openai/util.ts
@@ -328,6 +328,20 @@ const OPENAI_CHAT_MODELS = [
 			output: 14 / 1e6
 		}
 	})),
+	...["gpt-5.3-codex"].map((model) => ({
+		id: model,
+		cost: {
+			input: 1.75 / 1e6,
+			output: 14 / 1e6
+		}
+	})),
+	...["gpt-5.3-codex-spark"].map((model) => ({
+		id: model,
+		cost: {
+			input: .5 / 1e6,
+			output: 4 / 1e6
+		}
+	})),
 	...["gpt-audio", "gpt-audio-2025-08-28"].map((model) => ({
 		id: model,
 		cost: {
@@ -555,4 +569,4 @@ function formatOpenAiError(data) {
 
 //#endregion
 export { calculateOpenAICost as a, getTokenUsage as c, OPENAI_TRANSCRIPTION_MODELS as i, validateFunctionCall as l, OPENAI_COMPLETION_MODELS as n, failApiCall as o, OPENAI_REALTIME_MODELS as r, formatOpenAiError as s, OPENAI_CHAT_MODELS as t };
-//# sourceMappingURL=util-Bnw6EyZN.js.map
+//# sourceMappingURL=util-Cl0zfT3V.js.map

package/dist/src/{util-C7A-PlKK.cjs → util-DUYOvxAy.cjs}

@@ -1,10 +1,10 @@
-const require_logger = require('./logger-CMp-NS-e.cjs');
+const require_logger = require('./logger-BdZ-IqBc.cjs');
 const require_invariant = require('./invariant-CKcJAQ6M.cjs');
-const require_esm = require('./esm-gWVPXn3r.cjs');
-const require_pythonUtils = require('./pythonUtils-rOCm9w_5.cjs');
+const require_esm = require('./esm-BQkx5roy.cjs');
+const require_pythonUtils = require('./pythonUtils-DZ6EbdY4.cjs');
 const require_fileExtensions = require('./fileExtensions-DkJYkWUy.cjs');
-const require_types = require('./types-B-XUqfNs.cjs');
-const require_fetch = require('./fetch-MvWxCS5U.cjs');
+const require_types = require('./types-BIfttHrT.cjs');
+const require_fetch = require('./fetch-ZMn_oemb.cjs');
 let fs = require("fs");
 fs = require_logger.__toESM(fs);
 let path = require("path");
@@ -1458,4 +1458,4 @@ Object.defineProperty(exports, 'writeOutput', {
     return writeOutput;
   }
 });
-//# sourceMappingURL=util-C7A-PlKK.cjs.map
+//# sourceMappingURL=util-DUYOvxAy.cjs.map

package/dist/src/{util-WYC3rB_p.js → util-DiCePfDu.js}

@@ -1,10 +1,10 @@
-import { _ as getEnvBool, i as logger_default, w as cliState_default } from "./logger-q5I8CByj.js";
+import { _ as getEnvBool, i as logger_default, w as cliState_default } from "./logger-w8Ozp0Td.js";
 import { t as invariant } from "./invariant-1pAf2CD1.js";
-import { o as safeResolve, r as importModule, t as getDirectory } from "./esm-D40XsWlR.js";
-import { r as runPython } from "./pythonUtils-eNq6Wsfr.js";
+import { o as safeResolve, r as importModule, t as getDirectory } from "./esm-rDtG_2rg.js";
+import { r as runPython } from "./pythonUtils-vMlk9Qp5.js";
 import { i as isJavascriptFile, t as JAVASCRIPT_EXTENSIONS } from "./fileExtensions-ePDqouxn.js";
-import { A as isApiProvider, g as ResultFailureReason, h as OutputFileExtension, j as isProviderOptions } from "./types-BsU_PxR3.js";
-import { L as VERSION, M as TERMINAL_MAX_WIDTH, t as fetchWithProxy } from "./fetch-BIfcW-Fd.js";
+import { A as isApiProvider, g as ResultFailureReason, h as OutputFileExtension, j as isProviderOptions } from "./types-DMVjYLpx.js";
+import { L as VERSION, M as TERMINAL_MAX_WIDTH, t as fetchWithProxy } from "./fetch-18MuNu9i.js";
 import * as fs$1 from "fs";
 import * as path$1 from "path";
 import path from "path";
@@ -1253,4 +1253,4 @@ function printBorder() {
 
 //#endregion
 export { isAnthropicProvider as A, deduplicateTestCases as C, resultIsForTestCase as D, getTestCaseDeduplicationKey as E, isOpenAiProvider as M, isProviderAllowed as N, doesProviderRefMatch as O, setupEnv as S, filterRuntimeVars as T, extractVariablesFromTemplate as _, getNunjucksEngineForFilePath as a, loadFunction as b, maybeLoadFromExternalFile as c, maybeLoadToolsFromExternalFile as d, parsePathOrGlob as f, renderVarsInObject as g, renderEnvOnlyInObject as h, fetchCsvFromGoogleSheet as i, isGoogleProvider as j, getProviderDescription as k, maybeLoadFromExternalFileWithVars as l, readOutput as m, writeMultipleOutputs as n, getResolvedRelativePath as o, readFilters as p, writeOutput as r, maybeLoadConfigFromExternalFile as s, printBorder as t, maybeLoadResponseFormatFromExternalFile as u, extractVariablesFromTemplates as v, extractRuntimeVars as w, parseFileUrl as x, getNunjucksEngine as y };
-//# sourceMappingURL=util-WYC3rB_p.js.map
+//# sourceMappingURL=util-DiCePfDu.js.map

package/dist/src/{util-CV99ps44.cjs → util-DkFTvieG.cjs}

@@ -1,9 +1,16 @@
-const require_fetch = require('./fetch-MvWxCS5U.cjs');
-const require_transform = require('./transform-Bw1IstDE.cjs');
+const require_fetch = require('./fetch-ZMn_oemb.cjs');
+const require_transform = require('./transform-ivxEY4f7.cjs');
 
 //#region src/providers/anthropic/util.ts
 const ANTHROPIC_MODELS = [
-	...["claude-opus-4-6"].map((model) => ({
+	...["claude-sonnet-4-6", "claude-sonnet-4-6-latest"].map((model) => ({
+		id: model,
+		cost: {
+			input: 3 / 1e6,
+			output: 15 / 1e6
+		}
+	})),
+	...["claude-opus-4-6", "claude-opus-4-6-latest"].map((model) => ({
 		id: model,
 		cost: {
 			input: 5 / 1e6,
@@ -204,7 +211,12 @@ function parseMessages(messages) {
 	};
 }
 function calculateAnthropicCost(modelName, config, promptTokens, completionTokens) {
-	if (["claude-sonnet-4-5-20250929"].includes(modelName) && Number.isFinite(promptTokens) && Number.isFinite(completionTokens) && typeof promptTokens !== "undefined" && typeof completionTokens !== "undefined") {
+	if ([
+		"claude-sonnet-4-5-20250929",
+		"claude-sonnet-4-5-latest",
+		"claude-sonnet-4-6",
+		"claude-sonnet-4-6-latest"
+	].includes(modelName) && Number.isFinite(promptTokens) && Number.isFinite(completionTokens) && typeof promptTokens !== "undefined" && typeof completionTokens !== "undefined") {
 		const inputCost = config.cost ?? (promptTokens > 2e5 ? 6 / 1e6 : 3 / 1e6);
 		const outputCost = config.cost ?? (promptTokens > 2e5 ? 22.5 / 1e6 : 15 / 1e6);
 		return inputCost * promptTokens + outputCost * completionTokens;
@@ -314,4 +326,4 @@ Object.defineProperty(exports, 'processAnthropicTools', {
     return processAnthropicTools;
   }
 });
-//# sourceMappingURL=util-CV99ps44.cjs.map
+//# sourceMappingURL=util-DkFTvieG.cjs.map

package/dist/src/{util-CfepsNVK.js → util-mJ58qbbw.js}

@@ -1,9 +1,16 @@
-import { g as calculateCost } from "./fetch-BIfcW-Fd.js";
-import { b as parseDataUrl } from "./transform-BUrxadlA.js";
+import { g as calculateCost } from "./fetch-18MuNu9i.js";
+import { x as parseDataUrl } from "./transform-DMaxQwDx.js";
 
 //#region src/providers/anthropic/util.ts
 const ANTHROPIC_MODELS = [
-	...["claude-opus-4-6"].map((model) => ({
+	...["claude-sonnet-4-6", "claude-sonnet-4-6-latest"].map((model) => ({
+		id: model,
+		cost: {
+			input: 3 / 1e6,
+			output: 15 / 1e6
+		}
+	})),
+	...["claude-opus-4-6", "claude-opus-4-6-latest"].map((model) => ({
 		id: model,
 		cost: {
 			input: 5 / 1e6,
@@ -204,7 +211,12 @@ function parseMessages(messages) {
 	};
 }
 function calculateAnthropicCost(modelName, config, promptTokens, completionTokens) {
-	if (["claude-sonnet-4-5-20250929"].includes(modelName) && Number.isFinite(promptTokens) && Number.isFinite(completionTokens) && typeof promptTokens !== "undefined" && typeof completionTokens !== "undefined") {
+	if ([
+		"claude-sonnet-4-5-20250929",
+		"claude-sonnet-4-5-latest",
+		"claude-sonnet-4-6",
+		"claude-sonnet-4-6-latest"
+	].includes(modelName) && Number.isFinite(promptTokens) && Number.isFinite(completionTokens) && typeof promptTokens !== "undefined" && typeof completionTokens !== "undefined") {
 		const inputCost = config.cost ?? (promptTokens > 2e5 ? 6 / 1e6 : 3 / 1e6);
 		const outputCost = config.cost ?? (promptTokens > 2e5 ? 22.5 / 1e6 : 15 / 1e6);
 		return inputCost * promptTokens + outputCost * completionTokens;
@@ -279,4 +291,4 @@ function transformWebSearchTool(config) {
 
 //#endregion
 export { parseMessages as a, outputFromMessage as i, calculateAnthropicCost as n, processAnthropicTools as o, getTokenUsage as r, ANTHROPIC_MODELS as t };
-//# sourceMappingURL=util-CfepsNVK.js.map
+//# sourceMappingURL=util-mJ58qbbw.js.map

package/dist/src/{util-BZaMVBMq.js → util-vjscpUzy.js}

@@ -1,10 +1,17 @@
 #!/usr/bin/env node
-import { f as calculateCost } from "./fetch-Mb7SbeM-.js";
-import { b as parseDataUrl } from "./transform-Bn-lgBE2.js";
+import { f as calculateCost } from "./fetch-SRsE6Ctl.js";
+import { x as parseDataUrl } from "./transform-CoP2bJ7P.js";
 
 //#region src/providers/anthropic/util.ts
 const ANTHROPIC_MODELS = [
-	...["claude-opus-4-6"].map((model) => ({
+	...["claude-sonnet-4-6", "claude-sonnet-4-6-latest"].map((model) => ({
+		id: model,
+		cost: {
+			input: 3 / 1e6,
+			output: 15 / 1e6
+		}
+	})),
+	...["claude-opus-4-6", "claude-opus-4-6-latest"].map((model) => ({
 		id: model,
 		cost: {
 			input: 5 / 1e6,
@@ -205,7 +212,12 @@ function parseMessages(messages) {
 	};
 }
 function calculateAnthropicCost(modelName, config, promptTokens, completionTokens) {
-	if (["claude-sonnet-4-5-20250929"].includes(modelName) && Number.isFinite(promptTokens) && Number.isFinite(completionTokens) && typeof promptTokens !== "undefined" && typeof completionTokens !== "undefined") {
+	if ([
+		"claude-sonnet-4-5-20250929",
+		"claude-sonnet-4-5-latest",
+		"claude-sonnet-4-6",
+		"claude-sonnet-4-6-latest"
+	].includes(modelName) && Number.isFinite(promptTokens) && Number.isFinite(completionTokens) && typeof promptTokens !== "undefined" && typeof completionTokens !== "undefined") {
 		const inputCost = config.cost ?? (promptTokens > 2e5 ? 6 / 1e6 : 3 / 1e6);
 		const outputCost = config.cost ?? (promptTokens > 2e5 ? 22.5 / 1e6 : 15 / 1e6);
 		return inputCost * promptTokens + outputCost * completionTokens;
@@ -280,4 +292,4 @@ function transformWebSearchTool(config) {
 
 //#endregion
 export { parseMessages as a, outputFromMessage as i, calculateAnthropicCost as n, processAnthropicTools as o, getTokenUsage as r, ANTHROPIC_MODELS as t };
-//# sourceMappingURL=util-BZaMVBMq.js.map
+//# sourceMappingURL=util-vjscpUzy.js.map

package/dist/src/{utils-BryuD3vq.js → utils-CVzb4YiI.js}

@@ -1,5 +1,5 @@
-import { b as getEnvString } from "./logger-q5I8CByj.js";
-import { a as sha256 } from "./extractor-Da8IxCeS.js";
+import { b as getEnvString } from "./logger-w8Ozp0Td.js";
+import { a as sha256 } from "./extractor-Hs7la_19.js";
 
 //#region src/prompts/constants.ts
 const PROMPT_DELIMITER = getEnvString("PROMPTFOO_PROMPT_SEPARATOR") || "---";
@@ -94,4 +94,4 @@ function hashPrompt(prompt) {
 
 //#endregion
 export { PROMPT_DELIMITER as a, generateIdFromPrompt as i, maybeFilePath as n, normalizeInput as r, hashPrompt as t };
-//# sourceMappingURL=utils-BryuD3vq.js.map
+//# sourceMappingURL=utils-CVzb4YiI.js.map

package/dist/src/{utils-Dezi3MBH.cjs → utils-DFaZa6Rf.cjs}

@@ -1,5 +1,5 @@
-const require_logger = require('./logger-CMp-NS-e.cjs');
-const require_extractor = require('./extractor-D-RM2m6A.cjs');
+const require_logger = require('./logger-BdZ-IqBc.cjs');
+const require_extractor = require('./extractor-DbhlYEeo.cjs');
 
 //#region src/prompts/constants.ts
 const PROMPT_DELIMITER = require_logger.getEnvString("PROMPTFOO_PROMPT_SEPARATOR") || "---";
@@ -123,4 +123,4 @@ Object.defineProperty(exports, 'normalizeInput', {
     return normalizeInput;
   }
 });
-//# sourceMappingURL=utils-Dezi3MBH.cjs.map
+//# sourceMappingURL=utils-DFaZa6Rf.cjs.map

package/dist/src/{utils-GiyI2K4P.js → utils-JaY9veb5.js}

@@ -1,6 +1,6 @@
 #!/usr/bin/env node
-import { C as getEnvString } from "./logger-Bzi5o47S.js";
-import { a as sha256 } from "./extractor-BA90EtLq.js";
+import { C as getEnvString } from "./logger-BotXmWKW.js";
+import { a as sha256 } from "./extractor-LSYjrhK0.js";
 
 //#region src/prompts/constants.ts
 const PROMPT_DELIMITER = getEnvString("PROMPTFOO_PROMPT_SEPARATOR") || "---";
@@ -95,4 +95,4 @@ function hashPrompt(prompt) {
 
 //#endregion
 export { PROMPT_DELIMITER as a, generateIdFromPrompt as i, maybeFilePath as n, normalizeInput as r, hashPrompt as t };
-//# sourceMappingURL=utils-GiyI2K4P.js.map
+//# sourceMappingURL=utils-JaY9veb5.js.map