Skip to content

Enhanced ChatGPT Clone: Features Agents, MCP, Skills, DeepSeek, Anthropic, AWS, OpenAI, Responses API, Azure, Groq, o1, GPT-5, Mistral, OpenRouter, Vertex AI, Gemini, Artifacts, AI model switching, message search, Code Interpreter, langchain, DALL-E-3, OpenAPI Actions, Functions, Secure Multi-User Auth, Presets, open-source for self-hosting. Active

🧲 refactor: Consolidate Claude Prompt Cache and Context Checks (#15008)

Danny Avila81783b2d52ae
diff --git a/e2e/specs/mock/agents.spec.ts b/e2e/specs/mock/agents.spec.ts
index e935b06dbb..96b129e13f 100644
--- a/e2e/specs/mock/agents.spec.ts
+++ b/e2e/specs/mock/agents.spec.ts
@@ -20,7 +20,6 @@ const MODEL_PARAMETERS = {
topP: 0.8,
topK: 12,
resendFiles: false,
- promptCache: true,
thinking: true,
thinkingBudget: 2000,
web_search: true,
@@ -70,7 +69,9 @@ async function fillAnthropicStyleModelParameters(page: Page) {
await form.locator('#fileTokenLimit-dynamic-input').fill(`${MODEL_PARAMETERS.fileTokenLimit}`);
await setSwitch(form, 'Resend Files', MODEL_PARAMETERS.resendFiles);
- await setSwitch(form, 'Use Prompt Caching', MODEL_PARAMETERS.promptCache);
+ await expect(form.getByRole('switch', { name: 'Use Prompt Caching', exact: true })).toHaveCount(
+ 0,
+ );
await setSwitch(form, 'Thinking', MODEL_PARAMETERS.thinking);
await setSwitch(form, 'Web Search', MODEL_PARAMETERS.web_search);
@@ -107,9 +108,9 @@ async function expectAnthropicStyleModelParameters(page: Page) {
'aria-checked',
String(MODEL_PARAMETERS.resendFiles),
);
- await expect(
- form.getByRole('switch', { name: 'Use Prompt Caching', exact: true }),
- ).toHaveAttribute('aria-checked', String(MODEL_PARAMETERS.promptCache));
+ await expect(form.getByRole('switch', { name: 'Use Prompt Caching', exact: true })).toHaveCount(
+ 0,
+ );
await expect(form.getByRole('switch', { name: 'Thinking', exact: true })).toHaveAttribute(
'aria-checked',
String(MODEL_PARAMETERS.thinking),
diff --git a/packages/api/src/endpoints/anthropic/helpers.ts b/packages/api/src/endpoints/anthropic/helpers.ts
index 6e5c31c3da..6f2692e0e1 100644
--- a/packages/api/src/endpoints/anthropic/helpers.ts
+++ b/packages/api/src/endpoints/anthropic/helpers.ts
@@ -1,16 +1,14 @@
import { logger } from '@librechat/data-schemas';
import { AnthropicClientOptions } from '@librechat/agents';
import {
- EModelEndpoint,
ThinkingDisplay,
AnthropicEffort,
anthropicSettings,
- isMythosClassModel,
resolveThinkingDisplay,
supportsAdaptiveThinking,
+ supportsPromptCache,
requiresExplicitThinkingDisabled,
} from 'librechat-data-provider';
-import { matchModelName } from '~/utils/tokens';
const FINE_GRAINED_TOOL_STREAMING_BETA = 'fine-grained-tool-streaming-2025-05-14';
@@ -37,23 +35,7 @@ function appendAnthropicBetaHeader(
* @returns {boolean}
*/
function checkPromptCacheSupport(modelName: string): boolean {
- const modelMatch = matchModelName(modelName, EModelEndpoint.anthropic) ?? '';
- if (
- modelMatch.includes('claude-3-5-sonnet-latest') ||
- modelMatch.includes('claude-3.5-sonnet-latest')
- ) {
- return false;
- }
-
- return (
- /claude-3[-.]7/.test(modelMatch) ||
- /claude-3[-.]5-(?:sonnet|haiku)/.test(modelMatch) ||
- /claude-3-(?:sonnet|haiku|opus)?/.test(modelMatch) ||
- /claude-(?:sonnet|opus|haiku)-[4-9]/.test(modelMatch) ||
- /claude-[4-9]-(?:sonnet|opus|haiku)?/.test(modelMatch) ||
- /claude-4(?:-(?:sonnet|opus|haiku))?/.test(modelMatch) ||
- isMythosClassModel(modelMatch)
- );
+ return supportsPromptCache(modelName);
}
/**
diff --git a/packages/api/src/endpoints/anthropic/llm.spec.ts b/packages/api/src/endpoints/anthropic/llm.spec.ts
index a664038f37..b3321f6075 100644
--- a/packages/api/src/endpoints/anthropic/llm.spec.ts
+++ b/packages/api/src/endpoints/anthropic/llm.spec.ts
@@ -1864,6 +1864,12 @@ describe('getLLMConfig', () => {
shouldHaveHeaders: false,
shouldHavePromptCache: true,
},
+ {
+ model: 'claude-sonnet-6',
+ promptCache: true,
+ shouldHaveHeaders: false,
+ shouldHavePromptCache: true,
+ },
// Models that support prompt cache but have no additional beta headers needed
{
model: 'claude-3-opus',
diff --git a/packages/api/src/utils/tokens.spec.ts b/packages/api/src/utils/tokens.spec.ts
index 7890af6c7c..9c8eb9de9e 100644
--- a/packages/api/src/utils/tokens.spec.ts
+++ b/packages/api/src/utils/tokens.spec.ts
@@ -7,6 +7,12 @@ describe('getModelMaxTokens partial-override fallback', () => {
'custom-model': { prompt: 1, completion: 2, context: 32000, output: 4096 },
};
+ it('returns undefined for non-string model values from JavaScript consumers', () => {
+ for (const model of [undefined, null, 123]) {
+ expect(getModelMaxTokens(model as unknown as string)).toBeUndefined();
+ }
+ });
+
it('uses the override for a listed model', () => {
expect(getModelMaxTokens('custom-model', EModelEndpoint.openAI, partialOverride)).toBe(32000);
});
@@ -19,6 +25,18 @@ describe('getModelMaxTokens partial-override fallback', () => {
});
});
+describe('future Claude context windows', () => {
+ it('uses the 1M profile for future Sonnet and Opus model IDs', () => {
+ for (const model of ['claude-sonnet-6', 'claude-opus-6']) {
+ expect(getModelMaxTokens(model, EModelEndpoint.anthropic)).toBe(1000000);
+ }
+ });
+
+ it('keeps the safe Claude fallback for unsupported model families', () => {
+ expect(getModelMaxTokens('claude-haiku-4', EModelEndpoint.anthropic)).toBe(100000);
+ });
+});
+
describe('getModelMaxOutputTokens partial-override fallback', () => {
const partialOverride: EndpointTokenConfig = {
'custom-model': { prompt: 1, completion: 2, context: 32000, output: 4096 },
diff --git a/packages/api/src/utils/tokens.ts b/packages/api/src/utils/tokens.ts
index 8656c643b1..b3c543bb94 100644
--- a/packages/api/src/utils/tokens.ts
+++ b/packages/api/src/utils/tokens.ts
@@ -1,5 +1,5 @@
import z from 'zod';
-import { EModelEndpoint } from 'librechat-data-provider';
+import { EModelEndpoint, supportsContext1m } from 'librechat-data-provider';
import type { EndpointTokenConfig, TokenConfig } from '~/types';
/**
@@ -185,7 +185,7 @@ const anthropicModels = {
'claude-mythos-5': 1000000,
};
-const ANTHROPIC_SONNET_4_6_PLUS_CONTEXT = 1000000;
+const ANTHROPIC_CONTEXT_1M = 1000000;
const ANTHROPIC_SONNET_4_6_PLUS_OUTPUT = 128000;
const ANTHROPIC_SONNET_4_6_PLUS_PATTERN =
/(?:claude-sonnet[-.]?4[-.]?(?:[6-9]|\d{2})|claude[-.]?4[-.]?(?:[6-9]|\d{2})[-.]?sonnet)(?=$|[^0-9])/;
@@ -200,14 +200,11 @@ function usesAnthropicContextMap(endpoint: EModelEndpoint): boolean {
);
}
-function getAnthropicSonnet46PlusContext(
- modelName: string,
- endpoint: EModelEndpoint,
-): number | undefined {
- if (!usesAnthropicContextMap(endpoint) || !ANTHROPIC_SONNET_4_6_PLUS_PATTERN.test(modelName)) {
+function getAnthropicContext1m(modelName: string, endpoint: EModelEndpoint): number | undefined {
+ if (!usesAnthropicContextMap(endpoint) || !supportsContext1m(modelName)) {
return undefined;
}
- return ANTHROPIC_SONNET_4_6_PLUS_CONTEXT;
+ return ANTHROPIC_CONTEXT_1M;
}
function getAnthropicSonnet46PlusOutput(
@@ -673,6 +670,10 @@ export function getModelMaxTokens(
endpoint: EModelEndpoint = EModelEndpoint.openAI,
endpointTokenConfig?: EndpointTokenConfig,
): number | undefined {
+ if (typeof modelName !== 'string') {
+ return undefined;
+ }
+
/** A partial override only covers the models it lists; fall back to the
* built-in map for unlisted models instead of dropping to the default
* budget (matches buildTokenConfigMap and getMultiplier). */
@@ -682,9 +683,9 @@ export function getModelMaxTokens(
return overrideValue;
}
}
- const sonnet46PlusValue = getAnthropicSonnet46PlusContext(modelName, endpoint);
- if (sonnet46PlusValue != null) {
- return sonnet46PlusValue;
+ const context1mValue = getAnthropicContext1m(modelName, endpoint);
+ if (context1mValue != null) {
+ return context1mValue;
}
return getModelTokenValue(modelName, maxTokensMap[endpoint as keyof typeof maxTokensMap]);
}
@@ -753,7 +754,7 @@ export function matchModelName(
const matchedPattern = findMatchingPattern(modelName, tokensMap);
if (
(matchedPattern === 'claude-sonnet-4' || matchedPattern === 'claude-4') &&
- getAnthropicSonnet46PlusContext(modelName, endpoint) != null
+ getAnthropicContext1m(modelName, endpoint) != null
) {
return modelName;
}
diff --git a/packages/data-provider/src/bedrock.spec.ts b/packages/data-provider/src/bedrock.spec.ts
new file mode 100644
index 0000000000..05bd4ec1fb
--- /dev/null
+++ b/packages/data-provider/src/bedrock.spec.ts
@@ -0,0 +1,14 @@
+import { supportsContext1m, supportsPromptCache } from './bedrock';
+
+describe('Claude capability helpers', () => {
+ it('recognizes the 1M context window for future Sonnet and Opus model IDs', () => {
+ expect(supportsContext1m('claude-sonnet-6')).toBe(true);
+ expect(supportsContext1m('claude-opus-6')).toBe(true);
+ expect(supportsContext1m('claude-haiku-4')).toBe(false);
+ });
+
+ it('uses raw model IDs to resolve prompt-cache support', () => {
+ expect(supportsPromptCache('claude-sonnet-6')).toBe(true);
+ expect(supportsPromptCache('claude-3-5-sonnet-latest')).toBe(false);
+ });
+});
diff --git a/packages/data-provider/src/bedrock.ts b/packages/data-provider/src/bedrock.ts
index 12e444f259..114cec9164 100644
--- a/packages/data-provider/src/bedrock.ts
+++ b/packages/data-provider/src/bedrock.ts
@@ -275,6 +275,28 @@ export function supportsContext1m(model: string): boolean {
return false;
}
+/**
+ * Checks whether a native Anthropic Claude model supports prompt caching.
+ *
+ * This uses the configured model ID directly. Resolving it through a token
+ * map first can collapse a newly released Claude model to the generic
+ * `claude-` fallback and incorrectly disable cache control.
+ */
+export function supportsPromptCache(model: string): boolean {
+ if (model.includes('claude-3-5-sonnet-latest') || model.includes('claude-3.5-sonnet-latest')) {
+ return false;
+ }
+
+ return (
+ /claude-3[-.]7/.test(model) ||
+ /claude-3[-.]5-(?:sonnet|haiku)/.test(model) ||
+ /claude-3-(?:sonnet|haiku|opus)?/.test(model) ||
+ /claude-(?:sonnet|opus|haiku)[-.]?(?:[4-9]|\d{2,})/.test(model) ||
+ /claude-(?:[4-9]|\d{2,})(?:[-.](?:sonnet|opus|haiku))?/.test(model) ||
+ s.isMythosClassModel(model)
+ );
+}
+
/**
* A Bedrock Claude model ID may be prefixed (`anthropic.claude-*`,
* `us.anthropic.claude-*`, `global.anthropic.claude-*`) or bare (`claude-*`,
diff --git a/packages/data-provider/src/parameterSettings.spec.ts b/packages/data-provider/src/parameterSettings.spec.ts
index a8e269525a..a072459bc1 100644
--- a/packages/data-provider/src/parameterSettings.spec.ts
+++ b/packages/data-provider/src/parameterSettings.spec.ts
@@ -1,9 +1,12 @@
-import { EModelEndpoint } from './types';
-import { applyModelAwareDefaults, paramSettings } from './parameterSettings';
import type { SettingDefinition } from './generate';
+import { applyModelAwareDefaults, paramSettings } from './parameterSettings';
+import { EModelEndpoint } from './types';
const googleParams = paramSettings[EModelEndpoint.google] as SettingDefinition[];
+const anthropicParams = paramSettings[EModelEndpoint.anthropic] as SettingDefinition[];
const maxOut = (params: SettingDefinition[]) => params.find((p) => p.key === 'maxOutputTokens');
+const hasSetting = (params: SettingDefinition[], key: string) =>
+ params.some((param) => param.key === key);
describe('applyModelAwareDefaults', () => {
it('resolves the Google maxOutputTokens default for current Gemini models', () => {
@@ -25,13 +28,32 @@ describe('applyModelAwareDefaults', () => {
expect(maxOut(result)?.default).toBe(32768);
});
- it('returns settings unchanged for non-Google endpoints', () => {
+ it('returns settings unchanged for unrelated endpoints', () => {
+ const result = applyModelAwareDefaults(googleParams, EModelEndpoint.openAI, 'gemini-2.5-pro');
+ expect(result).toBe(googleParams);
+ });
+
+ it('keeps prompt-cache controls for future Claude models that support caching', () => {
const result = applyModelAwareDefaults(
- googleParams,
+ anthropicParams,
EModelEndpoint.anthropic,
- 'gemini-2.5-pro',
+ 'claude-sonnet-6',
);
- expect(result).toBe(googleParams);
+
+ expect(hasSetting(result, 'promptCache')).toBe(true);
+ expect(hasSetting(result, 'promptCacheTtl')).toBe(true);
+ });
+
+ it('hides prompt-cache controls for Anthropic models that do not support caching', () => {
+ const result = applyModelAwareDefaults(
+ anthropicParams,
+ EModelEndpoint.anthropic,
+ 'claude-3-5-sonnet-latest',
+ );
+
+ expect(hasSetting(result, 'promptCache')).toBe(false);
+ expect(hasSetting(result, 'promptCacheTtl')).toBe(false);
+ expect(hasSetting(result, 'temperature')).toBe(true);
});
it('returns settings unchanged when no model is provided', () => {
diff --git a/packages/data-provider/src/parameterSettings.ts b/packages/data-provider/src/parameterSettings.ts
index f2a65b7dff..af2b779119 100644
--- a/packages/data-provider/src/parameterSettings.ts
+++ b/packages/data-provider/src/parameterSettings.ts
@@ -16,6 +16,7 @@ import {
anthropicSettings,
} from './types';
import { SettingDefinition, SettingsConfiguration } from './generate';
+import { supportsPromptCache } from './bedrock';
// Base definitions
const baseDefinitions: Record<string, SettingDefinition> = {
@@ -1253,18 +1254,31 @@ export const agentParamSettings: Record<string, SettingsConfiguration | undefine
* Resolves model-aware defaults for a settings configuration before rendering.
* Google's `maxOutputTokens` default depends on the selected Gemini model so that
* current models (2.5 and 3+) surface their 64K output limit instead of the legacy 8K value.
+ * Anthropic prompt-cache controls are only surfaced for models that support them.
*/
export function applyModelAwareDefaults(
settings: SettingsConfiguration,
endpoint: string,
model?: string,
): SettingsConfiguration {
- if (endpoint !== EModelEndpoint.google || !model) {
+ if (!model) {
return settings;
}
- return settings.map((setting) =>
- setting.key === 'maxOutputTokens'
- ? { ...setting, default: googleSettings.maxOutputTokens.reset(model) }
- : setting,
+
+ const modelAwareSettings =
+ endpoint === EModelEndpoint.google
+ ? settings.map((setting) =>
+ setting.key === 'maxOutputTokens'
+ ? { ...setting, default: googleSettings.maxOutputTokens.reset(model) }
+ : setting,
+ )
+ : settings;
+
+ if (endpoint !== EModelEndpoint.anthropic || supportsPromptCache(model)) {
+ return modelAwareSettings;
+ }
+
+ return modelAwareSettings.filter(
+ (setting) => setting.key !== 'promptCache' && setting.key !== 'promptCacheTtl',
);
}