mirror of
https://github.com/danny-avila/LibreChat.git
synced 2026-11-21 15:33:45 +00:00
🎐 feat: Add Claude Haiku 5.5 Support (#17201)
* feat: Add Claude Haiku 5.5 Support Haiku 5.5 (`claude-haiku-5-5`) matched none of the Claude family gates, so every endpoint treated it as a legacy Claude 4 model: thinking went out as `enabled` with a budget (400), sampling parameters were sent, thinking blocks were not bound to the conversation prefix, and it resolved to a 200K context, 64K output and the default token rate. Add a Haiku version parser and `hasAdaptiveHaikuThinking` (5.5 or later), and include it in the adaptive thinking, block binding, omitted display, sampling omission, explicit disabled, effort cap, 1M context and 128K output gates. Unlike Sonnet 5.5, Haiku 5.5 still accepts `disabled` at effort high or below, so "thinking off" sends `disabled` and clamps xhigh/max to high. The settings panel keeps the thinking toggle and hides budget and sampling. Register it in the Anthropic, Bedrock (`global.anthropic.claude-haiku-5-5`) and Vertex default model lists, the 1M context map and the 128K output cap, and price it at $0.10 / $0.50 up to a 100K prompt and $0.50 / $2.50 above, with matching cache tiers. Related to #17078, #16911 and #17025. * test: Cover Claude Haiku 5.5 Parameter Panel * test: Pin Haiku Scenarios to Model Specs * fix: Keep Claude Haiku 5.5 Off Regional Vertex Endpoints
This commit is contained in:
parent
9f1fae7afb
commit
214362aff9
15 changed files with 446 additions and 29 deletions
82
e2e/specs/mock/scenarios/haiku-parameters.spec.ts
Normal file
82
e2e/specs/mock/scenarios/haiku-parameters.spec.ts
Normal file
|
|
@ -0,0 +1,82 @@
|
|||
import { expect, test } from '@playwright/test';
|
||||
import type { Page } from '@playwright/test';
|
||||
import { NEW_CHAT_PATH, selectModelSpec } from '../helpers';
|
||||
import { openPanel } from './panels';
|
||||
|
||||
const PROVIDER = 'Mock Provider A';
|
||||
const HAIKU_55 = 'claude-haiku-5-5';
|
||||
const HAIKU_45 = 'claude-haiku-4-5';
|
||||
|
||||
const SPEC_LABELS: Record<string, string> = {
|
||||
[HAIKU_55]: 'E2E Claude Haiku 5.5',
|
||||
[HAIKU_45]: 'E2E Claude Haiku 4.5',
|
||||
};
|
||||
|
||||
/** Mock Provider A takes Anthropic's parameter set and is reached through model specs, so
|
||||
* each Haiku model gets a spec pinned to it; the model name alone then decides which of
|
||||
* those settings the panel keeps. */
|
||||
async function openParametersFor(page: Page, model: string): Promise<void> {
|
||||
await page.route(
|
||||
(url) => url.pathname === '/api/models',
|
||||
async (route) => {
|
||||
const response = await route.fetch();
|
||||
const models = await response.json();
|
||||
models[PROVIDER] = [...(models[PROVIDER] ?? []), HAIKU_55, HAIKU_45];
|
||||
await route.fulfill({ response, json: models });
|
||||
},
|
||||
);
|
||||
await page.route(
|
||||
(url) => url.pathname === '/api/config',
|
||||
async (route) => {
|
||||
const response = await route.fetch();
|
||||
const config = await response.json();
|
||||
const list = Array.isArray(config.modelSpecs?.list) ? config.modelSpecs.list : [];
|
||||
const specs = [HAIKU_55, HAIKU_45].map((name) => ({
|
||||
name: `e2e-${name}`,
|
||||
label: SPEC_LABELS[name],
|
||||
preset: { endpoint: PROVIDER, model: name },
|
||||
}));
|
||||
await route.fulfill({
|
||||
response,
|
||||
json: {
|
||||
...config,
|
||||
modelSpecs: { ...(config.modelSpecs ?? {}), list: [...list, ...specs] },
|
||||
},
|
||||
});
|
||||
},
|
||||
);
|
||||
await page.goto(NEW_CHAT_PATH, { timeout: 10000 });
|
||||
await selectModelSpec(page, SPEC_LABELS[model]);
|
||||
await openPanel(page, 'parameters', 'Parameters');
|
||||
await expect(page.getByRole('region', { name: 'Reasoning' })).toBeVisible();
|
||||
}
|
||||
|
||||
test.describe('Claude Haiku model parameters', () => {
|
||||
test('Haiku 5.5 offers thinking and effort without a budget or sampling controls @scenario:haiku-55-hides-budget-and-sampling', async ({
|
||||
page,
|
||||
}) => {
|
||||
await openParametersFor(page, HAIKU_55);
|
||||
|
||||
const reasoning = page.getByRole('region', { name: 'Reasoning' });
|
||||
await expect(reasoning.getByRole('switch', { name: /^Thinking/ }).first()).toBeVisible();
|
||||
await expect(reasoning.getByText('Effort', { exact: true }).first()).toBeVisible();
|
||||
|
||||
await expect(page.getByText('Thinking Budget')).toHaveCount(0);
|
||||
await expect(page.getByText('Temperature', { exact: true })).toHaveCount(0);
|
||||
await expect(page.getByText('Top P', { exact: true })).toHaveCount(0);
|
||||
await expect(page.getByText('Top K', { exact: true })).toHaveCount(0);
|
||||
});
|
||||
|
||||
test('Haiku 4.5 keeps the thinking budget and sampling controls @scenario:haiku-45-keeps-budget-and-sampling', async ({
|
||||
page,
|
||||
}) => {
|
||||
await openParametersFor(page, HAIKU_45);
|
||||
|
||||
const reasoning = page.getByRole('region', { name: 'Reasoning' });
|
||||
await expect(reasoning.getByRole('switch', { name: /^Thinking/ }).first()).toBeVisible();
|
||||
await expect(reasoning.getByText('Thinking Budget')).toBeVisible();
|
||||
await expect(page.getByText('Temperature', { exact: true }).first()).toBeVisible();
|
||||
await expect(page.getByText('Top P', { exact: true }).first()).toBeVisible();
|
||||
await expect(page.getByText('Top K', { exact: true }).first()).toBeVisible();
|
||||
});
|
||||
});
|
||||
|
|
@ -1521,6 +1521,66 @@ describe('getLLMConfig', () => {
|
|||
});
|
||||
});
|
||||
|
||||
describe('Haiku 5.5', () => {
|
||||
const HAIKU_55_IDS = ['claude-haiku-5-5', 'claude-haiku-5.5'];
|
||||
|
||||
it.each(HAIKU_55_IDS)(
|
||||
'binds adaptive thinking blocks, omits sampling and caches prompts for %s',
|
||||
(model) => {
|
||||
const result = getLLMConfig('test-key', {
|
||||
modelOptions: {
|
||||
model,
|
||||
thinking: true,
|
||||
thinkingBudget: 5000,
|
||||
effort: AnthropicEffort.max,
|
||||
temperature: 0.7,
|
||||
topP: 0.9,
|
||||
topK: 40,
|
||||
},
|
||||
});
|
||||
|
||||
expect(result.llmConfig.thinking).toEqual({
|
||||
type: 'adaptive',
|
||||
display: 'summarized',
|
||||
block_binding: { prefix_mismatch_behavior: 'drop_block' },
|
||||
});
|
||||
expect(result.llmConfig.outputConfig).toEqual({ effort: AnthropicEffort.max });
|
||||
expect(result.llmConfig).not.toHaveProperty('temperature');
|
||||
expect(result.llmConfig).not.toHaveProperty('topP');
|
||||
expect(result.llmConfig).not.toHaveProperty('topK');
|
||||
expect(result.llmConfig.maxTokens).toBe(128000);
|
||||
expect(result.llmConfig).toHaveProperty('promptCache', true);
|
||||
const beta = (result.llmConfig.clientOptions?.defaultHeaders as Record<string, string>)[
|
||||
'anthropic-beta'
|
||||
];
|
||||
expect(beta).toContain('thinking-binding-controls-2026-08-01');
|
||||
},
|
||||
);
|
||||
|
||||
it.each([AnthropicEffort.xhigh, AnthropicEffort.max])(
|
||||
'sends disabled thinking and clamps effort %s to high when thinking is off',
|
||||
(effort) => {
|
||||
const { llmConfig } = getLLMConfig('test-key', {
|
||||
modelOptions: { model: 'claude-haiku-5-5', thinking: false, effort },
|
||||
});
|
||||
const Anthropic = getChatModelClass(Providers.ANTHROPIC);
|
||||
const payload = new Anthropic(llmConfig).invocationParams();
|
||||
|
||||
expect(payload.thinking).toEqual({ type: 'disabled' });
|
||||
expect(payload.output_config).toEqual({ effort: AnthropicEffort.high });
|
||||
expect(payload).not.toHaveProperty('temperature');
|
||||
},
|
||||
);
|
||||
|
||||
it('keeps Haiku 4.5 on budgeted thinking', () => {
|
||||
const result = getLLMConfig('test-key', {
|
||||
modelOptions: { model: 'claude-haiku-4-5', thinking: true, thinkingBudget: 2000 },
|
||||
});
|
||||
|
||||
expect(result.llmConfig.thinking).toEqual({ type: 'enabled', budget_tokens: 2000 });
|
||||
});
|
||||
});
|
||||
|
||||
it('should omit sampling parameters for Opus 5', () => {
|
||||
const result = getLLMConfig('test-key', {
|
||||
modelOptions: {
|
||||
|
|
@ -1852,13 +1912,8 @@ describe('getLLMConfig', () => {
|
|||
});
|
||||
});
|
||||
|
||||
it('should future-proof Claude 5.x Haiku models with 64K default', () => {
|
||||
const testCases = [
|
||||
'claude-haiku-5',
|
||||
'claude-haiku-5-0',
|
||||
'claude-haiku-5-2-20260101',
|
||||
'claude-haiku-5.5',
|
||||
];
|
||||
it('should future-proof Claude 5.x Haiku models before 5.5 with 64K default', () => {
|
||||
const testCases = ['claude-haiku-5', 'claude-haiku-5-0', 'claude-haiku-5-2-20260101'];
|
||||
|
||||
testCases.forEach((model) => {
|
||||
const result = getLLMConfig('test-key', {
|
||||
|
|
@ -1868,6 +1923,17 @@ describe('getLLMConfig', () => {
|
|||
});
|
||||
});
|
||||
|
||||
it('should give Claude Haiku 5.5+ models the 128K default', () => {
|
||||
const testCases = ['claude-haiku-5-5', 'claude-haiku-5.5', 'claude-haiku-5-6'];
|
||||
|
||||
testCases.forEach((model) => {
|
||||
const result = getLLMConfig('test-key', {
|
||||
modelOptions: { model },
|
||||
});
|
||||
expect(result.llmConfig.maxTokens).toBe(128000);
|
||||
});
|
||||
});
|
||||
|
||||
it('should future-proof Claude 5.x Opus models with 128K default', () => {
|
||||
const testCases = [
|
||||
'claude-opus-5',
|
||||
|
|
@ -1886,21 +1952,21 @@ describe('getLLMConfig', () => {
|
|||
|
||||
it('should future-proof Claude 6-9.x models with correct defaults', () => {
|
||||
const testCases = [
|
||||
// Claude 6.x - Sonnet/Opus get 128K, Haiku gets 64K
|
||||
// Claude 6.x - Sonnet, Opus and Haiku (5.5+) get 128K
|
||||
{ model: 'claude-sonnet-6', expected: 128000 },
|
||||
{ model: 'claude-haiku-6-0', expected: 64000 },
|
||||
{ model: 'claude-haiku-6-0', expected: 128000 },
|
||||
{ model: 'claude-opus-6-1', expected: 128000 },
|
||||
// Claude 7.x
|
||||
{ model: 'claude-sonnet-7-20270101', expected: 128000 },
|
||||
{ model: 'claude-haiku-7.5', expected: 64000 },
|
||||
{ model: 'claude-haiku-7.5', expected: 128000 },
|
||||
{ model: 'claude-opus-7', expected: 128000 },
|
||||
// Claude 8.x
|
||||
{ model: 'claude-sonnet-8', expected: 128000 },
|
||||
{ model: 'claude-haiku-8-2', expected: 64000 },
|
||||
{ model: 'claude-haiku-8-2', expected: 128000 },
|
||||
{ model: 'claude-opus-8-latest', expected: 128000 },
|
||||
// Claude 9.x
|
||||
{ model: 'claude-sonnet-9', expected: 128000 },
|
||||
{ model: 'claude-haiku-9', expected: 64000 },
|
||||
{ model: 'claude-haiku-9', expected: 128000 },
|
||||
{ model: 'claude-opus-9', expected: 128000 },
|
||||
];
|
||||
|
||||
|
|
|
|||
|
|
@ -305,6 +305,22 @@ describe('Opus 5.5 token limits', () => {
|
|||
});
|
||||
});
|
||||
|
||||
describe('Haiku 5.5 token limits', () => {
|
||||
it.each([
|
||||
'claude-haiku-5-5',
|
||||
'claude-haiku-5.5',
|
||||
'anthropic/claude-haiku-5-5',
|
||||
'global.anthropic.claude-haiku-5-5',
|
||||
])('resolves %s to the modern Claude profile', (model) => {
|
||||
expect(getModelMaxTokens(model)).toBe(1000000);
|
||||
expect(getModelMaxTokens(model, EModelEndpoint.anthropic)).toBe(1000000);
|
||||
});
|
||||
|
||||
it('keeps Haiku 4.5 at a 200K context', () => {
|
||||
expect(getModelMaxTokens('claude-haiku-4-5', EModelEndpoint.anthropic)).toBe(200000);
|
||||
});
|
||||
});
|
||||
|
||||
describe('Sonnet 5.5 token limits', () => {
|
||||
it.each([
|
||||
'claude-sonnet-5-5',
|
||||
|
|
|
|||
|
|
@ -168,6 +168,8 @@ const anthropicModels = {
|
|||
'claude-3.7-sonnet': 200000,
|
||||
'claude-4': 200000,
|
||||
'claude-haiku-4-5': 200000,
|
||||
'claude-haiku-5-5': 1000000,
|
||||
'claude-haiku-5.5': 1000000,
|
||||
'claude-opus-4': 200000,
|
||||
'claude-opus-4-5': 200000,
|
||||
'claude-sonnet-4': 200000,
|
||||
|
|
|
|||
|
|
@ -23,6 +23,7 @@ import {
|
|||
supportsPromptCache,
|
||||
supportsContext1m,
|
||||
hasAlwaysOnThinking,
|
||||
hasAdaptiveHaikuThinking,
|
||||
hasBetweenToolsThinkingFloor,
|
||||
supportsOutput128k,
|
||||
BEDROCK_FINE_GRAINED_TOOL_STREAMING_BETA,
|
||||
|
|
@ -569,6 +570,62 @@ describe('Sonnet 5.5 model gates', () => {
|
|||
});
|
||||
});
|
||||
|
||||
describe('Haiku 5.5 model gates', () => {
|
||||
const ids = [
|
||||
'claude-haiku-5-5',
|
||||
'claude-haiku-5.5',
|
||||
'anthropic/claude-haiku-5-5',
|
||||
'global.anthropic.claude-haiku-5-5',
|
||||
];
|
||||
|
||||
test.each(ids)('%s is detected across every family gate', (model) => {
|
||||
expect(hasAdaptiveHaikuThinking(model)).toBe(true);
|
||||
expect(hasAlwaysOnThinking(model)).toBe(false);
|
||||
expect(hasBetweenToolsThinkingFloor(model)).toBe(false);
|
||||
expect(bindsThinkingBlocks(model)).toBe(true);
|
||||
expect(supportsAdaptiveThinking(model)).toBe(true);
|
||||
expect(omitsThinkingByDefault(model)).toBe(true);
|
||||
expect(omitsSamplingParameters(model)).toBe(true);
|
||||
expect(supportsPromptCache(model)).toBe(true);
|
||||
expect(supportsContext1m(model)).toBe(true);
|
||||
expect(supportsOutput128k(model)).toBe(true);
|
||||
expect(requiresExplicitThinkingDisabled(model)).toBe(true);
|
||||
expect(capsEffortWhenThinkingDisabled(model)).toBe(true);
|
||||
expect(resolveThinkingOffConfig(model)).toEqual({ type: 'disabled' });
|
||||
expect(resolveThinkingDisplay(model)).toBe('summarized');
|
||||
});
|
||||
|
||||
test.each(['claude-haiku-5-6', 'claude-haiku-6', 'claude-6-haiku'])(
|
||||
'later Haiku release %s inherits the 5.5 contract',
|
||||
(model) => {
|
||||
expect(hasAdaptiveHaikuThinking(model)).toBe(true);
|
||||
expect(supportsAdaptiveThinking(model)).toBe(true);
|
||||
expect(bindsThinkingBlocks(model)).toBe(true);
|
||||
},
|
||||
);
|
||||
|
||||
test.each([
|
||||
'claude-haiku-4-5',
|
||||
'claude-haiku-4-5-20251001',
|
||||
'global.anthropic.claude-haiku-4-5-20251001-v1:0',
|
||||
'claude-3-5-haiku-20241022',
|
||||
'claude-haiku-5-20260101',
|
||||
])('leaves %s on the legacy budget path', (model) => {
|
||||
expect(hasAdaptiveHaikuThinking(model)).toBe(false);
|
||||
expect(supportsAdaptiveThinking(model)).toBe(false);
|
||||
expect(bindsThinkingBlocks(model)).toBe(false);
|
||||
expect(omitsSamplingParameters(model)).toBe(false);
|
||||
expect(supportsContext1m(model)).toBe(false);
|
||||
expect(resolveThinkingOffConfig(model)).toBeUndefined();
|
||||
});
|
||||
|
||||
test('clamps xhigh and max to high while thinking is disabled', () => {
|
||||
expect(clampEffortForDisabledThinking('claude-haiku-5-5', 'xhigh')).toBe('high');
|
||||
expect(clampEffortForDisabledThinking('claude-haiku-5-5', 'max')).toBe('high');
|
||||
expect(clampEffortForDisabledThinking('claude-haiku-5-5', 'medium')).toBe('medium');
|
||||
});
|
||||
});
|
||||
|
||||
describe('always-on thinking gates', () => {
|
||||
test.each([
|
||||
'claude-opus-5-5',
|
||||
|
|
@ -610,6 +667,39 @@ describe('always-on thinking gates', () => {
|
|||
});
|
||||
|
||||
describe('bedrockInputParser', () => {
|
||||
test('sends bound adaptive thinking for Haiku 5.5 and drops the legacy budget', () => {
|
||||
const result = bedrockInputParser.parse({
|
||||
model: 'global.anthropic.claude-haiku-5-5',
|
||||
thinking: true,
|
||||
thinkingBudget: 4000,
|
||||
effort: 'low',
|
||||
}) as Record<string, unknown>;
|
||||
const additionalFields = result.additionalModelRequestFields as Record<string, unknown>;
|
||||
|
||||
expect(additionalFields.thinking).toEqual({
|
||||
type: 'adaptive',
|
||||
display: 'summarized',
|
||||
block_binding: { prefix_mismatch_behavior: 'drop_block' },
|
||||
});
|
||||
expect(additionalFields.output_config).toEqual({ effort: 'low' });
|
||||
expect(additionalFields).not.toHaveProperty('thinkingBudget');
|
||||
expect(additionalFields.anthropic_beta).toEqual(['thinking-binding-controls-2026-08-01']);
|
||||
});
|
||||
|
||||
test('maps Haiku 5.5 thinking off to an explicit disabled config and caps effort', () => {
|
||||
const result = bedrockInputParser.parse({
|
||||
model: 'global.anthropic.claude-haiku-5-5',
|
||||
thinking: false,
|
||||
thinkingDisplay: 'summarized',
|
||||
effort: 'max',
|
||||
}) as Record<string, unknown>;
|
||||
const additionalFields = result.additionalModelRequestFields as Record<string, unknown>;
|
||||
|
||||
expect(additionalFields.thinking).toEqual({ type: 'disabled' });
|
||||
expect(additionalFields.output_config).toEqual({ effort: 'high' });
|
||||
expect(additionalFields).not.toHaveProperty('thinkingDisplay');
|
||||
});
|
||||
|
||||
test('maps Sonnet 5.5 thinking off to between_tools and caps effort', () => {
|
||||
const result = bedrockInputParser.parse({
|
||||
model: 'global.anthropic.claude-sonnet-5-5',
|
||||
|
|
@ -972,9 +1062,9 @@ describe('bedrockInputParser', () => {
|
|||
expect(amrf?.anthropic_beta).toEqual(['context-1m-2025-08-07']);
|
||||
});
|
||||
|
||||
test('should match anthropic.claude-haiku-6 model without context beta header', () => {
|
||||
test('should match anthropic.claude-haiku-4-5 model without context beta header', () => {
|
||||
const input = {
|
||||
model: 'anthropic.claude-haiku-6',
|
||||
model: 'anthropic.claude-haiku-4-5',
|
||||
};
|
||||
const result = bedrockInputParser.parse(input) as Record<string, unknown>;
|
||||
const additionalFields = result.additionalModelRequestFields as Record<string, unknown>;
|
||||
|
|
|
|||
|
|
@ -138,9 +138,13 @@ export const THINKING_BLOCK_BINDING = {
|
|||
prefix_mismatch_behavior: 'drop_block',
|
||||
} as const;
|
||||
|
||||
/** Models whose thinking blocks are bound to the conversation prefix (Opus 5.5+, Sonnet 5.5+, Fable/Mythos). */
|
||||
/** Models whose thinking blocks are bound to the conversation prefix (Opus 5.5+, Sonnet 5.5+, Haiku 5.5+, Fable/Mythos). */
|
||||
export function bindsThinkingBlocks(model: string): boolean {
|
||||
return hasAlwaysOnThinking(model) || hasBetweenToolsThinkingFloor(model);
|
||||
return (
|
||||
hasAlwaysOnThinking(model) ||
|
||||
hasBetweenToolsThinkingFloor(model) ||
|
||||
hasAdaptiveHaikuThinking(model)
|
||||
);
|
||||
}
|
||||
|
||||
export const BETWEEN_TOOLS_THINKING = { type: 'between_tools' } as const;
|
||||
|
|
@ -181,17 +185,46 @@ function parseSonnetVersion(model: string): ClaudeVersion | null {
|
|||
return null;
|
||||
}
|
||||
|
||||
/** Extracts haiku major/minor version from both naming formats. */
|
||||
function parseHaikuVersion(model: string): ClaudeVersion | null {
|
||||
const nameFirst = model.match(/claude-haiku[-.]?(\d+)(?:[-.](\d{1,2})(?!\d))?/);
|
||||
if (nameFirst) {
|
||||
return {
|
||||
major: parseInt(nameFirst[1], 10),
|
||||
minor: nameFirst[2] != null ? parseInt(nameFirst[2], 10) : 0,
|
||||
};
|
||||
}
|
||||
const numFirst = model.match(/claude-(\d+)(?:[-.](\d{1,2})(?!\d))?-haiku/);
|
||||
if (numFirst) {
|
||||
return {
|
||||
major: parseInt(numFirst[1], 10),
|
||||
minor: numFirst[2] != null ? parseInt(numFirst[2], 10) : 0,
|
||||
};
|
||||
}
|
||||
return null;
|
||||
}
|
||||
|
||||
/**
|
||||
* Haiku 5.5 and later run adaptive thinking on by default, bound to the
|
||||
* conversation prefix, and reject budgets and sampling parameters. Unlike
|
||||
* Sonnet 5.5 they still accept an explicit `disabled`, at effort `high` or below.
|
||||
*/
|
||||
export function hasAdaptiveHaikuThinking(model: string): boolean {
|
||||
return isAtLeast(parseHaikuVersion(model), 5, 5);
|
||||
}
|
||||
|
||||
/**
|
||||
* Mythos-class detection (Claude Fable / Mythos) lives in `schemas.ts` as
|
||||
* `isMythosClassModel` — the single source of truth for the family names.
|
||||
* The helpers below OR it in alongside the `opus`/`sonnet` version parsers.
|
||||
*/
|
||||
|
||||
/** Checks if a model supports adaptive thinking (Opus 4.6+, Sonnet 4.6+, Fable/Mythos) */
|
||||
/** Checks if a model supports adaptive thinking (Opus 4.6+, Sonnet 4.6+, Haiku 5.5+, Fable/Mythos) */
|
||||
export function supportsAdaptiveThinking(model: string): boolean {
|
||||
return (
|
||||
isAtLeast(parseOpusVersion(model), 4, 6) ||
|
||||
isAtLeast(parseSonnetVersion(model), 4, 6) ||
|
||||
hasAdaptiveHaikuThinking(model) ||
|
||||
s.isMythosClassModel(model)
|
||||
);
|
||||
}
|
||||
|
|
@ -210,6 +243,7 @@ export function omitsThinkingByDefault(model: string): boolean {
|
|||
return (
|
||||
isAtLeast(parseOpusVersion(model), 4, 7) ||
|
||||
isAtLeast(parseSonnetVersion(model), 5, 0) ||
|
||||
hasAdaptiveHaikuThinking(model) ||
|
||||
s.isMythosClassModel(model)
|
||||
);
|
||||
}
|
||||
|
|
@ -218,6 +252,7 @@ export function omitsSamplingParameters(model: string): boolean {
|
|||
return (
|
||||
isAtLeast(parseOpusVersion(model), 4, 7) ||
|
||||
isAtLeast(parseSonnetVersion(model), 5, 0) ||
|
||||
hasAdaptiveHaikuThinking(model) ||
|
||||
s.isMythosClassModel(model)
|
||||
);
|
||||
}
|
||||
|
|
@ -226,7 +261,7 @@ export function omitsSamplingParameters(model: string): boolean {
|
|||
* Whether disabling thinking requires sending an explicit `{ type: 'disabled' }`
|
||||
* config rather than simply omitting the `thinking` field.
|
||||
*
|
||||
* Sonnet 5 and Opus 5 treat an omitted `thinking` field as adaptive thinking ON
|
||||
* Sonnet 5, Opus 5 and Haiku 5.5+ treat an omitted `thinking` field as adaptive thinking ON
|
||||
* by default, so honoring a user who turns thinking off means sending the
|
||||
* disabled config explicitly. Opus 4.7/4.8 run without thinking when the field
|
||||
* is omitted, and Fable/Mythos, Opus 5.5+ and Sonnet 5.5+ reject an explicit
|
||||
|
|
@ -239,7 +274,11 @@ export function requiresExplicitThinkingDisabled(model: string): boolean {
|
|||
if (hasAlwaysOnThinking(model) || hasBetweenToolsThinkingFloor(model)) {
|
||||
return false;
|
||||
}
|
||||
return isAtLeast(parseSonnetVersion(model), 5, 0) || isAtLeast(parseOpusVersion(model), 5, 0);
|
||||
return (
|
||||
isAtLeast(parseSonnetVersion(model), 5, 0) ||
|
||||
isAtLeast(parseOpusVersion(model), 5, 0) ||
|
||||
hasAdaptiveHaikuThinking(model)
|
||||
);
|
||||
}
|
||||
|
||||
/** Effort levels Opus 5 rejects while thinking is explicitly disabled. */
|
||||
|
|
@ -254,8 +293,9 @@ const EFFORTS_REJECTED_WHEN_THINKING_DISABLED = new Set<string>([
|
|||
* Opus 5 rejects `xhigh`/`max` in that combination with a 400: "output_config
|
||||
* .effort 'xhigh' is not supported when thinking is disabled on this model. Use
|
||||
* effort 'high' or below, or enable thinking." Sonnet 5.5+ returns the same
|
||||
* error for `between_tools`. Opus 4.7/4.8, Sonnet 5, and Sonnet 4.6 accept
|
||||
* every effort level they otherwise support with thinking off. Always-on
|
||||
* error for `between_tools`, and Haiku 5.5+ for `disabled`. Opus 4.7/4.8,
|
||||
* Sonnet 5, and Sonnet 4.6 accept every effort level they otherwise support
|
||||
* with thinking off. Always-on
|
||||
* models cannot turn thinking off, so the cap never applies to them.
|
||||
*/
|
||||
export function capsEffortWhenThinkingDisabled(model: string): boolean {
|
||||
|
|
@ -265,7 +305,7 @@ export function capsEffortWhenThinkingDisabled(model: string): boolean {
|
|||
if (hasBetweenToolsThinkingFloor(model)) {
|
||||
return true;
|
||||
}
|
||||
return isAtLeast(parseOpusVersion(model), 5, 0);
|
||||
return isAtLeast(parseOpusVersion(model), 5, 0) || hasAdaptiveHaikuThinking(model);
|
||||
}
|
||||
|
||||
/**
|
||||
|
|
@ -323,20 +363,22 @@ export function isThinkingOffConfig(thinking: unknown): boolean {
|
|||
return thinking.type === 'disabled' || thinking.type === 'between_tools';
|
||||
}
|
||||
|
||||
/** Checks if a model has a 1M context window (Sonnet 4.6+, Opus 4.6+, Opus 5+, Fable/Mythos) */
|
||||
/** Checks if a model has a 1M context window (Sonnet 4.6+, Opus 4.6+, Opus 5+, Haiku 5.5+, Fable/Mythos) */
|
||||
export function supportsContext1m(model: string): boolean {
|
||||
return (
|
||||
isAtLeast(parseSonnetVersion(model), 4, 6) ||
|
||||
isAtLeast(parseOpusVersion(model), 4, 6) ||
|
||||
hasAdaptiveHaikuThinking(model) ||
|
||||
s.isMythosClassModel(model)
|
||||
);
|
||||
}
|
||||
|
||||
/** Opus 5.5+, Sonnet 5.5+ and Fable/Mythos emit up to 128k output tokens. */
|
||||
/** Opus 5.5+, Sonnet 5.5+, Haiku 5.5+ and Fable/Mythos emit up to 128k output tokens. */
|
||||
export function supportsOutput128k(model: string): boolean {
|
||||
return (
|
||||
isAtLeast(parseOpusVersion(model), 5, 5) ||
|
||||
isAtLeast(parseSonnetVersion(model), 5, 5) ||
|
||||
hasAdaptiveHaikuThinking(model) ||
|
||||
s.isMythosClassModel(model)
|
||||
);
|
||||
}
|
||||
|
|
|
|||
|
|
@ -4508,6 +4508,7 @@ const sharedAnthropicModels = [
|
|||
'claude-opus-4-6',
|
||||
'claude-sonnet-4-5',
|
||||
'claude-sonnet-4-5-20250929',
|
||||
'claude-haiku-5-5',
|
||||
'claude-haiku-4-5',
|
||||
'claude-haiku-4-5-20251001',
|
||||
'claude-opus-4-1',
|
||||
|
|
@ -4545,6 +4546,7 @@ export const bedrockModels = [
|
|||
'global.anthropic.claude-sonnet-4-6',
|
||||
'global.anthropic.claude-opus-4-6-v1',
|
||||
'global.anthropic.claude-sonnet-4-5-20250929-v1:0',
|
||||
'global.anthropic.claude-haiku-5-5',
|
||||
'global.anthropic.claude-haiku-4-5-20251001-v1:0',
|
||||
'us.anthropic.claude-opus-4-1-20250805-v1:0',
|
||||
// 'cohere.command-text-v14', // no conversation history
|
||||
|
|
|
|||
|
|
@ -108,6 +108,35 @@ describe('applyModelAwareDefaults', () => {
|
|||
},
|
||||
);
|
||||
|
||||
it.each(['claude-haiku-5-5', 'global.anthropic.claude-haiku-5-5'])(
|
||||
'keeps the thinking toggle but hides budget and sampling controls for %s',
|
||||
(model) => {
|
||||
const result = applyModelAwareDefaults(anthropicParams, EModelEndpoint.anthropic, model);
|
||||
|
||||
expect(hasSetting(result, 'thinking')).toBe(true);
|
||||
expect(result.find((setting) => setting.key === 'thinking')?.description).not.toBe(
|
||||
'com_endpoint_anthropic_thinking_between_tools',
|
||||
);
|
||||
expect(hasSetting(result, 'thinkingBudget')).toBe(false);
|
||||
expect(hasSetting(result, 'temperature')).toBe(false);
|
||||
expect(hasSetting(result, 'topP')).toBe(false);
|
||||
expect(hasSetting(result, 'topK')).toBe(false);
|
||||
expect(hasSetting(result, 'effort')).toBe(true);
|
||||
expect(hasSetting(result, 'promptCache')).toBe(true);
|
||||
},
|
||||
);
|
||||
|
||||
it('keeps budget and sampling controls for Haiku 4.5', () => {
|
||||
const result = applyModelAwareDefaults(
|
||||
anthropicParams,
|
||||
EModelEndpoint.anthropic,
|
||||
'claude-haiku-4-5',
|
||||
);
|
||||
|
||||
expect(hasSetting(result, 'thinkingBudget')).toBe(true);
|
||||
expect(hasSetting(result, 'temperature')).toBe(true);
|
||||
});
|
||||
|
||||
it('returns settings unchanged when no model is provided', () => {
|
||||
expect(applyModelAwareDefaults(googleParams, EModelEndpoint.google, '')).toBe(googleParams);
|
||||
});
|
||||
|
|
|
|||
|
|
@ -21,6 +21,7 @@ import {
|
|||
} from './types';
|
||||
import {
|
||||
hasAlwaysOnThinking,
|
||||
hasAdaptiveHaikuThinking,
|
||||
hasBetweenToolsThinkingFloor,
|
||||
supportsPromptCache,
|
||||
supportsAdaptiveThinking,
|
||||
|
|
@ -1773,6 +1774,12 @@ function applyModelFamilyDefaults(
|
|||
(setting) => !['thinkingBudget', 'temperature', 'topP', 'topK'].includes(setting.key),
|
||||
);
|
||||
}
|
||||
/** Haiku 5.5+ keeps the toggle ("off" sends `disabled`) but rejects budgets and sampling. */
|
||||
if (hasAdaptiveHaikuThinking(model)) {
|
||||
return settings.filter(
|
||||
(setting) => !['thinkingBudget', 'temperature', 'topP', 'topK'].includes(setting.key),
|
||||
);
|
||||
}
|
||||
const modelAwareSettings =
|
||||
endpoint === EModelEndpoint.google
|
||||
? settings.map((setting) => {
|
||||
|
|
|
|||
|
|
@ -110,8 +110,11 @@ describe('anthropicSettings', () => {
|
|||
expect(reset('claude-haiku-5')).toBe(64000);
|
||||
});
|
||||
|
||||
it('should return 64K for future versions like claude-haiku-9', () => {
|
||||
expect(reset('claude-haiku-9')).toBe(64000);
|
||||
it('should return 128K for claude-haiku-5-5 and later', () => {
|
||||
expect(reset('claude-haiku-5-5')).toBe(128000);
|
||||
expect(reset('claude-haiku-5.5')).toBe(128000);
|
||||
expect(reset('global.anthropic.claude-haiku-5-5')).toBe(128000);
|
||||
expect(reset('claude-haiku-9')).toBe(128000);
|
||||
});
|
||||
});
|
||||
|
||||
|
|
@ -336,6 +339,11 @@ describe('anthropicSettings', () => {
|
|||
it('should cap at 64K for claude-haiku-4-5 when value exceeds', () => {
|
||||
expect(set(80000, 'claude-haiku-4-5')).toBe(64000);
|
||||
});
|
||||
|
||||
it('should cap at 128K for claude-haiku-5-5 when value exceeds', () => {
|
||||
expect(set(100000, 'claude-haiku-5-5')).toBe(100000);
|
||||
expect(set(200000, 'claude-haiku-5-5')).toBe(128000);
|
||||
});
|
||||
});
|
||||
|
||||
describe('Claude Sonnet 4.6+ models (128K cap)', () => {
|
||||
|
|
|
|||
|
|
@ -647,6 +647,8 @@ const DEFAULT_MAX_OUTPUT = 8192 as const;
|
|||
const LEGACY_ANTHROPIC_MAX_OUTPUT = 4096 as const;
|
||||
const CLAUDE_SONNET_128K_OUTPUT_PATTERN =
|
||||
/claude-sonnet[-.]?(?:4[-.]?(?:[6-9]|\d{2})|[5-9]|\d{2,})(?=$|[^0-9])/;
|
||||
const CLAUDE_HAIKU_128K_OUTPUT_PATTERN =
|
||||
/claude-haiku[-.]?(?:5[-.]?(?:[5-9]|\d{2})|[6-9]|\d{2,})(?=$|[^0-9])/;
|
||||
|
||||
/**
|
||||
* Claude "Mythos-class" model families — new top-level classes (peers of
|
||||
|
|
@ -707,7 +709,10 @@ export const anthropicSettings = {
|
|||
return ANTHROPIC_MAX_OUTPUT;
|
||||
}
|
||||
|
||||
if (CLAUDE_SONNET_128K_OUTPUT_PATTERN.test(modelName)) {
|
||||
if (
|
||||
CLAUDE_SONNET_128K_OUTPUT_PATTERN.test(modelName) ||
|
||||
CLAUDE_HAIKU_128K_OUTPUT_PATTERN.test(modelName)
|
||||
) {
|
||||
return ANTHROPIC_MAX_OUTPUT;
|
||||
}
|
||||
|
||||
|
|
@ -740,7 +745,10 @@ export const anthropicSettings = {
|
|||
return value;
|
||||
}
|
||||
|
||||
if (CLAUDE_SONNET_128K_OUTPUT_PATTERN.test(modelName)) {
|
||||
if (
|
||||
CLAUDE_SONNET_128K_OUTPUT_PATTERN.test(modelName) ||
|
||||
CLAUDE_HAIKU_128K_OUTPUT_PATTERN.test(modelName)
|
||||
) {
|
||||
if (value > ANTHROPIC_MAX_OUTPUT) {
|
||||
return ANTHROPIC_MAX_OUTPUT;
|
||||
}
|
||||
|
|
|
|||
|
|
@ -52,6 +52,7 @@ describe('validateVertexConfig region gating', () => {
|
|||
'claude-opus-4-8',
|
||||
'claude-opus-4-7',
|
||||
'claude-sonnet-5',
|
||||
'claude-haiku-5-5',
|
||||
];
|
||||
const legacy = ['claude-sonnet-4-6', 'claude-3-7-sonnet-20250219', 'claude-3-opus@20240229'];
|
||||
|
||||
|
|
|
|||
|
|
@ -26,6 +26,7 @@ export const defaultVertexModels: string[] = [
|
|||
'claude-opus-4-6',
|
||||
'claude-sonnet-5',
|
||||
'claude-sonnet-4-6',
|
||||
'claude-haiku-5-5',
|
||||
'claude-3-7-sonnet-20250219',
|
||||
'claude-3-5-sonnet-v2@20241022',
|
||||
'claude-3-5-sonnet@20240620',
|
||||
|
|
@ -43,7 +44,7 @@ const MULTI_REGION_LOCATIONS = new Set(['global', 'us', 'eu']);
|
|||
* and return 404 for anything newer.
|
||||
*/
|
||||
const REQUIRES_MULTI_REGION =
|
||||
/^claude-(?:opus-(?:4-[7-9]|[5-9])|sonnet-[5-9]|fable-[5-9]|mythos-[5-9])/;
|
||||
/^claude-(?:opus-(?:4-[7-9]|[5-9])|sonnet-[5-9]|haiku-(?:5-[5-9]|[6-9])|fable-[5-9]|mythos-[5-9])/;
|
||||
|
||||
/**
|
||||
* Whether the configured location can serve models that require `global` or a
|
||||
|
|
|
|||
|
|
@ -3257,6 +3257,60 @@ describe('Opus 5.5 pricing', () => {
|
|||
});
|
||||
});
|
||||
|
||||
describe('Haiku 5.5 pricing', () => {
|
||||
const models = [
|
||||
'claude-haiku-5-5',
|
||||
'claude-haiku-5.5',
|
||||
'anthropic/claude-haiku-5-5',
|
||||
'global.anthropic.claude-haiku-5-5',
|
||||
];
|
||||
|
||||
it.each(models)('prices %s at its standard rates up to the prompt-size threshold', (model) => {
|
||||
const { threshold } = premiumTokenValues['claude-haiku-5-5'];
|
||||
for (const inputTokenCount of [1000, threshold]) {
|
||||
expect(getMultiplier({ model, tokenType: 'prompt', inputTokenCount })).toBe(
|
||||
tokenValues['claude-haiku-5-5'].prompt,
|
||||
);
|
||||
expect(getMultiplier({ model, tokenType: 'completion', inputTokenCount })).toBe(
|
||||
tokenValues['claude-haiku-5-5'].completion,
|
||||
);
|
||||
expect(getCacheMultiplier({ model, cacheType: 'write', inputTokenCount })).toBe(
|
||||
cacheTokenValues['claude-haiku-5-5'].write,
|
||||
);
|
||||
expect(getCacheMultiplier({ model, cacheType: 'read', inputTokenCount })).toBe(
|
||||
cacheTokenValues['claude-haiku-5-5'].read,
|
||||
);
|
||||
}
|
||||
});
|
||||
|
||||
it.each(models)('prices %s at its long-prompt rates above the threshold', (model) => {
|
||||
const inputTokenCount = premiumTokenValues['claude-haiku-5-5'].threshold + 1;
|
||||
expect(getMultiplier({ model, tokenType: 'prompt', inputTokenCount })).toBe(
|
||||
premiumTokenValues['claude-haiku-5-5'].prompt,
|
||||
);
|
||||
expect(getMultiplier({ model, tokenType: 'completion', inputTokenCount })).toBe(
|
||||
premiumTokenValues['claude-haiku-5-5'].completion,
|
||||
);
|
||||
expect(getCacheMultiplier({ model, cacheType: 'write', inputTokenCount })).toBe(
|
||||
premiumCacheTokenValues['claude-haiku-5-5'].write,
|
||||
);
|
||||
expect(getCacheMultiplier({ model, cacheType: 'read', inputTokenCount })).toBe(
|
||||
premiumCacheTokenValues['claude-haiku-5-5'].read,
|
||||
);
|
||||
});
|
||||
|
||||
it('does not fall back to the default rate', () => {
|
||||
expect(getMultiplier({ model: 'claude-haiku-5-5', tokenType: 'prompt' })).not.toBe(defaultRate);
|
||||
});
|
||||
|
||||
it('keeps Haiku 4.5 on its own flat rates', () => {
|
||||
expect(premiumTokenValues['claude-haiku-4-5']).toBeUndefined();
|
||||
expect(getMultiplier({ model: 'claude-haiku-4-5', tokenType: 'prompt' })).toBe(
|
||||
tokenValues['claude-haiku-4-5'].prompt,
|
||||
);
|
||||
});
|
||||
});
|
||||
|
||||
describe('Sonnet 5.5 pricing', () => {
|
||||
it.each([
|
||||
'claude-sonnet-5-5',
|
||||
|
|
|
|||
|
|
@ -181,6 +181,9 @@ export const tokenValues: Record<string, { prompt: number; completion: number }>
|
|||
'claude-3-7-sonnet': { prompt: 3, completion: 15 },
|
||||
'claude-3.7-sonnet': { prompt: 3, completion: 15 },
|
||||
'claude-haiku-4-5': { prompt: 1, completion: 5 },
|
||||
/* Rates up to a 100K prompt; the higher tier lives in `premiumTokenValues`. */
|
||||
'claude-haiku-5-5': { prompt: 0.1, completion: 0.5 },
|
||||
'claude-haiku-5.5': { prompt: 0.1, completion: 0.5 },
|
||||
'claude-opus-4': { prompt: 15, completion: 75 },
|
||||
'claude-opus-4-5': { prompt: 5, completion: 25 },
|
||||
'claude-opus-4-6': { prompt: 5, completion: 25 },
|
||||
|
|
@ -368,6 +371,8 @@ export const cacheTokenValues: Record<string, { write: number; read: number }> =
|
|||
'claude-3-5-haiku': { write: 1, read: 0.08 },
|
||||
'claude-3-haiku': { write: 0.3, read: 0.03 },
|
||||
'claude-haiku-4-5': { write: 1.25, read: 0.1 },
|
||||
'claude-haiku-5-5': { write: 0.125, read: 0.01 },
|
||||
'claude-haiku-5.5': { write: 0.125, read: 0.01 },
|
||||
'claude-sonnet-4': { write: 3.75, read: 0.3 },
|
||||
'claude-sonnet-4-5': { write: 3.75, read: 0.3 },
|
||||
'claude-sonnet-4-6': { write: 3.75, read: 0.3 },
|
||||
|
|
@ -456,6 +461,8 @@ export const premiumTokenValues: Record<
|
|||
string,
|
||||
{ threshold: number; prompt: number; completion: number }
|
||||
> = {
|
||||
'claude-haiku-5-5': { threshold: 100000, prompt: 0.5, completion: 2.5 },
|
||||
'claude-haiku-5.5': { threshold: 100000, prompt: 0.5, completion: 2.5 },
|
||||
'gemini-3.1': { threshold: 200000, prompt: 4, completion: 18 },
|
||||
'gpt-5.4': { threshold: 272000, prompt: 5, completion: 22.5 },
|
||||
'gpt-5.4-pro': { threshold: 272000, prompt: 60, completion: 270 },
|
||||
|
|
@ -486,6 +493,8 @@ export const premiumCacheTokenValues: Record<
|
|||
string,
|
||||
{ threshold: number; write: number; read: number }
|
||||
> = {
|
||||
'claude-haiku-5-5': { threshold: 100000, write: 0.625, read: 0.05 },
|
||||
'claude-haiku-5.5': { threshold: 100000, write: 0.625, read: 0.05 },
|
||||
'gpt-5.4': { threshold: 272000, write: 5, read: 0.5 },
|
||||
'gpt-5.5': { threshold: 272000, write: 10, read: 1 },
|
||||
'gpt-5.6': { threshold: 272000, write: 10, read: 0.8 },
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue