harlan-zw · Oct 10, 2026 · Oct 10, 2026 · Oct 10, 2026 · Oct 10, 2026
diff --git a/layers/identity/server/tasks/send-digests.ts b/layers/identity/server/tasks/send-digests.ts
@@ -17,7 +17,7 @@ import { renderDigest } from '../utils/digest-template'
 import { sendEmailWithEnv, signUnsubToken } from '../utils/email'
 
 const CRON = '0 9 1 * *'
-// Kill switch: the per-repo digest sentence runs `anthropic/claude-haiku-4.5`,
+// Kill switch: the per-repo digest sentence runs `anthropic/claude-sonnet-5.5`,
 // a partner model brokered through the Workers AI binding, so every call bills
 // against AI Gateway credits. Those credits are exhausted, so each monthly run
 // returned `provider_failure: 2021: Insufficient AI Gateway credits`, delivered

diff --git a/layers/identity/server/utils/digest-summary.ts b/layers/identity/server/utils/digest-summary.ts
@@ -15,15 +15,15 @@ export interface AiBinding {
     messages: Array<{ role: 'user' | 'assistant', content: string }>
     max_tokens: number
     system?: string
-    temperature?: number
+    thinking?: { type: 'between_tools' }
   }) => Promise<unknown>
 }
 
-// Anthropic Haiku 4.5 brokered through Workers AI. The binding handles auth;
+// Anthropic Sonnet 5.5 brokered through Workers AI. The binding handles auth;
 // no Anthropic key or gateway token needed. Swap to a llama/qwen if cost or
 // latency need to drop later — caller doesn't care which model produced
 // `response`.
-const MODEL = 'anthropic/claude-haiku-4.5'
+const MODEL = 'anthropic/claude-sonnet-5.5'
 
 export interface SubscriptionContext {
   owner: string
@@ -133,6 +133,7 @@ Output JSON only, no prose, with this exact shape:
     system: systemPrompt,
     messages: [{ role: 'user', content: userPrompt }],
     max_tokens: 1024,
+    thinking: { type: 'between_tools' },
   }).then(
     value => ({ _tag: 'response' as const, value }),
     error => ({

diff --git a/layers/registry/server/jobs/generate-faqs.ts b/layers/registry/server/jobs/generate-faqs.ts
@@ -73,7 +73,7 @@ export async function generateFaqs(ctx: FaqContext, skill: FaqSkill): Promise<Fa
   if (items.length < 3)
     return null
 
-  const payload: FaqPayload = { items: items.slice(0, 6), model: 'claude-haiku-4-5-20251001' }
+  const payload: FaqPayload = { items: items.slice(0, 6), model: 'claude-haiku-5-5' }
 
   await putGenerated(ctx.db, {
     owner: skill.owner,

diff --git a/layers/registry/server/jobs/generate-summary.ts b/layers/registry/server/jobs/generate-summary.ts
@@ -87,7 +87,7 @@ export async function generateSummary(ctx: SummaryContext, skill: SummarySkill):
     tagline: tagline.slice(0, 140),
     blurb: blurb.slice(0, 400),
     useCases,
-    model: 'claude-haiku-4-5-20251001',
+    model: 'claude-haiku-5-5',
   }
 
   await putGenerated(ctx.db, {

diff --git a/layers/registry/server/jobs/generate-tags.ts b/layers/registry/server/jobs/generate-tags.ts
@@ -64,7 +64,7 @@ export async function generateTags(ctx: TagContext, skill: TagSkill): Promise<Ta
   if (!tags.length)
     return null
 
-  const payload: TagPayload = { tags, model: 'claude-haiku-4-5-20251001' }
+  const payload: TagPayload = { tags, model: 'claude-haiku-5-5' }
 
   await putGenerated(ctx.db, {
     owner: skill.owner,

diff --git a/layers/registry/server/tasks/ai-generate-poll.ts b/layers/registry/server/tasks/ai-generate-poll.ts
@@ -39,9 +39,9 @@ interface BatchResultLine {
   }
 }
 
-// Haiku 4.5 batch pricing per million tokens (50% discount off realtime).
-const HAIKU_BATCH_INPUT_USD_PER_MTOK = 0.50
-const HAIKU_BATCH_OUTPUT_USD_PER_MTOK = 2.50
+// Haiku 5.5 batch pricing per million tokens for prompts under 100k tokens.
+const HAIKU_BATCH_INPUT_USD_PER_MTOK = 0.05
+const HAIKU_BATCH_OUTPUT_USD_PER_MTOK = 0.25
 
 interface BatchStatusResponse {
   id: string

diff --git a/layers/registry/server/tasks/ai-generate-submit.ts b/layers/registry/server/tasks/ai-generate-submit.ts
@@ -38,7 +38,7 @@ const BATCH_LIMIT = 50
 // ceiling observed in production; wall time goes 7s → ~18s, still well
 // under the scheduled-handler budget.
 const AI_CONCURRENCY = 3
-const HAIKU_MODEL = 'claude-haiku-4-5-20251001'
+const HAIKU_MODEL = 'claude-haiku-5-5'
 // Kill switch: pause Anthropic Haiku batch spend while we investigate the
 // scaled-content-abuse deindexing (10k AI pages → 0.6% indexed, sitewide
 // demotion late Apr 2026). Embeddings/abstractness (Workers AI, ~free) keep
@@ -265,6 +265,7 @@ async function runSubmit(db: D1Database, ai: AiBinding | undefined, vectorize: V
           params: {
             model: HAIKU_MODEL,
             max_tokens: kind === 'faq' ? 1024 : kind === 'tags' ? 256 : 512,
+            thinking: { type: 'disabled' },
             system: [{
               type: 'text',
               text: SHARED_SYSTEM_PROMPT,

diff --git a/layers/registry/server/utils/ai-prompts.ts b/layers/registry/server/utils/ai-prompts.ts
@@ -3,7 +3,7 @@
  * Anthropic's prompt caching (5-min ephemeral TTL) materially cuts cost when
  * a batch submits dozens of skills in one request.
  *
- * Three derivation kinds run through Haiku 4.5 Batch API:
+ * Three derivation kinds run through Haiku 5.5 Batch API:
  *  - `summary`   plain text, 2-3 sentences
  *  - `tags`      JSON array of 5-10 lowercase tags
  *  - `faq`       JSON array of 3-5 { question, answer } objects

diff --git a/shared/server/anthropic.ts b/shared/server/anthropic.ts
@@ -16,7 +16,7 @@ import { spawn } from 'node:child_process'
 const ANTHROPIC_URL = 'https://api.anthropic.com/v1/messages'
 const ANTHROPIC_VERSION = '2023-06-01'
 
-export type HaikuModel = 'claude-haiku-4-5-20251001'
+export type HaikuModel = 'claude-haiku-5-5'
 
 export interface ClaudeUsage {
   input_tokens: number
@@ -61,7 +61,7 @@ export async function callHaikuApi(opts: CallHaikuOpts & { apiKey: string }): Pr
     systemPrompt,
     userPrompt,
     apiKey,
-    model = 'claude-haiku-4-5-20251001',
+    model = 'claude-haiku-5-5',
     maxTokens = 1024,
     cacheSystem = true,
   } = opts
@@ -80,6 +80,7 @@ export async function callHaikuApi(opts: CallHaikuOpts & { apiKey: string }): Pr
     body: JSON.stringify({
       model,
       max_tokens: maxTokens,
+      thinking: { type: 'disabled' },
       system,
       messages: [{ role: 'user', content: userPrompt }],
     }),
@@ -105,7 +106,7 @@ export async function callHaikuApi(opts: CallHaikuOpts & { apiKey: string }): Pr
  * so we can read structured usage info. Returns the same shape as the API call.
  */
 export async function callHaikuCli(opts: CallHaikuOpts): Promise<ClaudeTextResponse> {
-  const { systemPrompt, userPrompt, model = 'claude-haiku-4-5-20251001' } = opts
+  const { systemPrompt, userPrompt, model = 'claude-haiku-5-5' } = opts
 
   // CC wraps/injects its own framing around --system-prompt, so Haiku tends
   // to stay conversational. For batch jobs we fold the system prompt into the

diff --git a/workers/skill-harness/src/gateway.ts b/workers/skill-harness/src/gateway.ts
@@ -90,9 +90,10 @@ export async function forwardSandboxRequest(
     if (!await options.consumeModelCall())
       return Response.json({ code: 'MODEL_CALL_LIMIT' }, { status: 403 })
     const standardInput = { ...input }
-    delete standardInput.speed
+    for (const field of ['speed', 'temperature', 'top_p', 'top_k'])
+      delete standardInput[field]
     const thinking = input.thinking && typeof input.thinking === 'object' && 'type' in input.thinking && input.thinking.type === 'enabled'
-      ? { type: 'enabled', budget_tokens: 2048 }
+      ? { type: 'adaptive' }
       : input.thinking
     return fetchClient('https://api.anthropic.com/v1/messages', {
       method: 'POST',

diff --git a/workers/skill-harness/test/gateway.test.ts b/workers/skill-harness/test/gateway.test.ts
@@ -132,18 +132,19 @@ describe('sandbox credential gateway', () => {
   })
 
   it('routes Anthropic with Worker credentials and enforced model limits', async () => {
-    const options = { ...fixture(), provider: 'anthropic' as const, model: 'claude-sonnet-4-6' }
+    const options = { ...fixture(), provider: 'anthropic' as const, model: 'claude-sonnet-5-5' }
     const response = await forwardSandboxRequest(new Request('https://api.anthropic.com/v1/messages', {
       method: 'POST',
       headers: { 'x-api-key': 'container-key', 'anthropic-beta': 'untrusted-beta' },
-      body: JSON.stringify({ model: 'expensive-model', max_tokens: 32000, service_tier: 'priority', speed: 'fast', messages: [], thinking: { type: 'enabled', budget_tokens: 16000 } }),
+      body: JSON.stringify({ model: 'expensive-model', max_tokens: 32000, service_tier: 'priority', speed: 'fast', temperature: 0.3, top_p: 0.9, top_k: 20, messages: [], thinking: { type: 'enabled', budget_tokens: 16000 } }),
     }), options)
     expect(response.status).toBe(200)
     const [, forwarded] = options.fetch.mock.calls[0]!
     expect(new Headers(forwarded?.headers).get('x-api-key')).toBe('worker-only-secret')
     expect(new Headers(forwarded?.headers).get('anthropic-beta')).toBeNull()
-    expect(JSON.parse(String(forwarded?.body))).not.toHaveProperty('speed')
-    expect(JSON.parse(String(forwarded?.body))).toMatchObject({ model: 'claude-sonnet-4-6', max_tokens: 8192, service_tier: 'standard_only', thinking: { type: 'enabled', budget_tokens: 2048 } })
+    for (const field of ['speed', 'temperature', 'top_p', 'top_k'])
+      expect(JSON.parse(String(forwarded?.body))).not.toHaveProperty(field)
+    expect(JSON.parse(String(forwarded?.body))).toMatchObject({ model: 'claude-sonnet-5-5', max_tokens: 8192, service_tier: 'standard_only', thinking: { type: 'adaptive' } })
   })
 
   it.each([{ tools: [{ type: 'web_search_20250305' }] }, { mcp_servers: [{ url: 'https://external.example' }] }])('blocks Anthropic hosted execution %j', async (body) => {