fix(think): extend 16K output headroom to OpenAI reasoning models (#3916)

maxOutputTokensFor() gave the 16000-token output budget only to
thinking-by-default Anthropic Claude 5 models; every other model kept the
4000 default. OpenAI reasoning models (the gpt-5 family and the numbered
o-series) bill internal reasoning tokens as output tokens counting against
max_tokens, so the 4000 cap starved think's actual answer the same way.

Add OPENAI_REASONING_MODEL_RE alongside the existing Claude-5 regex,
deliberately scoped to the gpt-5 family + numbered o-series with the
non-reasoning ChatGPT -chat snapshots (gpt-5-chat-latest,
gpt-5.2-chat-latest) excluded. Defaults (4000/16000) are unchanged and no
config knob is added; gpt-4o, other providers, and provider-prefixed
spellings outside openai:/openai/ keep 4000.

Tests pin the new 16000 cases (colon + slash forms), the chat-snapshot and
codex-mini-latest exclusions, and name/version boundaries (gpt-50, o3foo,
bare model names, openrouter-nested ids).

Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

src/core/think/index.ts

@@ -171,8 +171,19 @@ const DEFAULT_MAX_OUTPUT_TOKENS = 4000;
 // keeps 4000.
 const THINKING_DEFAULT_MAX_OUTPUT_TOKENS = 16000;
 const THINKING_BY_DEFAULT_MODEL_RE = /^anthropic[:/]claude-[a-z0-9]+-5(?:[.-]|$)/i;
+// OpenAI reasoning models spend output budget on internal reasoning tokens
+// the same way — reasoning tokens are billed as output and count against
+// `max_tokens` — so they get the same headroom. Deliberately scoped to the
+// gpt-5 family and the numbered o-series only; anything else (gpt-4o, the
+// non-reasoning `*-chat` snapshots like gpt-5-chat-latest, other providers'
+// reasoning models routed through their own recipes) keeps the conservative
+// 4000 default.
+const OPENAI_REASONING_MODEL_RE = /^openai[:/](?:gpt-5|o[0-9]+)(?:[.-]|$)/i;
+const OPENAI_CHAT_SNAPSHOT_RE = /-chat(?:-|$)/i; // gpt-5-chat-latest, gpt-5.2-chat-latest
 export function maxOutputTokensFor(modelStr: string): number {
-  return THINKING_BY_DEFAULT_MODEL_RE.test(modelStr)
+  const openaiReasoning =
+    OPENAI_REASONING_MODEL_RE.test(modelStr) && !OPENAI_CHAT_SNAPSHOT_RE.test(modelStr);
+  return THINKING_BY_DEFAULT_MODEL_RE.test(modelStr) || openaiReasoning
     ? THINKING_DEFAULT_MAX_OUTPUT_TOKENS
     : DEFAULT_MAX_OUTPUT_TOKENS;
 }

test/think-max-output-tokens.test.ts

@@ -1,9 +1,10 @@
 /**
  * Pins `maxOutputTokensFor` — the per-model output-token budget `runThink`
  * passes to `client.create`. Thinking-by-default Claude 5 models
- * (`anthropic:claude-*-5`) spend a large share of the budget on internal
- * reasoning before emitting an answer, so the 4000 default left `think` with
- * empty/truncated text. They now get 16000; everything else stays 4000.
+ * (`anthropic:claude-*-5`) and OpenAI reasoning models (gpt-5 family,
+ * o-series) spend a large share of the budget on internal reasoning before
+ * emitting an answer, so the 4000 default left `think` with empty/truncated
+ * text. They get 16000; everything else stays 4000.
  */
 import { describe, test, expect } from 'bun:test';
 import { maxOutputTokensFor } from '../src/core/think/index.ts';
@@ -17,12 +18,42 @@ describe('maxOutputTokensFor — thinking-default headroom', () => {
     expect(maxOutputTokensFor('anthropic/claude-sonnet-5')).toBe(16000); // slash form
   });
 
-  test('non-Claude-5 and non-Anthropic keep 4000', () => {
+  test('OpenAI reasoning models (gpt-5 family, o-series) get 16000', () => {
+    expect(maxOutputTokensFor('openai:gpt-5')).toBe(16000);
+    expect(maxOutputTokensFor('openai:gpt-5.2')).toBe(16000);
+    expect(maxOutputTokensFor('openai:gpt-5.5')).toBe(16000);
+    expect(maxOutputTokensFor('openai:gpt-5-mini')).toBe(16000);
+    expect(maxOutputTokensFor('openai:o1')).toBe(16000);
+    expect(maxOutputTokensFor('openai:o3')).toBe(16000);
+    expect(maxOutputTokensFor('openai:o4-mini')).toBe(16000);
+    expect(maxOutputTokensFor('openai/gpt-5.2')).toBe(16000); // slash form
+  });
+
+  test('non-Claude-5 and non-reasoning models keep 4000', () => {
     expect(maxOutputTokensFor('anthropic:claude-opus-4-8')).toBe(4000);
     expect(maxOutputTokensFor('anthropic:claude-haiku-4-5')).toBe(4000);
     expect(maxOutputTokensFor('anthropic:claude-sonnet-4-6')).toBe(4000);
     expect(maxOutputTokensFor('anthropic:claude-3-haiku')).toBe(4000);
     expect(maxOutputTokensFor('openai:gpt-4o')).toBe(4000);
+    expect(maxOutputTokensFor('openai:gpt-4o-mini')).toBe(4000);
+    expect(maxOutputTokensFor('openai:gpt-4.1')).toBe(4000);
+    // Non-reasoning ChatGPT snapshots of the gpt-5 family stay at 4000.
+    expect(maxOutputTokensFor('openai:gpt-5-chat-latest')).toBe(4000);
+    expect(maxOutputTokensFor('openai:gpt-5.2-chat-latest')).toBe(4000);
+    // Scope is the gpt-5 family + numbered o-series only — other OpenAI
+    // reasoning-capable ids (e.g. codex-mini-latest) keep the conservative
+    // default until deliberately added.
+    expect(maxOutputTokensFor('openai:codex-mini-latest')).toBe(4000);
+    // Version/name boundaries: `gpt-50` and `o3foo` are not gpt-5 / o3.
+    expect(maxOutputTokensFor('openai:gpt-50')).toBe(4000);
+    expect(maxOutputTokensFor('openai:o3foo')).toBe(4000);
+    // Other providers' reasoning models are out of scope here — the routed
+    // provider recipe, not this budget, is what changes for them.
     expect(maxOutputTokensFor('deepseek:deepseek-reasoner')).toBe(4000);
+    // Prefix must be the openai provider — a bare model name or another
+    // provider's gpt-5 spelling doesn't match.
+    expect(maxOutputTokensFor('o3')).toBe(4000);
+    expect(maxOutputTokensFor('gpt-5.2')).toBe(4000);
+    expect(maxOutputTokensFor('openrouter:openai/gpt-5.2')).toBe(4000);
   });
 });