chore(cognitive): update AI model catalog

[github-actions]

Model Update Summary

Updated 27 models across 7 providers (anthropic, google-ai, groq, cerebras, xai, openrouter, fireworks-ai). OpenAI docs were inaccessible (HTTP 403); no OpenAI changes made.

---

Anthropic

New Models Added

Model ID	Change	Details
claude-opus-4-6	ADDED	New production flagship model. 1M context, 128k output, $5/$25
claude-sonnet-4-6	ADDED	New production model. 1M context, 64k output, $3/$15. Updated defaultModel to this.
claude-opus-4-5-20251101	ADDED	Production (legacy). 200k context, 64k output, $5/$25
claude-opus-4-1-20250805	ADDED	Production (legacy). 200k context, 32k output, $15/$75
claude-opus-4-20250514	ADDED	Production (legacy). 200k context, 32k output, $15/$75

Lifecycle Changes

Model ID	Change	Old → New
claude-3-haiku-20240307	DEPRECATED	lifecycle: 'production' → lifecycle: 'deprecated', added deprecationDate: '2026-04-19', replacementModels: ['claude-haiku-4-5-20251001']

Config Changes

• defaultModel: claude-sonnet-4-5-20250929 → claude-sonnet-4-6

---

Google AI

New Models Added

Model ID	Change	Details
gemini-3.1-pro	ADDED	New preview model (internalModelId: gemini-3.1-pro-preview). 1M context, $2/$12
gemini-3.1-flash-lite	ADDED	New preview model (internalModelId: gemini-3.1-flash-lite-preview). 1M context, $0.25/$1.50
gemini-2.5-flash-lite	ADDED	New production model. 1M context, $0.10/$0.40

Lifecycle Changes

Model ID	Change	Old → New
gemini-3-pro	DISCONTINUED	lifecycle: 'preview' → lifecycle: 'discontinued', discontinuedDate: '2026-03-09' (shut down upstream). Added replacementModels: ['gemini-3.1-pro']
gemini-2.0-flash	DEPRECATED	lifecycle: 'production' → lifecycle: 'deprecated'. Added deprecationDate: '2026-01-01', replacementModels: ['gemini-2.5-flash']

---

Groq

New Models Added

Model ID	Change	Details
llama-4-scout-17b-16e-instruct	ADDED	New preview model (internalModelId: meta-llama/llama-4-scout-17b-16e-instruct). 131k context, 8k output, $0.11/$0.34
qwen3-32b	ADDED	New preview model (internalModelId: qwen/qwen3-32b). 131k context, 40k output, $0.29/$0.59

Pricing Changes

Model ID	Field	Old → New
gpt-oss-20b	inputCostPer1mTokens	0.1 → 0.075
gpt-oss-20b	outputCostPer1mTokens	0.5 → 0.3
gpt-oss-120b	outputCostPer1mTokens	0.75 → 0.6

Context Window / Limit Changes

Model ID	Field	Old → New
gpt-oss-20b	maxInputTokens	131_000 → 131_072
gpt-oss-20b	maxOutputTokens	32_000 → 65_536
gpt-oss-120b	maxInputTokens	131_000 → 131_072
gpt-oss-120b	maxOutputTokens	32_000 → 65_536
llama-3.3-70b-versatile	maxInputTokens	128_000 → 131_072
llama-3.1-8b-instant	maxInputTokens	128_000 → 131_072
llama-3.1-8b-instant	maxOutputTokens	8192 → 131_072

---

Cerebras

New Models Added

Model ID	Change	Details
qwen-3-235b-a22b-instruct-2507	ADDED	New preview model. 131k context, $0.60/$1.20
zai-glm-4.7	ADDED	New preview model. 131k context, $2.25/$2.75

---

xAI

New Models Added

Model ID	Change	Details
grok-4.20-beta-0309-reasoning	ADDED	New preview model. 2M context, $2/$6
grok-4.20-beta-0309-non-reasoning	ADDED	New preview model. 2M context, $2/$6
grok-4-1-fast-reasoning	ADDED	New production model. 2M context, $0.20/$0.50
grok-4-1-fast-non-reasoning	ADDED	New production model. 2M context, $0.20/$0.50

---

OpenRouter

Pricing Changes

Model ID	Field	Old → New
gpt-oss-120b	outputCostPer1mTokens	0.75 → 0.6

---

Fireworks AI

New Models Added

Model ID	Change	Details
deepseek-v3p2	ADDED	New production model (internalModelId: accounts/fireworks/models/deepseek-v3p2). 163k context, $0.56/$1.68
deepseek-v3p1	ADDED	New production model (internalModelId: accounts/fireworks/models/deepseek-v3p1). 163k context, $0.56/$1.68
kimi-k2-instruct-0905	ADDED	New production model (internalModelId: accounts/fireworks/models/kimi-k2-instruct-0905). 262k context, $0.60/$2.50

---

OpenAI

No changes made. The official OpenAI documentation page (platform.openai.com/docs/models) returned HTTP 403 Forbidden errors and could not be fetched programmatically. The existing config already includes models up to GPT-5.2 (December 2025). Manual verification recommended.