AI Common Sense

Use when user mentions model names, versions, pricing, API IDs, "which model should I use", "what's the latest model", "model comparison", "API pricing", "what models are available", "哪個模型最新", "模型比較", "API 定價", or when generating code that references specific AI model IDs.

AI Common Sense: Stop Hallucinating Model Names

LLMs frequently hallucinate model names, versions, pricing, and API identifiers because their training data has a cutoff date. This skill provides a verified quick reference and teaches AI agents how to self-verify when the reference may be stale.

Why This Exists

What LLMs Commonly Get Wrong	Example
Outdated flagship models	Saying "GPT-4o" when GPT-5.4 is current
Deprecated model IDs	Using claude-3-5-sonnet-20241022 (deprecated Jan 2026)
Wrong pricing	Quoting old rates that changed months ago
Phantom models	Referencing "GPT-4-turbo" or "Gemini Ultra" (deprecated/renamed)
Wrong API format	Using Authorization: Bearer for Anthropic (should be x-api-key)
Stale deprecation status	Not knowing DALL-E 3 is shutting down

Quick Reference (Last verified: 2026-04-12)

Current Flagship Models

Provider	Flagship	API ID	Input $/MTok	Output $/MTok	Released
OpenAI	GPT-5.4	gpt-5.4	$2.50	$15.00	2026-03-17
OpenAI	GPT-5.4 Mini	gpt-5.4-mini	$0.75	$4.50	2026-03-17
OpenAI	GPT-5.4 Nano	gpt-5.4-nano	$0.20	$1.25	2026-03-17
OpenAI	o3 (reasoning)	o3	$2.00	$8.00	2025-04
Anthropic	Claude Opus 4.6	claude-opus-4-6	$5.00	$25.00	2026-02-05
Anthropic	Claude Sonnet 4.6	claude-sonnet-4-6	$3.00	$15.00	2026-02-17
Anthropic	Claude Haiku 4.5	claude-haiku-4-5-20251001	$1.00	$5.00	2025-10
Google	Gemini 3.1 Pro	gemini-3-1-pro-latest	$2.00	$12.00	2026-02-19
Google	Gemini 3 Flash	gemini-3-flash-latest	$0.50	$3.00	2026-03
Google	Gemini 2.5 Flash-Lite	gemini-2.5-flash-lite	$0.10	$0.40	2025
Meta	Llama 4 Maverick	meta-llama/llama-4-maverick	varies	varies	2026-04-05
Mistral	Small 4	mistral-small-latest	$0.15	$0.60	2026-03-16
Mistral	Large 3	mistral-large-latest	$2.00	$6.00	2025-12
DeepSeek	V3	deepseek-chat	$0.32	$0.89	2025
DeepSeek	R1 (reasoning)	deepseek-reasoner	—	—	2025
Cohere	Command A Reasoning	command-a-reasoning	~$6.25	—	2026

Recently Deprecated (Do NOT Use)

Model	Provider	Deprecated	Replacement
GPT-4o	OpenAI	2026-02-13	GPT-5.4
GPT-4.1	OpenAI	2026-02-16	GPT-5.4
DALL-E 2 & 3	OpenAI	2026-05-12	gpt-image-1.5
Claude 3.5 Sonnet	Anthropic	2026-01-05	Sonnet 4.6
Claude Haiku 3.5	Anthropic	2026-02-19	Haiku 4.5
Claude Haiku 3	Anthropic	2026-04-20	Haiku 4.5

Upcoming / Announced (NOT Yet Available)

Model	Provider	Expected	Notes
GPT-5.5 "Spud"	OpenAI	2026 Q2	Pretraining done 2026-03-24.
Claude Mythos	Anthropic	Preview only	Invitation-only. Cybersecurity.
DeepSeek V4	DeepSeek	2026	~1T params, 1M context.
Grok 5	xAI	2026 Q2	~6T params.
Muse Spark	Meta	2026-04	Meta's first proprietary/closed-source model.

How to Use This Reference

For AI Agents

1. Before generating code with model IDs: Check this reference for the correct API ID.
2. Before quoting pricing: Check this reference. Pricing changes frequently.
3. Staleness check: If the "Last verified" date is older than 30 days, use WebSearch to verify before trusting these numbers.
4. When uncertain: Say "based on my last verified reference from [date]" rather than stating as current fact.

Staleness Detection

text

IF (today - last_verified_date) > 30 days:
  → WebSearch "[provider] models pricing [current year]"
  → Cross-reference with this table
  → Note any discrepancies
  → Flag to user: "My reference is [N] days old, I verified via web search"

When NOT to Trust This Reference

• Model pricing can change without notice
• New models may launch between updates
• Deprecation dates may shift
• "Upcoming" models may be delayed or cancelled

Verification Commands

When you need to verify current model information, use these tools:

Web search queries (use WebSearch tool):

• OpenAI models: site:platform.openai.com models
• Anthropic models: site:docs.anthropic.com models
• Google Gemini: site:ai.google.dev models
• Pricing (any provider): [provider] API pricing [current year]
• Specific model ID: "[exact-model-id]" API
• Deprecation status: [provider] model deprecation [current year]

SDK Version Check

bash

# OpenAI
npm info openai version
pip show openai | grep Version

# Anthropic
npm info @anthropic-ai/sdk version
pip show anthropic | grep Version

# Google
npm info @google/generative-ai version
pip show google-generativeai | grep Version

Cost Comparison (Budget → Premium)

Sorted by input cost per million tokens:

Rank	Model	Provider	Input $/MTok	Best For
1	Gemini 2.5 Flash-Lite	Google	$0.10	Cheapest multimodal
2	Mistral Small 4	Mistral	$0.15	Cheap + reasoning + vision
3	GPT-5.4 Nano	OpenAI	$0.20	Classification, extraction
4	DeepSeek V3	DeepSeek	$0.32	Coding, long context
5	Gemini 3 Flash	Google	$0.50	Balanced Google option
6	GPT-5.4 Mini	OpenAI	$0.75	OpenAI balanced
7	Claude Haiku 4.5	Anthropic	$1.00	Fast Anthropic option
8	Gemini 2.5 Pro	Google	$1.25	Advanced Google
9	Gemini 3.1 Pro	Google	$2.00	Frontier reasoning
10	Mistral Large 3	Mistral	$2.00	675B MoE
11	o3	OpenAI	$2.00	Complex reasoning
12	GPT-5.4	OpenAI	$2.50	OpenAI flagship
13	Claude Sonnet 4.6	Anthropic	$3.00	Anthropic balanced
14	Claude Opus 4.6	Anthropic	$5.00	Most capable coding/agents

Architecture Quick Facts

Architecture	Models Using It	Why It Matters
MoE (Mixture of Experts)	Mistral Large 3 (675B/41B), DeepSeek V3 (671B/37B), Llama 4 Maverick (17B/128 experts)	Massive total params but only a fraction active per token → cheaper inference.
Dense Transformer	GPT-5.4, Claude Opus 4.6, Gemini 3.1 Pro	All params active. Higher per-token cost but potentially more consistent.

Common Discount Mechanisms

Mechanism	Discount	Available On
Prompt Caching	75-90% on cached input	OpenAI, Anthropic, Google
Batch API	50% on all tokens	OpenAI, Anthropic, Google
Committed Use	Varies	Enterprise agreements

Per-Provider Deep Dives

For detailed model specs, deprecation timelines, cross-platform IDs, and API quick-start examples, see the references/ directory in the GitHub repo:

• references/openai.md — Full OpenAI model catalog + audio/image models
• references/anthropic.md — Cross-platform IDs (Bedrock, Vertex) + cache pricing
• references/google.md — Gemini 3.x + 2.5 + specialized models
• references/meta.md — Llama 4 MoE details + access methods
• references/mistral.md — Full specialist model catalog (Devstral, Voxtral, OCR)
• references/deepseek.md — V3 MoE details + V4 roadmap
• references/xai.md — Grok versions + corporate context
• references/cohere.md — Command A + open-source models (Transcribe, Tiny Aya)

How to Update This Reference

This reference gets stale. Here's how to help:

1. Found an error? Open an Issue on GitHub with the correction and source URL.
2. New model released? Submit a PR updating the relevant references/*.md file.
3. Pricing changed? Submit a PR with the new price and a link to the official pricing page.

Every update must include:

• The source URL (official docs preferred)
• The date you verified the information
• What changed and why

Tips

• The more confident an LLM sounds about a model name, the more likely it's hallucinating from training data.
• "I'm not sure which model is current — let me check" is always better than a confident wrong answer.
• Model IDs are exact strings. gpt-5.4 works; GPT-5.4 or gpt5.4 may not.
• Always test API calls with the actual model ID before deploying.