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 WrongExampleOutdated flagship modelsSaying "GPT-4o" when GPT-5.4 is currentDeprecated model IDsUsing claude-3-5-sonnet-20241022 (deprecated Jan 2026)Wrong pricingQuoting old rates that changed months agoPhantom modelsReferencing "GPT-4-turbo" or "Gemini Ultra" (deprecated/renamed)Wrong API formatUsing Authorization: Bearer for Anthropic (should be x-api-key)Stale deprecation statusNot knowing DALL-E 3 is shutting down

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

Current Flagship Models#

ProviderFlagshipAPI IDInput $/MTokOutput $/MTokReleasedOpenAIGPT-5.4gpt-5.4$2.50$15.002026-03-17OpenAIGPT-5.4 Minigpt-5.4-mini$0.75$4.502026-03-17OpenAIGPT-5.4 Nanogpt-5.4-nano$0.20$1.252026-03-17OpenAIo3 (reasoning)o3$2.00$8.002025-04AnthropicClaude Opus 4.6claude-opus-4-6$5.00$25.002026-02-05AnthropicClaude Sonnet 4.6claude-sonnet-4-6$3.00$15.002026-02-17AnthropicClaude Haiku 4.5claude-haiku-4-5-20251001$1.00$5.002025-10GoogleGemini 3.1 Progemini-3-1-pro-latest$2.00$12.002026-02-19GoogleGemini 3 Flashgemini-3-flash-latest$0.50$3.002026-03GoogleGemini 2.5 Flash-Litegemini-2.5-flash-lite$0.10$0.402025MetaLlama 4 Maverickmeta-llama/llama-4-maverickvariesvaries2026-04-05MistralSmall 4mistral-small-latest$0.15$0.602026-03-16MistralLarge 3mistral-large-latest$2.00$6.002025-12DeepSeekV3deepseek-chat$0.32$0.892025DeepSeekR1 (reasoning)deepseek-reasoner——2025CohereCommand A Reasoningcommand-a-reasoning~$6.25—2026

Recently Deprecated (Do NOT Use)#

ModelProviderDeprecatedReplacementGPT-4oOpenAI2026-02-13GPT-5.4GPT-4.1OpenAI2026-02-16GPT-5.4DALL-E 2 & 3OpenAI2026-05-12gpt-image-1.5Claude 3.5 SonnetAnthropic2026-01-05Sonnet 4.6Claude Haiku 3.5Anthropic2026-02-19Haiku 4.5Claude Haiku 3Anthropic2026-04-20Haiku 4.5

Upcoming / Announced (NOT Yet Available)#

ModelProviderExpectedNotesGPT-5.5 "Spud"OpenAI2026 Q2Pretraining done 2026-03-24.Claude MythosAnthropicPreview onlyInvitation-only. Cybersecurity.DeepSeek V4DeepSeek20261T params, 1M context.Grok 5xAI2026 Q26T params.Muse SparkMeta2026-04Meta'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:

RankModelProviderInput $/MTokBest For1Gemini 2.5 Flash-LiteGoogle$0.10Cheapest multimodal2Mistral Small 4Mistral$0.15Cheap + reasoning + vision3GPT-5.4 NanoOpenAI$0.20Classification, extraction4DeepSeek V3DeepSeek$0.32Coding, long context5Gemini 3 FlashGoogle$0.50Balanced Google option6GPT-5.4 MiniOpenAI$0.75OpenAI balanced7Claude Haiku 4.5Anthropic$1.00Fast Anthropic option8Gemini 2.5 ProGoogle$1.25Advanced Google9Gemini 3.1 ProGoogle$2.00Frontier reasoning10Mistral Large 3Mistral$2.00675B MoE11o3OpenAI$2.00Complex reasoning12GPT-5.4OpenAI$2.50OpenAI flagship13Claude Sonnet 4.6Anthropic$3.00Anthropic balanced14Claude Opus 4.6Anthropic$5.00Most capable coding/agents

Architecture Quick Facts#

ArchitectureModels Using ItWhy 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 TransformerGPT-5.4, Claude Opus 4.6, Gemini 3.1 ProAll params active. Higher per-token cost but potentially more consistent.

Common Discount Mechanisms#

MechanismDiscountAvailable OnPrompt Caching75-90% on cached inputOpenAI, Anthropic, GoogleBatch API50% on all tokensOpenAI, Anthropic, GoogleCommitted UseVariesEnterprise 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.