AI Resource
AI Model Reference Guide
A regularly updated reference covering the most relevant large language models from the major AI providers. Use this guide to quickly compare capabilities, pricing, and appropriate use cases when evaluating AI tools for your work.
Last updated: March 24, 2026
Provider Models
AI Models
Data sourced from official provider documentation and OpenRouter.ai. Pricing reflects list rates as of the last update date. Always verify with the provider before making purchasing decisions.
Weekly Brief
Latest in AI
Key announcements, releases, and shifts — updated every Monday.
Last Updated: March 24, 2026
Mar 22, 2026
Industry
OpenAI surpasses $25 billion in annualized revenue and is reportedly taking early steps toward a public listing, potentially as soon as late 2026. Rival Anthropic is approaching $19 billion in annualized revenue.
Mar 20, 2026
NVIDIA
NVIDIA GTC 2026 centers on agentic AI as the primary driver of next-generation infrastructure. Jensen Huang previews Kyber — a rack-scale architecture integrating 144 GPUs — planned for the Vera Rubin Ultra system in 2027.
Mar 19, 2026
Anthropic
Anthropic publishes findings from a large-scale study of 80,508 global participants examining public hopes and concerns about AI — one of the largest AI sentiment surveys conducted to date.
Mar 13, 2026
OpenAI
OpenAI strikes a deal to deploy its models in classified U.S. military contexts, breaking from the ethical red lines held by Anthropic. The announcement drives a 295% surge in ChatGPT uninstalls in a single day and pushes Claude to #1 on the App Store.
Mar 10, 2026
Anthropic
Stanford computer scientist Donald Knuth publishes Claude's Cycles, calling Claude Opus 4.6's solution to a complex open graph theory problem "a dramatic advance in automatic deduction and creative problem solving."
Mar 5, 2026
OpenAI
OpenAI releases GPT-5.4 — the first general-purpose model to beat human performance on desktop navigation tasks. Features native computer use, a 1M token context window (via API/Codex), and Tool Search, which reduces token costs by 47% in agentic workflows.
Quick Reference
Compare by Need
Top picks across common use-case categories.
| Largest context window | Gemini 3.1 Pro (2M extended) > Claude Opus 4.6 & GPT-5.4 (1M) > Llama 4 Scout (328K via API; 10M trained) |
| Best for coding / SWE | Claude Opus 4.6 or Sonnet 4.6 · GPT-5.4 · o3 / o4-mini (reasoning-heavy code) |
| Best overall value (frontier) | Gemini 3.1 Pro ($2/$12) · Claude Sonnet 4.6 ($3/$15) · GPT-5.4 ($2.50/$15) |
| Best budget / high-volume | Llama 4 Scout ($0.08/$0.30) · GPT-5.4 nano ($0.20/$1.25) · Gemini 3.1 Flash Lite ($0.25/$1.50) · Claude Haiku 4.5 ($1/$5) |
| Best multimodal (text + image + video + audio) | Gemini 3.1 Pro (only model natively supporting all four) · GPT-5.4 (text + image + computer use) |
| Best deep reasoning / math / science | OpenAI o3 · Gemini 3.1 Pro (ARC-AGI-2 leader at 77.1%) · o4-mini (budget reasoning) |
| Best safety & instruction following | Claude Opus 4.6 · Claude Sonnet 4.6 (Constitutional AI; strongest instruction adherence) |
| Best open-source / self-hostable | Llama 4 Maverick (frontier-class) · Llama 4 Scout (budget) · Phi-4 / Phi-4-mini (edge & on-device) |
| Best for M365 / enterprise ecosystem | Microsoft 365 Copilot (deep M365 integration with Microsoft Graph grounding) |
| Best computer use / desktop agents | GPT-5.4 (first to beat human on desktop tasks) · Claude Sonnet 4.6 (94% accuracy on insurance benchmarks) |
Sources
References
- [*]All API token pricing sourced from OpenRouter.ai model catalog as of March 23, 2026. OpenRouter passes through provider pricing without markup; actual costs billed through OpenRouter include a small purchase fee (~5.5% on credit top-ups). Verify pricing directly with providers: Anthropic, OpenAI, Google AI, Azure AI Foundry.
- [1]Phi-4-mini and Phi-4-multimodal are listed as free (zero cost per token) via OpenRouter with standard rate limits (20 req/min, 200 req/day). Both are open-weight (MIT license) and available for self-hosting. Verify current availability at openrouter.ai/microsoft.
- [2]Microsoft 365 Copilot pricing shown is the standalone add-on rate. Bundled M365 + Copilot plans may be lower (from ~$22/user/month). Pricing changes are effective July 1, 2026. Verify at microsoft.com.
- [3]Gemini 3.1 Pro, 3.1 Flash, and 3.1 Flash Lite are in Preview status. Pricing may change before GA; more restrictive rate limits apply. Gemini 3 Pro (predecessor) was deprecated March 9, 2026.
- [4]GPT-5.4’s 1M token context requires explicit opt-in in the API/Codex. Standard context is 272K tokens. Requests exceeding 272K are billed at double the input rate ($5.00/MTok). Not yet available in the ChatGPT interface.
- [5]Llama 4 Scout is trained on a 10M token context; most hosted providers expose a 328K window due to infrastructure constraints. Self-hosted deployments may expose the full context. Meta does not operate a hosted inference API.
- [6]Token Cost Bracket: LOW (<$1.00/MTok input), MED ($1.00–$5.00/MTok), HIGH (>$5.00/MTok). SEAT = subscription/per-seat pricing, not per-token.
- [7]Retire dates for current-generation models are TBD. GPT-5.2 Thinking is confirmed to retire June 5, 2026.