Dateline: March 26, 2026 | Next update: April 2, 2026
This guide covers four AI model families in equal depth. For each it explains not just the models but the full product ecosystem: tools, apps, integrations, pricing plans, and recent launches.
Claude (Anthropic)
Anthropic was founded in 2021 by former OpenAI researchers including Dario and Daniela Amodei. Their approach, called Constitutional AI, trains models to have internalised values rather than externally imposed rules. Claude is known for careful, nuanced responses and self-awareness about its limitations. As of March 2026 it holds the number-one position on the Chatbot Arena text leaderboard and leads on software engineering benchmarks. Claude Code, its coding agent, has surpassed $2.5 billion in annualised revenue as of March 2026, up from $1 billion in January.
Claude Haiku 4.5 — fast and cheap
The lightest and fastest Claude model. Best for high-volume tasks where speed matters more than depth: text classification, fact extraction, simple Q&A, and routing layers in automated pipelines that escalate harder queries to a stronger model.
Context: 200K tokens. SWE-bench: 73.3%. Batch API: 50% discount. Vision and tool use supported. Token consumption ~3–4x lower than Sonnet for comparable queries. Haiku 3 retires April 19 2026 — migrate to claude-haiku-4-5-20251001 before then.
Best for: High-volume pipelines, classification, simple Q&A
Claude Sonnet 4.6 — the everyday standard
The model most people should use for most tasks. Became the default free model on claude.ai in February 2026. Handles complex analysis, long documents, structured writing, code generation, and multi-step reasoning. Adaptive reasoning automatically calibrates thinking depth to query complexity.
Became the default free model on claude.ai in February 2026. In user preference studies, 70% of Claude Code users preferred it to the previous generation; 59% preferred it over the prior Opus generation for coding.
Context: 200K standard / 1M beta. Output: up to 64K tokens. Adaptive extended thinking replaces manual thinking budget controls. Elo 1,633 (professional writing leaderboard). Prompt caching: ~10% of base input cost on cache hits. Available on AWS Bedrock and Google Vertex AI. Default model for all Cowork sessions.
Best for: Most professional use — writing, analysis, coding, long documents
Claude Opus 4.6 — the flagship
Most powerful Claude model, for work demanding deep and sustained reasoning. Number one on Chatbot Arena (Elo 1504). Leads all published models on software engineering: 80.84% SWE-bench Verified. Task completion horizon: 14.5 hours of autonomous work (50th percentile), the longest published figure for any model. Introduced Agent Teams: native framework for coordinating multiple Claude instances simultaneously.
Agent Teams: native multi-agent orchestration in a single API call. 1M token context window now GA as of March 2026 (unified pricing). Code Security (Enterprise/Team preview): data flow reasoning for vulnerability detection.
Context: 200K standard / 1M GA (March 2026). Max output: 128K tokens. SWE-bench Verified: 80.84%. METR task horizon: 14h30m (50th percentile). Used by Norway sovereign wealth fund for ESG screening (Feb 2026); NASA Mars rover routing (Dec 2025).
Best for: Deep reasoning, long-horizon autonomous tasks, complex agentic workflows
Claude Code — agentic coding tool
A terminal-based coding agent giving Claude access to your codebase, terminal, and files. Writes code, runs tests, reads errors, and fixes bugs with minimal human input. Features include remote access, scheduled recurring tasks, a plugin marketplace (MCP integrations), parallel agents, and auto memory (persistent project architecture knowledge across sessions).
Auto mode (Mar 24): AI safety classifier approves routine developer actions automatically and screens for prompt injection attacks. Claude Code Review: automatic background code reviewer to catch bugs before they reach the codebase. Claude Code Channels (Mar 24): control Claude Code via Discord and Telegram.
SWE-bench terminal agent: ~75%. Remote access: browser/mobile session control. Plugin marketplace: MCP integrations with Team/Enterprise admin controls. Auto mode: safety classifier with prompt injection screening; small token/latency overhead. Revenue: $2.5B+ annualised March 2026, up from $1B in January.
Best for: Software engineering with full agentic control
Claude Cowork — for non-developers
Gives Claude persistent access to your local files in a sandboxed virtual machine. Connects to Gmail, Google Drive, Google Calendar, Slack, and 38+ services. Projects create isolated workspaces per context with their own files and history. Cross-session memory learns your preferences over time.
Computer use (Mar 23): Claude can now control your Mac — clicking, typing, scrolling, opening apps, navigating browsers, filling forms. Connector-first approach uses structured API integrations when available; falls back to keyboard-and-mouse only when no connector exists. Financial/trading/crypto apps blocked by default. macOS research preview; Pro and Max subscribers only.
Dispatch (Mar 17): Persistent conversation thread across phone and Claude Desktop. Assign tasks from phone; Claude executes on your desktop in the background. Scheduled recurring and on-demand tasks: set tasks to run automatically — weekly summaries, daily briefings, recurring reports — without prompting each time.
Sandboxed VM for file access; computer use runs OUTSIDE the VM on the actual macOS desktop. Per-app permission model; blocklist user-configurable. 38+ native MCP connectors. Projects: isolated context per folder. Memory excludes passwords, health, financial data by default. Requires Pro ($20/mo), Max ($100–200/mo), Team, or Enterprise. Computer use: macOS only, research preview. Sonnet 4.6 is default; user can switch to Opus 4.6.
Best for: Knowledge workers who want AI running multi-step tasks across files and apps
Claude in Excel and PowerPoint
Add-in for Microsoft Excel (Opus 4.6, native pivot table editing and conditional formatting) and add-in for PowerPoint (creates, edits, and revises slides). Both share the full open document as context and support custom skills and connectors via LLM gateway integrations with major cloud providers.
Best for: Microsoft Office users, enterprise knowledge workers
Health data on mobile
Claude can read and analyse health and fitness data on iOS and Android (activity patterns, workouts, sleep). Available on Pro and Max plans, currently US-only. Android requires Health Connect and Android 14+.
Plans and Pricing
| Plan | Price | Key Includes |
|---|---|---|
| Free | $0 | Sonnet 4.6, web search, memory, Artifacts, code execution, image analysis |
| Pro | $20/month | 5x usage, full Claude Code, unlimited Projects, Research mode, Cowork (incl. computer use preview) |
| Max | $100–200/month | All Pro features, higher limits, early feature access, priority at peak |
| Team | $25/seat/month (annual) | Min 5 seats. Admin controls, shared Projects, SSO, MS365 + Slack |
| Enterprise | Custom | 500K context, HIPAA readiness, SCIM, audit logs. Self-serve purchase available |
200K rule: input above 200K tokens billed at 2x ($3→$6 input, $15→$22.50 output per MTok). Batch API: 50% discount. Prompt caching: up to 90% off repeated context. Developer math: $200 Max is ~18x cheaper than equivalent direct API token spend for heavy Claude Code use. Claude 5 (codename Fennec) has appeared in Google Vertex AI infra logs — expected mid-2026.
ChatGPT / OpenAI
ChatGPT is OpenAI's general-purpose AI assistant. It is a single place where people ask questions, draft and edit writing, analyse files, search the web, generate images and video, write code, talk by voice, and delegate multi-step work to an agent. The product routes between several models and tools depending on what you ask it to do.
GPT-5.3 Instant — the default everyday model
The fast general-purpose default model for chat, everyday writing, search, and tool use. Supports the widest set of ChatGPT tools and is the baseline experience on Free as well as paid plans.
16 March 2026: update improves follow-up tone and reduces teaser-style phrasing.
Rate limits: Free — 10 messages per 5 hours, then fallback to mini. Plus and Go — 160 messages per 3 hours. Business and Pro described as effectively unlimited. Context: 16K (Free), 32K (Plus/Business), 128K (Pro/Enterprise). Tools: web search, data analysis, image analysis, file analysis, canvas, image generation, memory, custom instructions.
Best for: General chat, fast drafting, routine search, file Q&A, broadest day-to-day tool-compatible use
GPT-5.4 Thinking — the flagship reasoning model
OpenAI's current flagship reasoning model. Slower and more expensive than the default Instant model, but better at multi-step reasoning, coding, document-heavy work, and difficult research questions. The model to pick when quality matters more than speed.
Launched 5 March 2026, replacing GPT-5.2 Thinking as the current flagship reasoning path.
API ID: gpt-5.4. Context: 1M tokens. Max output: 128K tokens. API pricing: Standard $2.50 input / $0.25 cached / $15 output per 1M tokens. Flex: $1.25 / $0.13 / $7.50. Batch: half standard. Priority: $5 / $0.50 / $30. Benchmarks: OSWorld 75.0; SWE-bench Pro 57.7; GPQA Diamond 92.8; ARC-AGI-2 73.3; Harvey BigLaw Bench 91%; Humanity's Last Exam 52.1 with tools.
Best for: Complex writing, multi-step analysis, coding, legal and research work
GPT-5.4 Pro — the highest-quality option
The version for users willing to spend far more for a better answer on especially difficult tasks. Not a general default — the 'use this when the stakes are high and the problem is hard' option. Note: Apps, Memory, Canvas, and image generation are not available with the Pro model in ChatGPT.
Launched 5 March 2026 alongside GPT-5.4 Thinking, with higher benchmark scores on several hard reasoning tests but a narrower ChatGPT tool set.
API ID: gpt-5.4-pro. Context: 1,050,000 tokens. Max output: 128K tokens. API pricing: $30 input / $180 output per 1M tokens standard. For prompts above 272K tokens: 2x input and 1.5x output. Regional endpoints: 10% uplift. Benchmarks: GPQA Diamond 94.4; ARC-AGI-2 83.3; Humanity's Last Exam 58.7 with tools; BrowseComp 89.3. Some API requests can take minutes — background mode recommended.
Best for: Very hard reasoning problems, high-stakes review, cases where cost matters less than quality
GPT-5.4 mini — the smaller reasoning model
A cheaper API model and practical fallback in the ChatGPT user experience. For Free and Go users it appears through the Thinking feature. For most paid users it is the model ChatGPT falls back to when GPT-5.4 Thinking limits are reached.
Launched 17–18 March 2026 as both a user-facing Thinking option for Free/Go and a fallback model for paid users hitting GPT-5.4 Thinking limits.
API ID: gpt-5.4-mini. Context: 400K tokens. API pricing: $0.75 input / $0.075 cached / $4.50 output per 1M tokens. Batch: half standard. Benchmarks: SWE-bench Pro 54.4; Terminal-Bench 60.0; MCP Atlas 57.7.
Best for: Budget-sensitive reasoning, graceful fallback, high-volume API workloads needing a strong reasoning model
GPT-5.4 nano — the API-only lowest-cost model
Not a mainstream ChatGPT model. Exists for developers who need a very cheap, very fast model for automation, classification, extraction, and high-volume backend tasks.
Launched 17 March 2026 as the cheapest current GPT-5.4-class API model.
API ID: gpt-5.4-nano. Context: 400K tokens. API pricing: $0.20 input / $0.02 cached / $1.25 output per 1M tokens. Benchmarks: SWE-bench Pro 52.4; Terminal-Bench 46.3; MCP Atlas 56.1.
Best for: Developer automation, classification, extraction, cost-sensitive inference pipelines
Legacy models — what remains and until when
GPT-5.2 Thinking remains in Legacy Models for Plus and Pro users for 90 days after the 5 March 2026 launch — retiring 5 June 2026. GPT-4o in Custom GPTs: Business, Enterprise, and Edu retain it until 3 April 2026. Already retired from ChatGPT (13 Feb 2026): GPT-4o, GPT-4.1, GPT-4.1 mini, OpenAI o4-mini, and GPT-5 (Instant and Thinking).
Agent Mode — ChatGPT that can act
Instead of just replying with advice, Agent Mode can browse websites, work through forms, inspect files, use tools, and carry out multi-step tasks. Operates through OpenAI's agent environment. When a site needs a login or sensitive confirmation, control is handed back to the user. Typical task duration: 5–30 minutes.
Limits: Plus 40 agent messages/month; Pro 400/month; Business/Enterprise 40/month. Business flexible pricing: 30 credits per message. User remains in control for logins — screenshots not captured while user controls browser.
Best for: Delegating long web tasks, repetitive information gathering, form-heavy work
Canvas — shared editing workspace
A persistent canvas where text or code can be revised, marked up, and iterated on directly. Supports Python code execution, can be enabled inside GPTs, and allows sharing of canvas assets. Most useful when a task has become too large for ordinary chat bubbles.
Best for: Long drafts, iterative editing, structured code review
Sora 2 — video generation
OpenAI's current video generation system. Consumer limits: Plus/Business — unlimited images and video, up to 480p and 10 seconds, 1 concurrent generation. Pro — unlimited, faster, up to 1080p and 20 seconds, up to 5 concurrent, watermark-free.
12 March 2026: Sora API expanded with reusable character references, longer 20-second generations, 1080p output for sora-2-pro, video extensions, and Batch API support.
API IDs: sora-2 and sora-2-pro. API pricing: sora-2 at 720p is $0.10/sec standard. sora-2-pro: $0.30/sec at 720p, $0.50/sec at 1024p, $0.70/sec at 1080p; batch is half standard prices.
Best for: Creative video generation, concept visualization, marketing drafts, storyboarding
Deep Research — long-form, citation-backed research
Creates a research plan, works through sources over time, and returns a structured report with citations. Connected apps are read-only. Users can edit the research plan before the run starts. Legacy Deep Research mode removed 26 March 2026.
19 March 2026: OpenAI added editable research plan and improved report view. Legacy Deep Research mode removed 26 March 2026.
Best for: Policy briefs, market research, literature-style reviews, multi-source synthesis
Memory, Projects, and Library — the persistence layer
Memory lets ChatGPT remember past details. Projects group chats and files together. Library is the new filing cabinet for uploaded and generated files — reusable across chats.
23 March 2026: Library launched for saved files, making uploaded and created files reusable across chats. Currently web-only; recent files in composer and file search on iOS and Android. Available to Plus, Pro, and Business users outside the EEA, Switzerland, and UK.
Best for: Ongoing projects, recurring work, personalized assistance
OpenAI Codex — the coding platform
A dedicated coding environment with multiple parallel agents, isolated worktrees, reviewable diffs, and delegation to cloud agents. Integrations: GitHub, Slack, Linear.
4 March 2026: Expanded on Windows for Business workspaces. 10 March: auto top-up for shared credits used by Codex and Sora. GPT-5.3-Codex remains the specialized coding model launched on 5 February 2026.
Best for: Serious software work, parallel coding tasks, teams wanting agentic help in real development workflows
Plans and Pricing
| Plan | Price | Key Includes |
|---|---|---|
| Free | $0 | Limited GPT-5.3, limited uploads and image generation, limited Deep Research |
| Go | $8/month (US) | More GPT-5.3, more uploads and images, longer memory. May include ads. |
| Plus | $20/month | Advanced reasoning models, projects, tasks, Codex, Sora, expanded Deep Research and Agent Mode |
| Pro | $200/month | GPT-5.4 Pro, maximum Deep Research and Agent Mode, expanded Sora, higher-priority Codex |
| Business (formerly Team) | $25/user/month annual | Shared workspace, admin controls, Codex, agent access. Min 2 users. |
| Enterprise | Custom | Bigger context/file support, SCIM, RBAC, EKM, data residency, SLAs |
API pricing (current GPT-5 family): gpt-5.4 $2.50/$0.25 cached/$15 output per 1M tokens; gpt-5.4-pro $30/$180; gpt-5.4-mini $0.75/$0.075/$4.50; gpt-5.4-nano $0.20/$0.02/$1.25. API billing is separate from ChatGPT subscriptions. Self-serve API usage tiers scale from ~$100/month at low tiers to $200,000/month at Tier 5.
Gemini (Google)
Gemini is Google's primary artificial intelligence system — a single "brain" that can see, hear, read, and create. It is multimodal: it understands text, images, audio, and video at once. It powers everything from a simple chat box on your phone to professional tools used by scientists and developers.
Gemini 3.1 Flash Lite — the efficiency specialist
The newest and most affordable member of the family. Built for high-volume work where you need the AI to do a simple task thousands of times per second without costing a fortune: sorting, translating short messages, labeling photos.
Launched March 3, 2026, as the first "Lite" model in the 3.1 generation, optimized for latency-sensitive production pipelines. 2.5x faster Time to First Token (TTFT) than Gemini 2.5 Flash.
Context Window: 1M tokens (32K via API Preview). API ID: gemini-3.1-flash-lite-preview. API Pricing: $0.25 per 1M input / $1.50 per 1M output. Benchmarks: GPQA Diamond: 86.9% | MMMU-Pro: 76.8% | MMMLU: 88.9%.
Best for: High-volume data processing, latency-sensitive production pipelines
Gemini 3.1 Flash — the balanced workhorse
The default version of Gemini for most people. Strikes the best balance between being clever and being quick. Handles vacation planning, PDF summaries, email drafting, and real-time conversation.
Context Window: 1M tokens. API Pricing: $0.50 per 1M input / $3.00 per 1M output. API ID: gemini-3.1-flash. Benchmarks: ARC-AGI-2: 45.2% | MMLU-Pro: 82.1%.
Best for: General-purpose chat, speed-sensitive everyday tasks
Gemini 3.1 Pro — the flagship
Google's most capable model for reasoning. Designed for hard problems: debugging complex code, synthesizing conflicting data, creating intricate live dashboards or animated graphics. Features "Extended Thinking" — pauses to reason before answering difficult logic puzzles.
Context Window: 1M to 2M tokens. Output Window: 65,536 tokens. API Pricing: $2.00 per 1M input / $12.00 per 1M output (up to 200k context); rates double above 200k. Benchmarks: ARC-AGI-2: 77.1% (SOTA) | GPQA Diamond: 94.3% | Terminal-Bench 2.0: 68.5%.
Best for: Complex logic and coding, scientific research, multi-source synthesis
Gemini 3.1 Ultra — the powerhouse
The top-tier model, exclusive to the most expensive plan. Used for the most demanding creative tasks: generating professional 4K video with Veo 3.1, running advanced Deep Research reports that take up to 20 minutes to compile.
Currently holds #1 on Artificial Analysis Intelligence Index (Score: 61). API access restricted — primarily via Vertex AI for enterprise and through the Gemini App for Ultra subscribers.
Best for: State-of-the-art research, 4K video generation, highest-stakes creative work
NotebookLM — personalized research assistant
Upload your own documents (PDFs, transcripts, websites), and NotebookLM creates a private world where the AI only answers based on your specific information. Audio Overviews: two AI "hosts" discuss your documents like a podcast. Study Suite: quizzes, flashcards, and study guides from your sources.
Cinematic Video Overviews (March 15, 2026): launched for Ultra tier users to create 4K visual summaries. Mind Maps and Artifacts: automatically generates visual diagrams of how your ideas connect.
Source Limits: up to 50 sources per notebook; 500k words per source. Export: supports .pptx and .docx artifact creation.
Best for: Research and study, private knowledge-base Q&A
Jules — the AI coding agent
An asynchronous coding partner. You give Jules a big task, close your laptop, and Jules works in the background — cloning your code, testing its own work, and sending you a pull request when it's finished.
Base Model: Gemini 2.5 Pro / 3.1 Pro. Limits: Free (15 tasks/day), Pro (100 tasks/day), Ultra (300 tasks/day).
Best for: Autonomous coding, background task execution
Veo 3.1 and Lyria 3
- Veo 3.1: Google's 4K video generation model. Creates 8-second cinematic clips with synchronized audio.
- Lyria 3: High-fidelity music generation. Creates 3-minute songs with vocals and lyrics from a text prompt.
Lyria 3 (March 25, 2026): updated to support 3-minute professional arrangements with realistic vocals.
Plans and Pricing
| Plan | Price | Key Includes | Not Included |
|---|---|---|---|
| Basic (Free) | $0.00 | Gemini 3.1 Flash, 32k Context | Deep Research, Jules, 4K Video |
| AI Plus | $7.99/mo | 200GB Storage, 200 Credits | 3.1 Pro Model, High Jules limits |
| AI Pro | $19.99/mo | 2TB Storage, Gemini 3.1 Pro | Veo 3.1 Standard (4K) |
| AI Ultra | $249.99/mo | 30TB Storage, Gemini 3.1 Ultra | — |
Microsoft Copilot
Microsoft Copilot is not a single AI model. It is a family of AI-powered features built on top of models from other companies (primarily OpenAI GPT-5 and Anthropic Claude). What makes Copilot distinctive is its Work IQ, its integration with Microsoft Graph, and its ability to live inside Office apps like Word, Excel, and Teams. The underlying models provide raw reasoning and language capabilities, while Copilot adds organizational context, compliance, and productivity integration.
Copilot Free
Basic chat and writing assistance inside Microsoft Edge and Office web apps. Limited to general-purpose queries and does not access organizational data.
Uses GPT-4.5. No Microsoft Graph integration. No compliance frameworks. No admin controls.
Best for: Individual users wanting lightweight AI help without enterprise integration
Copilot Pro
Unlocks advanced writing, summarization, and design features in Word, Excel, PowerPoint, and Outlook. Requires a separate Microsoft 365 subscription for full app integration.
Access to GPT-5 and Claude 3. Limited Graph integration. No enterprise compliance. Admin controls not available.
Best for: Power users and freelancers who need deeper AI integration in Office apps
Copilot Business
Integrates with Microsoft Graph, enabling context-aware assistance across Teams, SharePoint, and OneDrive. Includes meeting transcription, AI notes, and live translation.
Graph integration: SharePoint, OneDrive, Teams. Compliance: ISO 27001, SOC 2. Admin portal with role-based controls.
Best for: Small and medium businesses needing collaboration and compliance features
Copilot Enterprise
Adds Semantic Index, Purview governance, compliance APIs, SCIM provisioning, and vertical add-ons like Copilot for Sales and Service.
Full Graph integration. Semantic Index for organizational knowledge. Purview compliance. SCIM support. API access for enterprise developers.
Best for: Large organizations requiring compliance, governance, and extensibility
E7 Tier
Bundles Microsoft 365 E7 with Copilot Enterprise. Includes advanced security, compliance, and analytics.
Best for: Enterprises standardizing on Microsoft's highest security and compliance tier
Agent 365
A lightweight agent framework for workflow automation. Enables Copilot to act across apps and services without full Microsoft 365 integration. Claude 3-based agent.
Best for: Teams needing automation without full enterprise licensing
Copilot Cowork
A collaborative AI agent built with Anthropic. Emphasizes Work IQ — the ability to track tasks, checkpoints, and plan-to-action loops.
March 18, 2026: Cowork added inbox and calendar awareness in Copilot Chat.
Claude Cowork architecture. Graph integration optional. Compliance: SOC 2, ISO 27001. Enterprise Cowork adds governance APIs.
Best for: Teams experimenting with collaborative AI workflows
GitHub Copilot
Assists with code completion, multi-file reasoning, and documentation. Enterprise plans add custom knowledge bases and fine-tuned models.
Multi-file agent. Enterprise supports private model fine-tuning. Compliance: SOC 2, ISO 27001.
Best for: Developers and enterprises needing AI-powered coding assistance
Copilot inside Microsoft 365 Apps
- Word: Drafting, summarization, rewriting
- Excel: Data analysis, formula generation, chart creation
- PowerPoint: Slide generation, design suggestions
- Outlook: Email drafting, summarization
- Teams: Meeting transcription, AI notes, live translation
- SharePoint: Content summarization, metadata tagging
- OneDrive: File summarization, search assistance
Excel: predictive trend analysis added.
Best for: Microsoft 365 users wanting AI embedded directly in productivity apps
Plans and Pricing
| Tier | Price | Requirements | Models |
|---|---|---|---|
| Copilot Free | $0 | Microsoft account | GPT-4.5 |
| Copilot Pro | $20/month | Microsoft 365 subscription | GPT-5, Claude 3 |
| Copilot Business | $30/user/month | 5 seats minimum | GPT-5, Claude 3 |
| Copilot Enterprise | $30/user/month | E3/E5 licence | GPT-5, Claude 3 |
| E7 Tier | $57/user/month | Enterprise licence | GPT-5 |
| Agent 365 | $15/user/month | None | Claude 3 |
| GitHub Copilot Pro | $10/month | GitHub account | GPT-4.5 |
| GitHub Copilot Enterprise | $39/user/month | Enterprise licence | GPT-5 |
This document is kept current as of March 2026. For the most up-to-date figures, please request a re-run of this guide monthly.