Skip to main content

What's New in AI (2026 Edition)

In short: Since 2024, AI shifted from chat assistants to autonomous agents that complete multi-step tasks on their own. The headline changes: agents and built-in tool use (web, code, files) are now standard, context windows reached 1M+ tokens, skills and MCP let you extend models, and the frontier lineup is Claude (Opus 5.5 and Fable 5.1), OpenAI (GPT-5.6 still running regular Chat, topped by the newer GPT-6 family — Astra, Sol, and Luna), and Gemini (3.1 Pro and 3.8 Flash). Here's what changed — and what to do about it.

For Returning Users

If you learned AI in 2023-2024, welcome back! The landscape has transformed dramatically. This page gets you up to speed quickly.

Snapshot: September 2026

This page captures the state of AI as of early September 2026, with partial updates for releases on September 22, 29, and 30. The agent ecosystem and model lineup move quickly, so expect specifics (versions, prices, tool standings) to drift between updates and this page to need a fuller refresh soon. The current frontier lineup — Claude Opus 5.5 (September 22, superseding the July 24 Opus 5), GPT-5.6 (July 2026, still running ChatGPT's regular Chat), Claude Fable 5.1 (September 1), Gemini 3.8 Flash (September 2), GPT-6 Astra (September 3), GPT-6 Sol / GPT-6 Luna (September 22), GPT-6.1 Sol (September 29, upgrading GPT-6 Sol), and Gemini 4 Argon (September 30, restricted access only) — was verified current as of this update.

The Biggest Shifts​

1. Chat → Agents 🤖​

Then (2024):

  • You: "Write a marketing plan"
  • AI: [Writes text]
  • You: [Copies, pastes, uses manually]

Now (2026):

  • You: "Create and execute a marketing campaign"
  • Agent: [Researches competitors, generates content, creates calendar, schedules posts, monitors performance]
  • You: [Reviews automated work]

The shift: AI now completes entire workflows autonomously.

Tools to try:

2. Tiny Context → Massive Context 📏​

Then:

  • GPT-4 launched at 8K tokens; the 32K and 128K variants that followed still felt tight
  • Claude: 100K was revolutionary
  • Workarounds needed for long documents

Now:

  • Standard: 200K tokens (≈500 pages)
  • Claude Fable 5.1, Opus 5.5, and Sonnet 5.5: 1M tokens (≈750,000 words) at standard pricing
  • Gemini 3.1 Pro/3.8 Flash: 1M+ tokens with multimodal input (Google's new flagship, Gemini 4 Argon, was announced September 30 but is limited to a restricted cyber-defense program for now; Gemini 3.5 Pro was cancelled)
  • You can give AI entire codebases or document libraries

What this means:

  • No more splitting documents
  • AI understands full project context
  • Better, more coherent long-form work

3. Skills & Customization Ecosystem 🧩​

Then:

  • Base model only
  • Some custom GPTs
  • Limited customization

Now:

  • Skills.sh - Marketplace of AI capabilities
  • Custom skills for any use case
  • MCP servers for tool integration
  • Skills work across platforms

Example: Instead of prompting "analyze data like an expert", install a "Data Analysis" skill that knows your preferred methods, tools, and output format.

4. MCP Becomes Standard 🔧​

Model Context Protocol (MCP) is now the universal way to connect AI to tools.

One MCP server → Works with Claude, ChatGPT, Gemini

Popular servers:

  • GitHub (repos, issues, PRs)
  • Slack (messages, search)
  • Postgres (database queries)
  • Filesystem (read/write files)
  • Google Drive (docs, sheets)

Impact: Build integrations once, use everywhere.

Learn more: MCP Introduction

5. Multi-Modal Is Default 🎨​

Then:

  • Text models separate from image models
  • Had to switch modes
  • Limited integration

Now:

  • Text, images, audio, video in one interface
  • No mode switching
  • Seamless multi-modal reasoning

Example: "Analyze this chart [image], compare to this data [file], and create a video explanation [generation]" - all in one conversation.

Major Model Updates​

Claude (Fable 5.1, the Claude 5.5 family — Opus 5.5 and Sonnet 5.5 — plus Haiku 4.5)​

What's new:

  • Fable 5.1 sits at the top of the range, Sonnet 5.5 is the balanced daily driver, and Haiku 4.5 remains the fast, low-cost option in the current lineup
  • Fable 5.1 (released September 1, 2026) succeeded Fable 5, alongside a new Mythos 5.1 controlled-access release — Anthropic now ships both as different safety configurations of the same underlying model rather than as separate products
  • Opus 5 (released July 24, 2026) replaced Opus 4.8 as Anthropic's flagship Opus model — same $5/$25 per-million-token pricing, but closer to Fable 5.1's performance on coding and knowledge-work tasks, plus an adjustable "effort" setting to trade intelligence for speed
  • Opus 5.5 (released September 22, 2026) superseded Opus 5 in turn — roughly 40% lower typical workload cost and roughly 30% faster output, at $4/$20 per million input/output tokens (cache reads $0.20/MTok). Haiku 5.5 is expected "in the coming weeks" but hasn't shipped yet, so Haiku 4.5 is still current.
  • Sonnet 5.5 (released September 28, 2026) superseded Sonnet 5 (June 30, 2026) as the second model in the Claude 5.5 family — 30%+ faster output and up to 30% lower cost per task in Anthropic's testing, at the same $2/$10 per million input/output tokens as Sonnet 5. Sonnet 5 remains available as a legacy model.
  • Computer use (controls browsers, apps)
  • Extended thinking mode
  • 1M-token context window at standard pricing on Fable 5.1, Opus 5.5, and Sonnet 5.5
  • Significantly better coding and agentic task completion
  • Skills and Claude Code now production-grade

Best for: Autonomous agents, complex coding, long-context research

Try it: claude.ai

GPT-6: Astra, Sol, and Luna (Released September 3, 22, and 29, 2026)​

What's new:

  • OpenAI's new flagship reasoning model, GPT-6 Astra, launched September 3, 2026 — its largest training run to date — topping the GPT-5.6 lineup below for coding, computer use, and complex multi-step work
  • Confusingly, Astra goes by two names depending on where you access it: GPT-6 Astra in ChatGPT's Work and Codex modes, and GPT-6 Pro in the regular Chat interface, where it's limited to Pro ($100/$200), Business, and Enterprise plans
  • On September 22, 2026, OpenAI expanded the family with GPT-6 Sol ($2/$10 per million input/output tokens, a permanent ~50% cut from GPT-5.6 Sol's $4/$20) and GPT-6 Luna ($0.10/$0.50 per million tokens, versus $0.20/$1.20 for GPT-5.6 Luna). There's no GPT-6 Terra — this generation is only three tiers.
  • GPT-6 Sol and GPT-6 Luna are live in ChatGPT's Work and Codex modes for Plus/Pro/Business/Enterprise/Edu plans; GPT-6 Luna is also in the desktop app for Free/Go users. Neither is in regular browser Chat yet.
  • One week later, at DevDay on September 29, 2026, OpenAI released GPT-6.1 Sol, an upgrade to GPT-6 Sol at the same $2/$10 price (API: gpt-6.1-sol, 1.05M-token context). OpenAI says it nearly matches GPT-6 Astra on agentic coding and computer use at about a fifth of Astra's token prices. Like GPT-6 Sol, it's live in Work and Codex modes for Plus/Pro/Business/Enterprise/Edu plans but not yet in regular Chat.
  • Plus subscribers ($20/month) still default to GPT-5.6 Sol in ordinary Chat, and Free/Go users to GPT-5.6 Luna — see the GPT-5.6 section below
  • 1.05M-token context window and API access for Astra (as gpt-6-astra), priced at $10/$50 per million input/output tokens (higher above 272K tokens of input)

Best for: Complex agentic and computer-use work on higher-tier ChatGPT plans or via the API

GPT-5.6 (Released July 2026) — predecessor lineup, still running regular Chat​

What's new:

  • OpenAI's previous frontier family, replacing GPT-5.5 (released April 2026), split into three named tiers instead of a single model plus a Pro add-on: Sol (the default in regular ChatGPT Chat for Plus and below, complex reasoning/coding/agentic work), Terra (mid-tier, competitive with GPT-5.5 at roughly half the cost, with no GPT-6 successor), and Luna (optimized for speed, the default for Free/Go in regular Chat)
  • Strong gains in coding efficiency and agentic task completion
  • Available in ChatGPT and the OpenAI API
  • Superseded at the top of OpenAI's overall lineup by GPT-6 Astra, Sol, and Luna (above), though GPT-5.6 Sol and Luna remain the default for most ChatGPT users today because regular browser Chat hasn't gotten GPT-6 Sol/Luna yet

Best for: General use, custom GPTs, agentic coding via Codex

Gemini 3.1 Pro / Gemini 3.8 Flash (Released early 2026 / September 2026)​

What's new:

  • Gemini 3.8 Flash (GA September 2, 2026) replaced Gemini 3.7 Flash — itself only three weeks old — as the default fast-tier model, with further gains in coding and agentic tasks. Per Google's own benchmarks it still outperforms Gemini 3.1 Pro on some of those tasks, a reminder that "Pro" and "Flash" no longer map cleanly to "smarter" and "faster"
  • Gemini 4 Argon (announced September 30, 2026) is Google's first Gemini 4 model and new frontier flagship, built for deep reasoning across long, multi-step workflows such as software engineering and cyber defense. Access starts with governments and trusted cyber defenders in Google's Fairwind Program; Google says it will widen access to developers, enterprises, and consumers later but has given no date. Google also cancelled the long-delayed Gemini 3.5 Pro, so 3.1 Pro stays the top generally available Gemini model for now
  • Major capability jump in reasoning and multimodal understanding over the Gemini 2.x family
  • Deep Workspace integration with native multi-modal input

Best for: Research, massive documents, Google Workspace users

New Tool Categories​

AI-First Code Editors​

Winners:

  1. Cursor - Most popular, VS Code fork with AI
  2. Windsurf → Devin Desktop - Cognition folded its 2025 Windsurf acquisition into an agent cockpit for running Devin and other coding agents side-by-side
  3. VS Code + Copilot - Still relevant but losing ground

What changed: AI isn't a plugin anymore, it's the core interface.

Natural Language to App​

Game changers:

  • v0.app (renamed from v0.dev in January 2026) - Describe UI → Get React components
  • Bolt.new - Full-stack apps from descriptions
  • Replit Agent - Entire deployed apps in minutes

Impact: Non-developers building functional software.

No-Code Agent Builders​

Platform tier:

  • Relevance AI - Build "AI employees"
  • Lindy - Personal AI assistant
  • Salesforce Agentforce - Enterprise-grade no-code agent workflows
  • Zapier Agents - Natural language automation

Impact: Anyone can build custom agents without coding.

Browser Agents​

  • Claude in Chrome - Research and automation
  • Perplexity Comet - AI-native browser for research and multi-step tasks
  • Browserbase - Headless browser for agents

Use case: "Research these 10 companies and create comparison spreadsheet" - agent handles everything.

Deprecated / Less Relevant​

What's fading:

  • ChatGPT Plugins → Replaced by GPT Actions and MCP
  • Jasper/Copy.ai → Base models got so good, specialized tools less needed
  • Basic code completion only tools → Agents do more than autocomplete now
  • Tabnine → Pivoted to enterprise self-hosted; Cursor and Copilot dominate the general dev market

Pricing Changes​

Standard pricing (2026):

  • Claude Pro: $20/month
  • ChatGPT Plus: $20/month
  • Google AI Pro (formerly Gemini Advanced): $20/month
  • Cursor Pro: $20/month
  • Copilot: $10/month (still cheapest entry price, but shifted to usage-based AI Credit billing on June 1, 2026)

Enterprise:

  • Most platforms: $30-40/user/month
  • More usage limits, admin controls, SSO

Free tiers:

  • Still available but limited
  • Claude Free: Basic access to current Sonnet
  • ChatGPT Free: Limited access to current GPT models
  • Gemini Free: Gemini 3.8 Flash with generous daily limits

Skills You Should Learn (2026)​

For Everyone​

  1. Agent delegation - How to give tasks vs ask questions
  2. Skill creation - Extending AI with custom capabilities
  3. Basic MCP - Understanding tool connections

For Technical Users​

  1. LangGraph or CrewAI - Agent orchestration frameworks
  2. MCP server creation - Build custom integrations
  3. Prompt caching - Optimize costs and speed
  4. Agent safety - Sandboxing and guardrails

For Business Users​

  1. Workflow design - Thinking in agent processes
  2. No-code platforms - Relevance AI, Lindy, Respell
  3. ROI measurement - Proving agent value
  4. Team training - Getting team to think agentically

Common Mistakes (From 2024 Thinking)​

❌ Mistake 1: Still Treating AI Like Chat​

Old: "AI, write me an email" → Copy/paste → Send New: "Handle email responses for X type of inquiry" → Agent does it

❌ Mistake 2: Not Using Skills​

Old: Long, detailed prompts every time New: Create a skill once, reuse forever

❌ Mistake 3: Ignoring MCP​

Old: Copying data in and out manually New: Connect AI directly to your tools via MCP

❌ Mistake 4: Single Agent for Everything​

Old: One ChatGPT for all tasks New: Specialized agents for different domains (research agent, coding agent, writing agent)

❌ Mistake 5: Not Thinking in Workflows​

Old: One-off requests New: "What repeating workflow can an agent handle?"

Quick Start Guide (2026)​

Week 1: Foundation​

  • Try Claude (Fable 5.1, Sonnet 5.5, or Opus 5.5), GPT-5.6 or GPT-6, or Gemini 3.1 Pro / 3.8 Flash
  • Give it a complete task (not just a question)
  • Watch it use tools autonomously
  • Browse skills.sh for ideas

Week 2: Skills & Customization​

  • Create your first custom skill
  • Install an MCP server
  • Build a simple agent workflow

Week 3: Development (if technical)​

  • Install Cursor or try v0.app
  • Use Claude Code for a feature
  • Experiment with Bolt.new

Week 4: Deployment​

  • Pick one recurring task
  • Build an agent to handle it
  • Measure time saved

Resources for Catch-Up​

Essential Reading​

  1. AI Agents Overview - Core concepts
  2. 50+ Agent Use Cases - Practical examples
  3. Tools & Platforms 2026 - What to use
  4. Skills & Instructions - Customization

Communities​

Keep Learning​

  • AI changes fast - follow key platforms on social media
  • Join relevant Discord/Slack communities
  • Build something small each month
  • Share what you learn (teaching solidifies knowledge)

What's Coming Next (2026-2027)​

Predictions based on current trajectory:

  1. Agent-to-Agent Communication - Your agents talk to others' agents
  2. Persistent Agents - Agents that work 24/7 in background
  3. Specialized Hardware - Agent-optimized compute
  4. Agent Marketplaces - Buy/sell trained agents
  5. Regulation - Laws specifically for autonomous agents. This one is already moving: in late July 2026, more than 1,100 employees across OpenAI, Anthropic, Google, and Meta signed an open letter urging governments to build an international "pacing mechanism" — a way to verifiably slow deployment if capability outruns oversight. By mid-September, OpenAI, Anthropic, and Google DeepMind confirmed talks on a voluntary industry safety standards body of their own, reportedly targeted for late 2026 or early 2027. Whatever comes of either effort, the debate has moved inside the labs. See the Global Regulatory Landscape for where policy stands today.

The Bottom Line​

2024: AI was a powerful assistant 2026: AI is an autonomous teammate

The question changed from "What should I ask AI?" to "What should I delegate to my agents?"

Next Steps​

Now that you're caught up, here's where to go to start putting 2026's AI to work:

Welcome to 2026. Time to build some agents.

SponsoredYour product hereReach people actively learning AI — flat monthly rates, no ad networks