MAM5020F — Generative AI for Research
Current Generative AI Landscape
A practical, research-oriented snapshot (updated through May 2026).
Contents
Key terms
An AI system trained on vast amounts of text that can generate, summarise, translate, and reason about language. LLMs such as Claude, ChatGPT, and Gemini are the core of most modern AI assistants.
Describes an AI model that can process or generate more than one type of data, for example text and images together, or text, audio, and video, rather than working with text alone. Most large models since 2023 are multimodal.
An AI model whose trained parameters are publicly released so anyone can download, run, or modify it, in contrast to closed, proprietary models whose weights stay private and are reachable only through an API.
An architecture where a large model is split into many specialised sub-networks (experts), but each input is routed through only a small subset, keeping computation per query low despite a very large total parameter count.
Track the landscape by model families, modalities, context limits, and agent/tool reliability, rather than by a single leaderboard.
This page prioritizes tools/models with clear availability and/or credible release notes.
Large Language Models (LLMs)
OpenAI (ChatGPT / API)
- GPT-5.5 / GPT-5.5 Pro: flagship family for "work + agents" (May 2026); Instant/Thinking/Pro variants; tops the FrontierMath leaderboard
- GPT-5.5-Codex: GPT-5.5 variant optimized for long-horizon, agentic coding
- GPT-5.2 / GPT-4o: earlier models; still widely used but no longer "default latest"
Practical differentiator: sustained tool-calling + long-context work and repo-scale coding agents.
Anthropic Claude
- Claude Opus 4.7: flagship (generally available 16 April 2026); strong long-context reasoning + coding; current default for production research workflows
- Claude Mythos (preview): limited research preview (May 2026); leads GPQA Diamond and HLE at time of writing
- Claude Sonnet 4.5: high-utility "daily driver" for agents/coding
- Claude Haiku 4.5: fast, cost-efficient small model
Often preferred for instruction-following stability and long-document workflows.
Google Gemini
- Gemini 3.1 Pro / Deep Think: multimodal + Google ecosystem integrations; IMO 2025 gold-medal-level maths reasoning
- NotebookLM: grounded Q&A on your sources; audio/video "overviews" expanding
Strongest "current web + Workspace/Drive" integration path for many research teams.
Meta Llama (Open-weight)
- Llama 4 Scout: compact open-weight model; long-context positioning
- Llama 4 Maverick: larger open-weight model; aimed at stronger reasoning/coding
- Llama 4 Behemoth (preview): very large MoE model previewed by Meta
Use when you need controlled deployment, fine-tuning, or "open-weight first" constraints.
China-focused frontier & fast-moving families
- Kimi K2.6 (Moonshot AI): open-weights MoE (~1T total / 32B active), native multimodal + "agentic" positioning (April 2026); coding near closed-frontier parity
- Qwen 3.x (Alibaba): open-weight family; rapid iteration cadence (watchpoint for 2026)
- DeepSeek V4 Pro / V4 Flash: open-weights MoE (Pro: 1.6T total / 49B active, 1M-token context; MIT licence, April 2026); world-leading competitive-programming score at roughly 10× lower cost
- Doubao 2.0 (ByteDance): "agent era" positioning; very large consumer footprint
- ERNIE 4.5 / X1 (Baidu): multimodal + reasoning variants; rolling into product ecosystem
- GLM-5 (Zhipu), M2.5 (MiniMax), Hunyuan 2.0 (Tencent), Spark X2 (iFlytek): additional major families
China's ecosystem now has parallel "frontier stacks" (models + apps + cloud hosting) with very fast deployment cycles.
Other important families
- Mistral Large 2: strong multilingual model family (widely hosted in enterprise stacks)
- Qwen3: also widely used outside China via open weights + hosted endpoints
- GLM: Zhipu's family continues to iterate quickly (watch open-source drops)
Image Generation Models
General-purpose image models
- DALL·E 3: strong prompt interpretation + editing workflows
- Midjourney (v6+): aesthetic quality and style control
- Stable Diffusion (SD3 / SD3.5): local deployment + custom pipelines
- Adobe Firefly: Creative Cloud integration + commercial positioning
Google image generation (Imagen)
- Imagen 3: rolled into Gemini experiences; also accessible via Gemini API (Feb 2025)
- Imagen 4: improved quality + text rendering; available via Gemini API / AI Studio (Jun 2025)
- Imagen 4 Ultra: higher-end variant for demanding creative/control needs
Practical note: Imagen models are often chosen for prompt adherence + typography improvements in later versions.
FLUX (Black Forest Labs)
- FLUX.2 [dev]: high-quality text-to-image + editing; multi-reference control
- FLUX.2 [klein]: "sub-second" variants aimed at real-time workflows (Jan 2026 updates)
- FLUX.2 [flex]: ongoing performance updates
Notable: explicit focus on controllable editing + multiple references.
Video Generation Models
Frontier text-to-video
- Sora 2 (OpenAI): video + synchronized dialogue/sound effects
- Veo 3 / Veo 3 Fast (Google): native audio; provenance via SynthID watermarking
- Runway: creator-oriented suite; strong tooling for iteration and control
- Pika / Luma: competitive creative tools; often best for rapid ideation
2025→2026 shift: controllability, consistency, and editability matter as much as realism.
Avatar / translation video tools
- Synthesia: training/presentation avatars
- HeyGen: dubbing + lip-sync translation workflows
Code Generation & Agentic Development
Agentic coding assistants
- OpenAI Codex (GPT-5.2-Codex): long-horizon refactors/migrations; agent workflows
- Claude Code: terminal-centric agent for repo-scale work
- GitHub Copilot: mainstream IDE integration
- Cursor: AI-first editor with codebase-wide operations
"Best" depends on tool permissions/sandboxing, repo size, and multi-file edit reliability.
Agent platforms & orchestration (incl. OpenClaw)
- OpenClaw: open-source agent platform: multi-channel integrations + large "skills" ecosystem + model-agnostic config
- Why it matters: abstracts "agents" away from one vendor/model; lets you plug GPT/Claude/Gemini/Llama into the same automation layer
- Security note: self-hosted agent platforms can be risky if exposed/misconfigured; guidance has emphasized access controls + auditing
Treat it like running a web app with plugin execution: sandbox skills, least-privilege permissions, and lock down network exposure.
Key Trends (2025 → early 2026)
Agent reliability beats raw scores: long-running tool use, memory/compaction, and safe sandboxing are the real differentiators.
Parallel "frontiers" emerged: the China stack (models + apps + cloud) is now a fast-moving peer ecosystem rather than a lagging one.
Provenance becomes normal: watermarking/provenance systems (e.g., SynthID for some Google media) increasingly ship by default.
Image models are getting judged on typography: later Imagen family entries explicitly target text rendering improvements.
Source anchors for the updates above include official Google Developers notes on Imagen (Imagen 3/4), Moonshot's Kimi K2.5 release info, and reporting on OpenClaw's rapid adoption + security cautions.
Drafted with Claude (Anthropic) and reviewed by Jonathan Shock before publication. AI-generated errors are possible — if you spot one, please email jonathan.shock@uct.ac.za. Full detail: AI Content Disclaimer.
© 2026 Jonathan Shock · MAM5020F: Generative AI for Research · CC BY 4.0