ChatGPT alternatives

Perplexity

A search engine that answers the question instead of listing pages that might. Every claim carries a numbered citation you can follow, which makes it useful for research you intend to defend later. Focus modes narrow a search to academic papers, forums or video. The trade-off is the usual one for answer engines: you read a synthesis rather than the sources, and the synthesis is only as good as what it retrieved.

Claude

Anthropic's assistant, reached for most often when the input is long or the reasoning has to be careful, contracts, codebases, research papers, anything where skimming produces a confident wrong answer. Projects keep a set of documents in context across conversations, and Artifacts render code and documents beside the chat rather than inside it. Also available as an API and a desktop app.

Pinecone

Managed vector search, so retrieval stops being an operational problem. You upsert embeddings with metadata and query by similarity with filters, and the indexing, sharding and scaling are handled. Serverless indexes bill on usage rather than provisioned pods. The decision is nearly always this against pgvector: a managed service that scales, or one less system to run.

Suno

Generates complete songs from a text prompt (vocals, instrumentation, structure and a title), rather than loops or stems. Useful for demos, background music and figuring out whether an idea works before booking anyone. You can supply your own lyrics or let it write them. Commercial rights depend on the plan, and that detail matters more than the output quality for most professional use.

ElevenLabs

Speech synthesis good enough for published audio rather than prototypes: audiobooks, dubbing, narration and voice-overs that hold up to close listening. It clones a voice from a short sample, speaks across dozens of languages, and exposes the whole thing as an API. Voice cloning is exactly as sensitive as it sounds, which is why consent verification sits in front of it.

Hugging Face

The hub open-source machine learning is organised around. Hundreds of thousands of models and datasets with versioning, model cards and licences, plus Spaces for hosting a live demo of any of them. The `transformers` library is the default way to load a model in Python. If a paper released weights, this is almost always where they are.

Runway

Video generation and editing aimed at people who already edit video. Text and image prompts produce clips, and the surrounding tools (background removal, motion tracking, inpainting, camera control) are the parts that get used daily once the novelty of generation wears off. Widely used in advertising and pre-visualisation, where an approximation of a shot is worth more than a description of it.

LangChain

Open source, unlike ChatGPT

Ollama

Open source, unlike ChatGPT

Replicate

Runs open-source models behind an HTTP API so you never provision a GPU. Push a model with Cog and it becomes an endpoint; pull someone else's and it is one call. Billing is per second of compute, which suits bursty work and gets expensive under sustained load, at which point self-hosting is the comparison to make.

Midjourney

Image generation with a distinctive aesthetic, lighting, composition and texture that other models approximate but rarely match. Prompts are refined through variations and upscales rather than tuned parameters, which suits art direction more than precise control. It runs in Discord and on the web, and the learning curve is mostly about learning what its house style does well.

Glide

Generates mobile and web apps from a table your team already maintains. Rows become screens, columns become fields, and the result is installable and works offline. Best for internal tools (inventory, inspections, field data, directories), where the alternative is a spreadsheet somebody edits on a phone.