ALL AI
News, money, models, tools, and what actually works — in one place.
Auto-updated Saturday, September 26, 2026 • 4:01 PM EDT — news and markets refresh with every site update
The public companies whose numbers move with the AI buildout. Live, refreshed with the site.
Not investment advice. See The Economy, Stupid for the wider market picture.
Pulled automatically from tech and wire feeds, filtered for AI relevance.
| Product | Free tier | Paid (individual) | Best for |
|---|---|---|---|
| ChatGPT | Yes, limited | ~$20/mo Plus | General use, image generation, wide plugin ecosystem |
| Claude | Yes, limited | ~$20/mo Pro | Long documents, writing, coding, analysis |
| Google Gemini | Yes | ~$20/mo tier | Google Workspace integration, long context |
| Microsoft Copilot | Yes | ~$20/mo | Office documents, Windows integration |
| Perplexity | Yes | ~$20/mo Pro | Search with citations, research |
| GitHub Copilot | Limited | ~$10–19/mo | In-editor code completion |
| Ollama (local) | Free, open | $0 — your hardware | Privacy, offline use, no per-token cost |
If you write or analyse long documents, try Claude first. If you want the broadest general-purpose tool, ChatGPT. If you live in Google or Microsoft documents, the bundled option usually wins on convenience alone. If you need sourced answers, Perplexity. If the data cannot leave your building, run a local model — that is a genuine option now, not a compromise.
Benchmark tables go stale within weeks and every vendor cherry-picks. Rather than publish numbers that will be wrong by next month, here are the trackers that update continuously and test independently:
Updated weekly. Previous weeks stay available below — nothing gets quietly deleted.
Models weight early tokens heavily. 'Answer in under 100 words, no bullet points, cite a source' placed before your question beats the same instructions tacked on the end.
For anything analytical, 'work through this step by step, then give me a one-line verdict' reliably outperforms asking for the verdict alone — you also get to check where it went wrong.
'You are an expert tax accountant' helps because it selects for vocabulary and conventions. 'You are a genius' does nothing measurable. Skip the flattery, keep the domain.
Models hallucinate far less when the source material is in the context window. Pasting a 2,000-word document beats summarizing it yourself and asking about your summary.
Ask one model to produce, another to critique. Cross-checking catches confident errors that a single model will defend when challenged.
Routing simple classification, extraction, and formatting to a small fast model and reserving the frontier model for genuine reasoning typically cuts spend substantially at the same output quality.
If you send the same long system prompt or document repeatedly, providers that support caching charge a fraction for the cached portion. Restructure so the stable content comes first.
Most providers discount asynchronous batch processing sharply. Overnight jobs — summarization, tagging, enrichment — rarely need real-time latency.
Output tokens usually cost several times input tokens. Asking for 'three sentences' rather than letting it run is a direct cost control, not just a style preference.
Models fabricate plausible-looking references and links with real confidence. Every citation needs clicking. This is the single most common way AI-assisted work embarrasses people publicly.
Individual calculations are usually fine; chains of them drift. For anything financial, have it show the numbers and verify with a calculator or spreadsheet.
Models are least reliable about the months just before and after their training cutoff, where partial information produces confident half-truths. Use a model with live search for anything current.
Biographical detail, quotes, and career specifics are frequently wrong and carry real reputational and legal risk if published. Verify against a primary source every time.