AI Tool Pricing: Every Plan, Every Gotcha

Claude Pro at $20/month or ChatGPT Plus at $20/month remain the best single-seat value for most knowledge work, writing, analysis, brainstorming, coding assistance. Pick Claude if you do long-document work or want the best writing quality; pick ChatGPT if you need image generation and the Advanced Voice experience. Heavy coders should add Cursor ($20/month) on top. Gemini Advanced at $19.99 makes more sense if you live in Google Workspace. Anyone spending under $20 should use the free tiers of Claude, ChatGPT, and Gemini together before paying for anything.

Four culprits. (1) API usage on top of seat pricing, Cursor Business, Claude Code, and most agentic tools bill usage above the included cap. (2) Per-seat creep as the team adopts the tool, a $20 seat becomes $1,000 when 50 people get it. (3) Enterprise add-ons: SSO, audit logs, DLP, admin console, often only on the top tier. (4) Image and video generation sold as credits that burn fast at meaningful resolution. Always build a 90-day actuals model before signing an annual enterprise contract.

For most users, no. The $200/month tiers (ChatGPT Pro, Claude Max, Gemini Ultra, Grok Heavy) are priced for power users who are already saturating the $20 tier's rate limits and need priority access, the newest models first, and extended-thinking runs. If you hit "you've used your Claude limit" a few times per week, they're worth it. If you don't, the delta is compute you won't consume. Try the $20 tier hard for a month before stepping up.

Cheaper for bursty or low-volume use, more expensive for heavy daily chat use. A heavy chat user burns 10-30M tokens a month; on API at Claude Sonnet 4.6 or GPT-5 rates that runs $30-120+, versus a flat $20 Plus/Pro subscription. For programmatic or batched workloads (summarisation, classification, enrichment), the API is usually 3-10x cheaper than any seat-based product equivalent. The honest rule: interactive chat → subscription; automation → API.

Open-weight models run on OpenRouter, Together, Fireworks, Groq, or DeepInfra give you Llama 4, DeepSeek V3/R1, and Kimi K2 Thinking at 5-20% the cost of proprietary frontier models for most tasks. For a small business, a $50/month OpenRouter balance with careful model routing outperforms $200+ in mixed subscriptions. The trade-off is tool integration, you lose the native ChatGPT/Claude apps and canvas features.

Rough baseline: $30-80 per person per month all-in. That covers one primary LLM seat ($20), a shared agent/coding tool at usage-based pricing (Cursor, Claude Code, Replit Agent), image gen credits if you do marketing work, and a slice of API budget for automations. Expect 2-3x that for engineering-heavy teams. The worst budgeting mistake is signing a per-seat enterprise contract before the team has a 60-day usage baseline, you'll lock in seats you don't need at a tier you didn't have to buy.

Free tiers in 2026 are genuinely useful. Claude's free tier gives you Sonnet 4.x. ChatGPT's free tier gives you GPT-5 with limits. Gemini's free tier includes Gemini 2.5 Flash and some Pro access. Grok's free tier shipped with real capability. For any single-modality task (text chat, basic code questions, simple research), rotating across 3-4 free tiers hits zero marginal cost and covers 80% of personal use. The paid tiers earn their money on rate limits, the newest models, and long-context work.

Image-generation credits are the classic trap, Midjourney, Runway, and Sora credits burn faster than the marketing implies, especially at higher resolutions or with multiple variants. Sales intelligence tools (Clay, Apollo, Instantly-adjacent) have per-export and per-enrichment fees that stack. Enterprise agent platforms charge per action, per tool call, and per seat. Read the unit of billing before the advertised plan price.

Quarterly. The frontier moves fast enough that a $20 tool in April is a free-tier capability in July, and the model you're paying for has a successor you haven't tried. Block a calendar recurring every 90 days. Cancel anything you haven't opened in 30 days, even if it's cheap. Re-benchmark your primary LLM on the three tasks you use it for most. The cost of a bad autorenewal compounds fast across a team.

Yes: (1) Is the price per seat, per usage, per agent action, or per output? (2) Are the newest models behind a higher tier than the one being pitched? (3) Does rate-limit headroom match your actual daily volume? (4) Are SSO, audit logs, and admin controls on this tier or a more expensive one? (5) Are the evaluations the sales team showed you on the tier you'd buy, or the top tier? (6) What's the annual-commit penalty if you want to downgrade mid-term? If any answer is ambiguous, the deal is not ready to sign.