LIVE
US Military Avoids Close Call After AI-Generated Intelligence Hallucination18/09/26|Study finds gender bias in GPT models is not reduced but reshaped across generations17/09/26 · OpenAI|Architectural tweaks may break conventional scaling law exponents16/09/26|OpenAI publishes a framework for reporting model misalignment16/09/26 · OpenAI|NVIDIA's Vera Rubin NVL72 Debuts in MLPerf Inference v6.116/09/26 · NVIDIA|OpenAI expands ChatGPT advertising with Sponsored Agents16/09/26 · OpenAI|OpenAI moves into advertising with 'Sponsored Agents'16/09/26 · OpenAI|Google DeepMind Introduces Gemini 3.8 Live and Its Extended Thinking Variant15/09/26 · Google DeepMind|What's at stake in AI's trillion-dollar infrastructure bet15/09/26|A Flaw in Chain-of-Thought Safety Monitoring14/09/26|Stellar Colosseum: A Multi-Agent System for Long-Horizon Mathematical Research14/09/26|Apple Code Hints Siri Could Be Swapped for ChatGPT or Claude14/09/26 · Apple|US Military Avoids Close Call After AI-Generated Intelligence Hallucination18/09/26|Study finds gender bias in GPT models is not reduced but reshaped across generations17/09/26 · OpenAI|Architectural tweaks may break conventional scaling law exponents16/09/26|OpenAI publishes a framework for reporting model misalignment16/09/26 · OpenAI|NVIDIA's Vera Rubin NVL72 Debuts in MLPerf Inference v6.116/09/26 · NVIDIA|OpenAI expands ChatGPT advertising with Sponsored Agents16/09/26 · OpenAI|OpenAI moves into advertising with 'Sponsored Agents'16/09/26 · OpenAI|Google DeepMind Introduces Gemini 3.8 Live and Its Extended Thinking Variant15/09/26 · Google DeepMind|What's at stake in AI's trillion-dollar infrastructure bet15/09/26|A Flaw in Chain-of-Thought Safety Monitoring14/09/26|Stellar Colosseum: A Multi-Agent System for Long-Horizon Mathematical Research14/09/26|Apple Code Hints Siri Could Be Swapped for ChatGPT or Claude14/09/26 · Apple|
Beginner💰

Pay or not for an AI?

Comparison of free and paid versions, what really changes, and when the subscription (20-30€/month) becomes profitable. With real cost calculations.

10 min readPublished May 3, 2026· Updated September 17, 2026

In one sentence

For 90% of users, free versions are more than sufficient (Claude.ai, ChatGPT, Gemini are all free with limits). But if you use AI every day for work, a paid subscription at £20-30/month pays for itself quickly. And if you integrate AI into a product, the API is a different story altogether.

💰
The 30-second summary
- Occasional use: free version is enough, no need to pay - Daily professional use: £20-30/month for Claude Pro, ChatGPT Plus or Gemini Advanced, profitable from the 2nd hour saved per month - Product integration: API is pay-per-token, can range from £5 to £5,000/month depending on volume - Self-hosting for free: possible with Llama 4, DeepSeek V3, Mistral Small, but hidden costs (GPU, electricity, expertise)

The 3 ways to pay (or not)

There are 3 different economic models for using AI:

The 3 ways to use AI

 👤Individual🏢Enterprise / Dev
Consumer subscription (Pro/Plus)£20-30/month for unlimited usageNot suitable for large scale
API per tokenToo technical for 99%Standard for product integration
Self-hosting (open-source)Complex, GPU requiredSovereign, full control, but expertise needed

Option 1: Free versions

All the leaders have a free version. They're very powerful but with limits.

Free version limits in 2026

Mistral Le Chat (Free)100 msg/day
Gemini Free (3 Flash)50 msg/day
Claude Free (Sonnet)15 msg/day
ChatGPT Free (GPT-5 mini)10 msg/3h
DeepSeek Chat (Free)0 unlimited

⚠️ Warning: free versions often have a less powerful model. You use GPT-5 mini instead of full GPT-5, Claude Sonnet instead of Opus, etc.

Tip: combine multiple free services
If you use AI occasionally, you can combine 3-4 free services to never hit the limits. Claude for writing, ChatGPT for code, Gemini for web searches, Mistral for French. You'll have the best of the best without paying a penny.

When free is more than enough

  • You use AI occasionally (several times a week, not per day)
  • You mostly do simple tasks (summaries, emails, research, translations)
  • You don't need advanced features (large files, long documents, bulk image generation, agents that take actions)

When free becomes frustrating

  • You hit message limits several times a week
  • You need files/images regularly (limited or absent in free)
  • You want the latest model available as soon as it's released
  • You do long prompts (free model often has shorter context)

Option 2: Consumer subscriptions

£20 to 30/month to have the most powerful model without limits. This is the standard offering from all leaders.

Consumer subscription prices (£/month)

ChatGPT Plus23£
Gemini Advanced22£
Perplexity Pro20£
Claude Pro18£
Mistral Le Chat Pro15£

What you unlock with a subscription

  • The best model (GPT-5 vs GPT-5 mini, Opus vs Sonnet)
  • No (or very few) usage limits
  • Longer context (can analyse large documents)
  • Unlimited image generation (on certain plans)
  • Early access to new features
  • Voice mode, agents, custom chatbot creation

The economic calculation

When is a Pro subscription profitable?
If you use AI to save time at work, do this simple calculation: - Hourly cost (day rate/8): if you charge £400/day → £50/h - Subscription: £20-30/month - Therefore: profitable as soon as AI saves you 30-40 minutes per month For 99% of knowledge workers, it's profitable with just 1 well-used prompt.

2026 subscription comparison

Claude Pro vs ChatGPT Plus

 🟣Claude Pro (£18/month)🟢ChatGPT Plus (£23/month)
Model includedClaude Opus 4.7 + Sonnet 4.6GPT-5 + GPT-5 thinking
French qualityExcellentVery good
Image generationNo (integrated soon)DALL-E 3 unlimited
Voice modeNoYes (advanced)
Message limit5x more than freeVirtually unlimited
Web searchYesYes
Agent creationYes (Projects)Yes (GPTs)
GDPR complianceYesYes (with opt-out)

My personal recommendation

  • Claude Pro if you write a lot in French, do complex code, or want the best writing quality
  • ChatGPT Plus if you want the widest palette (image, voice, agents, multimodal)
  • Mistral Le Chat Pro if European sovereignty is important
  • Gemini Advanced if you're in the Google ecosystem (Workspace, Android)

Option 3: API (for developers)

The API allows you to integrate AI into your own applications. You pay per token (≈ 1 word = 1.3 tokens in French).

Token, what's that?
A token is a unit of text. For French: 1,000 tokens ≈ 750 words ≈ 1.5 A4 pages. You pay for input (what you send) and output (what the AI generates). Output is generally 3-5x more expensive than input.

Current API prices (per million tokens)

API input price - 2026 ($/M tokens)

GPT-5 (OpenAI)15$
Claude Opus 4.715$
Gemini 3 Pro7$
Mistral Large 36$
Claude Sonnet 4.63$
GPT-5 mini0,4$
Gemini 3 Flash0,3$
DeepSeek V30,27$

What does it really cost?

Here are some typical use cases to help you anticipate:

Estimated monthly cost by usage

 Use caseMonthly cost
SaaS chatbot, 1,000 conversations/day30K tokens/conv≈ £200-500
SEO article generation, 100/day20K tokens/article≈ £300-800
Internal assistant, 50 employeesModerate usage≈ £50-150
B2C mobile app, 10K active usersAverage usage≈ £1,000-3,000
Document analysis (RAG)1,000 docs/month≈ £50-200
API pitfalls to avoid
- No caching: if you send the same context 10 times, you pay 10 times → use prompt caching (Claude/GPT offer this at -50 to -90%) - No monitoring: you could end up with a £5,000 bill from an infinite loop bug - Wrong model: using GPT-5 for simple tasks costs 50x more than Claude Sonnet or Gemini Flash - Output too long: limit max_tokens to avoid surprises

Option 4: Self-hosting (open-source)

You can download open-source models and run them on your own server. Models to consider:

  • Llama 4 (Meta), top quality, but huge
  • DeepSeek V3 : excellent quality/price ratio
  • Mistral Small 3 : lightweight, perfect for specific use cases
  • Mixtral 8x22B : power/cost balance

Self-hosting vs API

 🏠Self-hosting☁️API
Initial costHigh (GPU)£0
Recurring costElectricity + maintenancePer volume
PrivacyTotalDepends on provider
Model performanceLimited by GPUMarket leading
Expertise requiredHigh (DevOps, ML)Low
UpdatesManual, complexAutomatic
SovereigntyTotalDepends on provider country

When self-hosting makes sense

  • Ultra-sensitive data (health, finance, defence)
  • Huge volume (millions of requests/day) where API becomes too expensive
  • Critical latency (video games, HFT finance)
  • Sovereignty required (government, certain EU companies)

When self-hosting does NOT make sense

  • ❌ You're just starting out and want to test
  • ❌ You don't have a DevOps + ML team
  • ❌ You want the latest model always available
  • ❌ Your volume is low (< £100/month in API)
📚Going further: hidden economics of LLMs

The true cost of an AI product

Many startups underestimate the cost of integrating AI. Here's the complete checklist:

Direct costs

  • API tokens (input + output)
  • Embeddings for semantic search
  • Vectorstore (Pinecone, Weaviate, etc.): £50-500/month
  • Image model if needed

Hidden costs

  • Latency: an LLM call takes 2-10 seconds → impacts UX, therefore churn
  • Caching: implementing a caching system to save 50-80%
  • Monitoring: prompt observability (Langfuse, Helicone, etc.): £50-200/month
  • Evaluation: automated response evaluation system
  • Moderation: content filters to prevent abuse
  • User support: 30% of tickets will be about hallucinations or AI refusals

Human costs

  • An ML/AI engineer: £80-150K/year loaded
  • Prompt engineering time: 2-5 days per feature
  • Documentation, team training: ongoing

The 10x rule

For every pound spent on API tokens, budget £10 in ancillary costs (infrastructure, monitoring, human resources, support).

This is why "£5 per customer in API so I charge £10" calculations are always wrong. Operating margin is closer to 30% than 50% on current AI products.

Smart strategy: multi-model

Instead of using one model for everything, route intelligently:

  • Simple tasks (classification, extraction, short summary) → GPT-5 mini or Gemini Flash (50x cheaper)
  • Complex tasks (reasoning, code) → GPT-5 or Claude Opus
  • Verification (eval, moderation) → Claude Sonnet (quality/price ratio)
  • Embeddings → text-embedding-3-small (very cheap)

This strategy can divide your bill by 5-10.

Moving between thresholds

Many companies follow this path as they grow:

  1. MVP: ChatGPT Plus at £23 (whole team shares 1 account)
  2. Early product: basic API, ~£200/month
  3. Product-market fit: optimisations + caching, multi-model, ~£2,000/month
  4. Scale: price negotiation with OpenAI/Anthropic, ~£10K/month
  5. Hyperscale: self-host for high volumes, keep API for premium cases

Practical cases: how much should I pay?

Student, research, occasional writing

Verdict: free version is enough. Combine ChatGPT Free + Claude Free + Gemini Free. Cost: £0

Freelancer, knowledge worker, daily usage

Verdict: 1 Pro subscription (Claude Pro or ChatGPT Plus depending on preference). Cost: £18-23/month

Small startup, AI product integration

Verdict: API + monitoring + caching. Budget £200-500/month to start. Cost: £200-500/month

Company > 50 people

Verdict: combination of Pro accounts + enterprise API + potential self-hosted for confidentiality. Cost: £500-5,000/month depending on usage

Very large company, massive volumes

Verdict: dedicated team, enterprise pricing negotiation, partial self-hosting on Llama 4 or DeepSeek. Cost: £10K-100K/month

🎯 Find the perfect model for YOUR budget

Filter by price, free tier, GDPR, open-source. 30 models compared in detail.

Compare models

Quiz: your AI economics instinct

🧠 Quiz
Question 1 of 5

For occasional personal use, what's the best strategy?

Going further

  • 🎯 Compare models with price filter: find the best quality/price ratio
  • 🧠 Understand benchmarks: MMLU, HumanEval, Arena Elo... decode what really matters
  • Learn prompt engineering: a good prompt with a small model often beats a bad prompt with a large model

My final advice: start free, upgrade when you feel the friction, and don't hesitate to combine several services to get the best from each. ✨

Tags
ÉconomiePricingAPIComparatifROI

Read next