How Much Do LLM APIs Really Cost? I Ran the Numbers for 5 Common Workloads in 2026

"How much will we actually pay?" is the question every team asks before committing to an LLM API. Published price-per-token tables are accurate but hard to translate into real monthly bills. This piece runs the numbers for five common workloads — chat app, code assistant, support bot, document Q&A, and batch summarisation — and gives you rough monthly cost bands for OpenAI, Anthropic, Google, and xAI in 2026. No fluff; the goal is to help developers and founders in the US, UK, Europe, India, and Australia budget realistically.

The Five Workloads (and Assumptions)

1. Consumer chat app (light use)

~50,000 input + 20,000 output tokens per user per month. Assumes a small B2C product with a few thousand active users. Mix of short turns and occasional long threads.

2. Code assistant / dev tool

~200,000 input + 80,000 output tokens per developer per month. Assumes daily use for completions, explanations, and refactors. Heavy on code context.

3. Customer support bot

~500,000 input + 150,000 output tokens per month per agent. Assumes the bot handles a meaningful share of tier-1 support; multi-turn conversations and knowledge-base retrieval.

4. Document Q&A / RAG

~1M input + 200,000 output tokens per month. Assumes internal docs or help-center RAG; repeated retrieval and medium-length answers.

5. Batch summarisation

~2M input + 400,000 output tokens per month. Assumes nightly or weekly jobs over reports, emails, or logs. Output-heavy.

Exact numbers depend on model choice (e.g. GPT-4o vs GPT-4o mini, Claude 3.5 Sonnet vs Haiku). The ranges below use typical "mid-tier" models where most teams land.

Workload 1: Consumer Chat App (Light)

Rough scale: 50K in / 20K out per user/month.

OpenAI (GPT-4o): ~$0.25–0.40 per user/month.

Anthropic (Claude 3.5 Sonnet): ~$0.20–0.35.

Google (Gemini 1.5 Pro): ~$0.15–0.30.

xAI (Grok): ~$0.01–0.02 (Grok is orders of magnitude cheaper per token).

At 5,000 users, you are in the $1,250–2,000/month range for OpenAI/Anthropic/Google, or well under $100 for xAI at similar usage. Switching to "mini" or "Haiku" tiers can cut these by 50–70%.

Workload 2: Code Assistant

Rough scale: 200K in / 80K out per dev/month.

OpenAI: ~$1.50–2.50 per dev/month.

Anthropic: ~$1.20–2.00.

Google: ~$1.00–1.80.

xAI: ~$0.05–0.10.

For a team of 20 developers, that is roughly $30–50/month (OpenAI/Anthropic/Google) or a few dollars for xAI. Code-assistant workloads are often among the most predictable; many teams lock in a single provider and optimise later with caching and model tiers.

Workload 3: Customer Support Bot

Rough scale: 500K in / 150K out per "agent"/month.

OpenAI: ~$4–7 per agent/month.

Anthropic: ~$3.50–6.

Google: ~$3–5.

xAI: ~$0.15–0.25.

At 10 equivalent agents, you are looking at $40–70/month for the big three, or about $1.50–2.50 for xAI. Support bots often need strong instruction-following and safety; Claude and GPT-4o are common choices even when cost is higher.

Workload 4: Document Q&A / RAG

Rough scale: 1M in / 200K out/month.

OpenAI: ~$8–14/month.

Anthropic: ~$6–12.

Google: ~$5–10.

xAI: ~$0.30–0.50.

RAG workloads are input-heavy (retrieval context). Long-context models (Claude, Gemini) can reduce round-trips; xAI and smaller tiers keep cost low if quality is acceptable.

Workload 5: Batch Summarisation

Rough scale: 2M in / 400K out/month.

OpenAI: ~$16–28/month.

Anthropic: ~$12–22.

Google: ~$10–18.

xAI: ~$0.60–1.00.

Batch jobs are where per-token price matters most. Many teams use the cheapest capable model (e.g. Haiku, Gemini Flash, or Grok) for summarisation and reserve premium models for user-facing features.

How to Use These Numbers

Treat these as order-of-magnitude estimates. Your mix of models, caching, and prompt length will shift the numbers. The point is to get a feel for which workload dominates your bill and which provider is in the right ballpark for your region and quality bar.

For a quick side-by-side of 2026 token pricing across providers, use the free LLM API Pricing Tracker on this site. For a deeper dive into when to choose which model, see OpenAI vs Anthropic vs Google vs xAI API Pricing 2026.