Published 2026-06-30 · reference prices, verify before budgeting
The 2026 LLM price war has a new shape: xAI's Grok has muscled into the conversation as a cost leader, OpenAI and Google have split into clear budget and premium tiers, and Anthropic's Claude still sits at the top of the price ladder. To make the comparison honest we did what we always do — fixed one workload and priced it on every major model. The job: 1,000,000 requests, each 1,000 input tokens and 500 output tokens (a typical short chat or generation call). That's 1 billion input tokens and 500 million output tokens.
| Model | In $/1M | Out $/1M | Cost for 1M requests |
|---|---|---|---|
| Grok 4.1 Fast cheapest frontier | $0.20 | $0.50 | $450 |
| DeepSeek V3 | $0.27 | $1.10 | $820 |
| GPT-5.4 mini | $0.40 | $1.60 | $1,200 |
| Gemini 2.5 Flash | $0.30 | $2.50 | $1,550 |
| Gemini 3.1 Pro | $2.00 | $12.00 | $8,000 |
| GPT-5.4 | $2.50 | $15.00 | $10,000 |
| Grok 4 (flagship) | $3.00 | $15.00 | $10,500 |
| Claude Opus 4.6 | $5.00 | $25.00 | $17,500 |
Prices are reference estimates, July 2026. Report outdated price →
Same workload, $450 on Grok 4.1 Fast versus $17,500 on Claude Opus 4.6 — a ~39× spread. The single biggest decision on your AI bill is still which model you route to, and in 2026 the cheap end of the frontier moved to xAI.
Grok went aggressive on price. xAI's Fast tier at $0.20/$0.50 undercuts even DeepSeek on output and sits well below every Western frontier model — a deliberate land-grab, bundled with a huge context window and live X data. OpenAI and Google bifurcated. Each now has a genuinely cheap small model (GPT-5.4 mini, Gemini 2.5 Flash) and a pricey flagship (GPT-5.4, Gemini 3.1 Pro), so "which is cheaper, OpenAI or Google?" depends entirely on which tier you mean. Claude stayed premium. Opus remains the most expensive per token of the majors — Anthropic is pricing on capability and reliability, not on cost leadership. If your budget is the constraint, Claude Haiku (much cheaper) is the tier to look at, not Opus.
The number that wins this table doesn't automatically win your project. A cheaper model that needs a retry, writes longer answers, or fails on hard prompts can cost more in practice than a pricier one that nails it first time. Output length matters too — output is billed 2.5–5× input across every model here, so a verbose cheap model can lose to a terse expensive one. The pattern most teams settle on is tiering: route routine, high-volume traffic to a cheap model (Grok Fast, Gemini Flash, DeepSeek) and escalate only the genuinely hard requests to a flagship. Even sending 20% of calls to a premium model keeps your blended cost near the cheap end of the table.
Once the model is picked, three things move the bill: output caps (set max_tokens — output is the expensive half), prompt caching (repeated system prompts are billed at a steep discount on OpenAI, Anthropic, Grok and DeepSeek), and routing (cheap-first, escalate on uncertainty). Each compounds with the others. Model the savings with the prompt caching calculator and the model routing calculator.
Your tokens aren't ours, so the ranking can flip on output-heavy or input-heavy traffic. Plug your real input/output lengths and volume into the cheapest LLM API ranker — it sorts every model for your numbers — or the AI API cost calculator. Building something specific? Use the AI app cost estimator, and if you resell AI in a subscription, the AI wrapper pricing calculator shows whether your price actually clears the API cost. Per-provider guides: Grok, OpenAI, Claude, Gemini, DeepSeek.
Open the Cheapest LLM API ranker → · OpenAI vs Claude vs Gemini →
Reference prices (June 2026), gathered from public pricing pages and provider comparisons; model names and rates change often — verify before budgeting. Not affiliated with any provider.