When a model that charges $25 per million output tokens sits beside one that charges $1.2, the difference shows up on the bill before you even finish a prompt. Anthropic’s claude‑opus‑5 lists $5 input and $25 output per 1 M tokens, xAI’s grok‑4.6 lists $2 input, $0.5 cached input, and $6 output, while OpenAI’s gpt‑5.6‑luna lists $0.2 input, $0.02 cached input, and $1.2 output. The gaps are not subtle: the output cost of Claude is more than twenty times that of OpenAI, and Grok’s output sits at half of Claude’s but five times OpenAI’s.
## Token economics in a nutshell
| Vendor | Model | Input | Cached input | Output |
|--------|-------|-------|--------------|--------|
| OpenAI | gpt‑5.6‑luna | $0.2 per 1 M tokens | $0.02 per 1 M tokens | $1.2 per 1 M tokens |
| Anthropic | claude‑opus‑5 | $5 per 1 M tokens | $0.5 per 1 M tokens | $25 per 1 M tokens |
| xAI | grok‑4.6 | $2 per 1 M tokens | $0.5 per 1 M tokens | $6 per 1 M tokens |
The numbers reveal that OpenAI’s pricing is laser‑focused on keeping per‑token costs low, especially when a prompt can be cached for only $0.02. xAI’s model offers a modest cache discount but its baseline input price is four times higher than OpenAI’s, and its output sits at half of Claude’s yet still five times OpenAI’s. Anthropic’s model is the most expensive in every column, reflecting a premium on safety and compliance features.
For a developer who needs to generate millions of tokens daily, the choice comes down to cost per token: OpenAI’s gpt‑5.6‑luna delivers the lowest bill, Grok‑4.6 offers a middle ground for those who value its multi‑agent design, and Claude‑Opus‑5 is the best fit for teams that prioritize Anthropic’s safety guarantees and can absorb the higher price. In short, if budget and volume are the main concerns, gpt‑5.6‑luna wins; if compliance and safety are paramount, claude‑opus‑5 is worth the extra spend; and if you need a hybrid of performance and multi‑agent capability, grok‑4.6 sits comfortably between the two.
Pricing verified on 2026-09-12.