OpenAI API Pricing Hits $0.05 to $180/M Tokens

OpenAI API users now face a spread from $0.05 to $180 per million tokens, depending on model and meter. BenchLM’s August 7, 2026 pricing tracker says the newest GPT-5.6 tiers and older Pro models can differ by hundreds of times on the same workload.

BenchLM updated OpenAI API pricing for August 2026, covering token rates, cache discounts, batch pricing, and long-context charges.

What changed

BenchLM’s registry now reflects the GPT-5.6 family, which went GA on July 9, 2026. Sol, Terra, and Luna all ship with a 1.05M-token context window, while Terra and Luna saw price cuts on July 30: Terra fell 20% and Luna fell 80%.

The pricing page also keeps older OpenAI families live, including GPT-5.5, GPT-5.4, GPT-5.1, GPT-4o, and the o-series. That matters because the cheapest listed options still start at GPT-5 nano’s $0.05 per million input tokens, while GPT-5.5 Pro and o1-pro sit at the top end of the chart.

• GPT-5.6 Sol: $5 input / $30 output per million tokens
• GPT-5.6 Terra: $2 input / $12 output per million tokens
• GPT-5.6 Luna: $0.20 input / $1.20 output per million tokens
• GPT-5.5 Pro: $30 input / $180 output per million tokens
• Long-context meters raise flagship rates above short-context pricing

Why it matters

For developers, the bill is now shaped by more than prompt length. Cached input bills at 10% of standard input rates, Batch API cuts both directions in half, and GPT-5.6 cache writes cost 1.25x standard input with a 30-minute minimum cache life.

That means the same app can swing from cheap to expensive based on routing, prefix reuse, and whether it runs asynchronously. BenchLM says a cached and batched Sol workload can land near a quarter of list price, which makes model selection and request design part of cost control, not just ops cleanup.

OpenAI’s API is still billed separately from ChatGPT plans, so a paid consumer subscription does not buy API credits. The practical question for teams is simple: do you need flagship quality, or can Terra, Luna, or an older model hit the target at a much lower rate?

The takeaway is not just that OpenAI prices vary widely, but that the cheapest viable model and the request pattern around it now decide the real bill.