benchrPricing breakdown
| Tier | Rate / 1M tokens |
|---|---|
| Standard input | $1.250 |
| Cached input | $0.125 |
| Standard output | $10.000 |
| Context window | 400,000 tokens |
| Max output | 128,000 tokens |
The 400K context ceiling
GPT-5's 400K limit is below the current 1M standard set by Claude, DeepSeek, and Gemini. For most workloads this doesn't matter. Where it does: analyzing full large repositories, processing entire legal contracts, or running extended agent conversations without truncation. If you're regularly approaching 300K tokens of context, the limit is a functional constraint.
A 90% cached-input discount
OpenAI prices GPT-5 cached input at $0.125/1M tokens, 90% below the $1.25 standard input rate. For agentic workloads with large repeated system prompts, eligible cache hits can sharply cut the repeated-prefix portion of the input bill.
Capabilities to test
GPT-5 is positioned for general-purpose text work such as writing, summarization, Q&A, and analysis. OpenAI reports 74.9% on SWE-bench. Treat that figure as one coding signal and test assisted-development tasks with the same review process you will use in production.
Cost scenarios
At 10M input tokens + 3M output tokens per month: $12.50 input + $30 output = $42.50/month. At 100M + 30M: $425/month. These are far lower than Claude Sonnet ($540/month) or GPT-5.5 ($1,250/month) at the same volume, which is why most teams default here.
Use-case fit
Consider for: General-purpose production traffic, broad text tasks, and coding agents that fit within the 400K context limit.
Skip if: You need the absolute deepest reasoning — use 5.5; Very large context windows.
At $1.25/1M input and $10/1M output, GPT-5 should be judged on total task cost rather than headline input price alone. Long prompts, retries, and output-heavy generations can make output pricing dominate the bill. Its 400,000-token context limit is useful only if retrieval quality and latency remain acceptable on your own documents. For coding-heavy evaluation, compare the 74.9% SWE-bench figure against your expected human-review cost.
Decision checklist
Before routing production traffic to GPT-5, run a small evaluation that matches the workload you actually pay for: prompt length, expected output length, cache reuse, retry behavior, and human-review rate. A public benchmark is a useful signal, but it does not tell you whether this model is cheaper for your own calls.
Benchmark it against one cheaper option and one stronger option; this tier is where routing rules usually save more money than choosing a single default. For OpenAI specifically, verify the live model ID, rate limits, and any regional availability constraints before treating the published token price as the final production cost.
The main operational check for this page: output-heavy workloads can cost far more than the input headline suggests.