Claude Opus 4.8 can be 10x faster than OpenAI GPT-5

A blogger reports that Anthropic's Claude Opus 4.8 is nearly 10 times faster than OpenAI's GPT-5 in rewriting blog comments, with both models producing equally good results. The blogger also notes a 3x price difference between GPT-5.4 ($2.50 per input token) and GPT-5.4-mini ($0.75), and is switching from the litellm wrapper to native SDKs due to security concerns.

This picture summarizes it well: Here on my blog, for this popular blog post /plog/blogitem-040601-1 I get a lot of comments. 28k blog comments over the years. Some of them are terribly written and hard to understand, so I let AI suggest a rewrite. That code that sends the blog post comment to AI, I actually fire off three times: once with OpenAI gpt-5 , once with OpenAI gpt-5-mini , and once with Claude claude-opus-4.8 . I use my human eyes and judgement to evaluate the results, and I can tell you they do equally well. Only the slightest differences. The surprising thing is how amazingly slow OpenAI's gpt-5 is It's nearly 10x slower than claude-opus-4.8 . What's up with that ? It's also clear that the latency difference between gpt-5-mini and gpt-5 is significant. At the time of writing, the input token price difference between gpt-5.4 and gpt-5.4-mini is $2.50 compared to $0.75 That's a 3x difference. Conclusion conclusion - If you're constructing a prompt the API, use Claude. - If you have to use OpenAI, consider the mini model because it's both cheaper and faster. Bonus bonus Before I added Claude, I used to use litellm to wrap OpenAI's models. The code looks like this: response = litellm.completion model="openai-gpt-5", api key=settings.OPENAI API KEY, messages=my prompt messages, Unlike, if you use the native OpenAI Python SDK the invocation looks like this: client = openai.OpenAI api key=settings.OPENAI API KEY response = client.responses.create model="gpt-5, input=my prompt messages, I measured the difference, in speed, where I compare using the OpenAI SDK versus the litellm wrapper and the difference looks like this: Granted, in June I "only" did a bit over 30 of these calls, but strangely there's a difference I don't have the intricate knowledge to understand why the litellm makes the total time different from using the SDK. not sure I care either Either way, I'm moving away from litellm and only use the SDKs provided by Claude and OpenAI. Feels safer given the CVEs we've seen this year on litellm .