← sources

Real world comparison of Terra 6.1 vs Sol 6.1 vs Sonnet 5.5 vs Luna 6

source: reddit (reddit:OpenAI)published 2026-10-28 01:30ingested 2026-10-29 02:39

retained text

*Processing img 2raj791o24yh1...*

tldr; Terra 6.1 launched and is fantastic. Our old baseline of using Sol 6.1/Sonnet 5.5 for chat is now replaced with Terra 6.1/low reasoning/balanced output with better results and lower cost. Our voice baseline of Luna 6, is significantly better with Terra 6.1/no reasoning/concise. Well worth the added cost.

First of all, this is by no means a highly scientific test. It's a real world use case of what I'm testing for my specific company as a real world user looking at models. Normally, our baseline doesn't change but I thought I'd share since this time it did.

Quick context. I am founder of [ChatRep](https://chatrep.com/) and we are basically an AI native BPO. So blend AI and human. Our focus is on quality as opposed to trying to eliminate human interaction. I say this because as we evaluate models, for us, we don't care too much about the model cost. (Chat isn't too heavy on token usage and we charge $2000+/mo and not $25/mo for context).

Also, with benchmarks, they often ignore things like reasoning level or settings like concise/balanced in response. Those are very important for CX type applications.

We also specify different models for chat vs voice. Latency is critical in voice so models that allow "none" for reasoning are better for us. We experimented with Opus/Fable and Astra but those were overkill. Took way too long and often gave way too long an answer. Even when it's tailored to a specific use case like technical support area with technical documents.

So our standard was Tossup between Sol 6.1 or Sonnet 5.5 for chat applications. These models were significantly better than Luna and balanced speed, accuracy and quality of response really well. We use objective measures like pure accuracy, speed and word count. But we also look at things a bit qualitatively... were the words useful, natural, engaging. Did it include appropriate links when necessary and not excessively.

Costs are standardized so I can compare models but is not a key driver. (more of a tie-breaker)

Voice was quite different. Sol doesn't have a "none" option for reasoning and both Sol and Sonnet were bit worse on latency. (I'm not including the TTS/STT models... those were separate tests and things like English vs foreign language play a big role). So it was down to Luna 6 or Haiku 5.5. Luna was a tad better overall in our tests (barely) so that became our default for voice.

Then Terra 6.1 came out. On the surface, we weren't impressed. Seemed like another average LLM. Bit lower cost than Sol. probably slotted in-between Luna and Sol. But we are quite impressed.

For us, Terra 6.1 with low or medium reasoning gave great responses. 100% accurate in our test. Low was faster and consistently as good as medium. What surprised us is that on the qualitative side, it was bit better than Sol. Sol tends to over do things for chat. Terra was better at getting to the point in a natural way. Even in concise mode, SOL tended to answer longer than Terra in balanced.

Then to me, the most exciting part was how well it performed with voice. Terra 6.1 has "none" as a reasoning option and it still scored 100%. It was significantly better than Luna 6. Anthropic models don't have an option to disable reasoning so we tend not to use them. But I'm looking forward to the next Haiku model coming out and will test against that. Luna is still much cheaper than Terra if that is important to you.

I'm thrilled with how well Terra performed for us. This is specific to fairly simple rag database type interactions with low output. It's not for coding, 3d rendering, etc. Although word has it the new Haiku model will be out any day now so who knows.

document fetched 2026-10-29 00:29 · canonical https://reddit.com/r/OpenAI/comments/1xhb1hu/real_world_comparison_of_terra_61_vs_sol_61_vs