Back to Posts / GPT-5.6 vs Mimo-2.5-pro

GPT-5.6-terra used 48.5% more context than Mimo-2.5-pro

By Max Trivedi

Tl;Dr: I dug up the transcripts of GPT-5.6 Terra vs Mimo-2.5-pro from my agent's task history and found gpt-5.6-terra used on average 48.5% more transcript tokens than Mimo in my tasks (sample size >80 for each). Mostly due to pulling unnecessary reads, running noisy commands, and poking at Git when nobody asked.

OpenAI had given me free credits that expire soon so I spent the last few days testing gpt-5.6 as my main model. I used mostly Terra in my work.

My overall feeling was, it does get things right most of the time but results in a much larger context window than necessary. This was obviously anecdotal, so I thought I should check against the previous model I had been using (mimo-2.5-pro) for several days.

The setup: Same repo (https://github.com/dirac-run/dirac), relatively averaged out task distribution due to a large enough sample size, Dirac was the harness used.

Data: Only the transcript size at the task end was picked, to eliminate cache effects entirely (Mimo cached significantly better than Terra, for reference).

Terra vs Mimo: Task Metrics Comparison

Average Final Transcript Size

Model Tasks Avg Transcript Tokens
gpt-5.6-terra 95 91,794
xiaomi/mimo-v2.5-pro 81 61,831

gpt-5.6-terra's average transcript is 48.5% larger than xiaomi/mimo-v2.5-pro's (91,794 vs. 61,831 tokens).

Average Tool Calls per Task

Tool gpt-5.6-terra xiaomi/mimo-v2.5-pro
read_file 6.99 12.68
execute_command 6.36 6.31
search_files 4.24 7.32
edit_file 3.74 2.63
get_function 2.36 3.31
list_files 1.01 1.26
write_to_file 0.59 0.99
get_file_skeleton 0.63 0.96
diagnostics_scan 0.56 0.56
say 0.51 0.70
find_symbol_references 0.18 0.06
replace_symbol 0.02 0.05

Observations

  • read_file: Mimo emitted a lot more file reads but they were very frugal reads with disciplined line ranges; Terra preferred at least a few hundred lines every call.
  • execute_command: Anecdotally, this was the biggest token waste in Terra’s runs. It showed little to no harness awareness. Executed a bunch of commands which resulted in very large outputs. For instance, npm run test:unit 2>&1 (Terra version) vs npm run test:unit 2>&1 | tail -20 (typical Mimo version).

Terra also showed a surprisingly high number of edit-and-check loops, even for trivial changes such as, make one line change, and make sure things compile/test before making the next change. This caused more edit_file tool calls on average. Mimo on the other hand produced all changes first and then went on fixing loops once changes were done.

Most annoying parts

  1. Terra showed a tendency to maximize context window, even when the things it read were evidently not relevant (the "you are right, I overstepped" case). These qualities of maximal exploration and measured small incremental change/test loops make models score high on benchmarks but also make the model very annoying to use.
  2. Terra showed an obsession with Git. Used Git as a debugging tool by constantly examining Git commits and Git diffs. So much so that I had to change the AGENTS.md to instruct it not to do that too much; it ignored the instructions.

Overall, Terra is not a bad model by any means, but I wouldn’t use it at the current price - mainly because of the volume cost and to a larger extent, the amount of time it takes to get tasks done. The experience felt a bit more sluggish than expected.