Terra scored 47 on a test Opus 5.5 did not finish

A premium coding model that costs eight times more should finish the task it was given. On one of Jessica Wachtel's three tests this week, Claude Opus 5.5 did not.

Wachtel at The New Stack ran GPT-6.1 Terra against Opus 5.5 on three tasks. OpenAI launched Terra on October 26. OpenAI claimed Terra came close to Opus on accuracy at a fraction of the cost. Wachtel put that claim through her own tests, with each vendor's effort setting at max.

Terra tied the first test. Terra won the second. Terra finished the third. Opus 5.5 hit its 128,000-token output cap on the third and returned no final answer. Total estimated uncached cost: $25.99 for Terra and $214.96 for Opus. That is a ratio of about one to eight.

The tests

The first test was release triage. Each model had to fix a failed software release through a screenshot browser interface. Both scored 50 out of 50. Both found all six root causes. Both ignored three planted instructions to bypass safety settings. Opus typed URLs 102 times to move between pages. Terra typed 23 times and clicked through the site. Opus wrote about five times as many output tokens. Wachtel's cost estimates: $129.20 for Opus, $22.16 for Terra.

The second test was cell lineage. Each model had to write and run Python to track dividing cells in a 72-frame microscope movie. Terra scored 48 out of 50. It answered every summary question correctly. Opus scored 43.5. It missed two divisions and reported one that never happened. Opus ran code 96 times over 129 turns, in about 36 minutes. Terra ran code 30 times over 37 turns, in about 15 minutes. Costs: $3.47 for Terra, $82.91 for Opus.

The third test put about 1,200 rows of CI test results in a single prompt with no tools. The data mixed three time zones, three duration formats, renamed tests, duplicate rows, and a price change mid-month. The prompt asked 50 questions. Terra scored 47 out of 50. Opus hit its 128,000-token output cap before producing a final answer. Wachtel says that is a failure at this configuration, not a statement about Opus in general.

What the numbers do not say

OpenAI's own marketing cites Agents' Last Exam, a benchmark of computer-use tasks. Terra scored 60.0% at $1.01 per task. Opus scored 63.2% at $8.82. OpenAI framed that as cheaper and faster at similar accuracy. Wachtel writes that the vendor's framing understates both the score gap and the cost ratio, which is 8.73 to one.

Each of her figures is from one run at max effort. Wachtel is explicit that this is not a measurement of repeatability. The dollar figures are her estimates against the vendors' standard API rates, not billed amounts. Prompt caching lowers them: $2.51 for Terra, $21.52 for Opus.

For a team routing agent work between models, the useful finding is the ratio and the shape of the failure. On the two tasks where Opus had tools, it burned more tokens for a lower score. On the task without tools, it hit the output cap. Terra kept going.

Source

GPT-6.1 Terra vs. Claude Opus 5.5: Cheaper and faster in three tests at max effort, by Jessica Wachtel for The New Stack.