Head-to-headTested AIBy DanPublished Tested Updated
GPT-6.1 Terra vs Opus 5.5: three real projects, and the designs I picked blind
The report
Verdict
Code: Terra, narrowly · Designs: Opus, picked blind
Terra finished every solo job fastest and won the code fix narrowly because it handled uncertain data more honestly; neither fix met every requirement. Opus's fix idled on 3.6× less CPU than Terra's, and in both design trials the page I picked blind was Opus's.
On the tier list: Linked from the Claude Opus 5.5 and GPT-6.1 Terra entries on my tier list. It moves no agent; Terra's B placement is my own verdict, not a score from these trials.
| Trial | Claude Opus 5.5 | GPT-6.1 Terra | Result |
|---|---|---|---|
| 1 · Fix GIT_REAL's idle CPU | 1.12% of a core idle · 1 h 45 min | 4.02% of a core idle · 38 min | Judge: Terra, narrowly. Neither met every requirement |
| 2 · Redesign the GIT_REAL page | 96/100 · 3 h 24 min, with my feedback | 85/100 · 21 min, one pass | My blind pick and the third judge: Opus |
| 3 · Redesign Terminal Ward | Solo: 2 h 52 min | Solo: 36 min | Team of Opus + 6 Terra: 38 min. My blind pick: Opus solo |
Three real jobs from my workshop
On October 29 I gave Claude Opus 5.5 and OpenAI's GPT-6.1 Terra the same three jobs from my own workshop, one run each. Opus worked in Claude Code and Terra in the Codex CLI, both at max effort, each in its own folder with the same brief and told to work alone, with no helper agents. Trial 1 was a bug fix in GIT_REAL, my dashboard for the 74 Git repositories in my workshop. Trials 2 and 3 were redesigns: the GIT_REAL page, then Terminal Ward, the page where I watch my agents and terminals. Trial 3 added a third entry: Opus directing six Terra workers, each in its own terminal.
Everything ran on sample data: 74 made-up repositories for GIT_REAL and fixture sessions for Terminal Ward, never my live projects or agents.
Trial 1: the dashboard that spun the fans
GIT_REAL re-read all 74 repositories from scratch every cycle, even when nothing had changed. On the sample fleet the untouched code used 35.1% of a CPU core while idle, and on my real fleet the page said it refreshed every 30 seconds while it actually took about a minute. The job: make it at least 5× cheaper when nothing changes, show every real change within 30 seconds, and never show old data as fresh.
Both beat 5×. In the judge's shared two-minute check, the untouched code used 38.97% of a core, Opus's fix 1.12% and Terra's 4.02%. Opus's version checks a repository without running Git at all when it can prove nothing changed; Terra's still starts one Git command per check.
Then the judge, GPT-6.1 Sol, built four valid Git setups, one ordinary and three unusual, and made a real change in each. Both fixes caught the ordinary branch within about 10 seconds. Opus's still reported two of the other three as unchanged and healthy after 36 seconds, including a repository whose object store was broken. Terra missed one. Both also shared an older design flaw: one slow repository can hold back fresh results for all the others. The judge gave Trial 1 to Terra, narrowly, because the brief said never to trade truth for speed, and failed both against the full requirements. Nothing was merged.
Each lane first had a team of GPT-6 Astra reviewers. After Opus had finished, its review lead messaged the Opus session. The judge found no new code or test from Opus after that message and did not count it for or against Opus.
Trial 2: redesigning the GIT_REAL page
Both agents had to rebuild the GIT_REAL page in my FABLE design style, using only real data from the sample fleet. Terra delivered in 21 minutes in one pass. Opus took longer, and I stepped in twice: early on I said yes to installing tools it needed to see the page, and after seeing it at my real window size I approved a fix and polish pass.
Two Kimi K3.1 Pro reviewers scored the pages separately, Opus's at 96/100 and Terra's at 85/100, each with its own weights. A third Kimi read both reviews and chose Opus's: "A wins decisively." With the builders' names hidden, I picked Opus's page too. By Opus's own measurement on my GPU, its page ran at 59–60 fps on 44–48% of the browser's CPU, while the FABLE reference page it was asked to beat used 97–98% for 37–38 fps. At my window size, 7 to 10 of the sample's deliberately long repository names, 23 characters or more, end in "…".
Trial 3: Terminal Ward, solo and as a team
All three entries started at 1:02 PM ET. Terra working alone finished in 36 minutes, and the team in 38: I told the Opus orchestrator to open six Terra workers in their own terminals, and it split the page between them. Opus working alone took 2 hours 52 minutes.
All three pages fit my 1688×1100 window with no page scroll in their Habitat and List views, threw no page errors and kept StarNet, a part of the page they weren't allowed to touch, byte for byte. After looking through all three with the names hidden, I picked Opus's solo page. When I opened an agent inside my narrower FABLE window, its Habitat columns squeezed together and some labels overlapped.
Limits of this report
- One run per agent per trial. My picks are what I would keep looking at; the review scores came from separate reviews with different weights, not one shared scoring sheet.
- Times run from each agent's first prompt to its last message. Opus's Trial 2 time includes the fix and polish pass I approved, and max effort doesn't mean equal compute across the two products.
- Opus's page figures in Trial 2 are its own measurement, and Terra's page wasn't measured the same way.
- Trial 3 had no reviewer scores, and the team ran as a separate method, so its result is not a solo result for either model.
Follow the lab
New test reports and tier-list changes get announced in my Discord's #news channel. The full videos are on YouTube.
Receipts
DanManREAL · AI lab
Test receipt
GPT-6.1 Terra vs Claude Opus 5.5
- Trial 1 verdict, GPT-6.1 Sol as final judge
- Terra, narrowly · both failed full acceptancethe judge put truth before speed: Terra refused to trust data it couldn't prove unchanged more often
- Trial 1 idle CPU, same 74-repo sample fleet
- 38.97% → Opus 1.12% · Terra 4.02%share of one CPU core over about 120 s with nothing changing, Git included; 34.9× and 9.7× below the untouched code
- Trial 1 changes still hidden after 36 s
- Opus 2 of 4 · Terra 1 of 4the judge's four valid Git setups, one ordinary and three unusual; a change should show within 30 s
- Trial 1 time to finish
- Terra 38 min · Opus 1 h 45 minboth started at 2:35 AM ET; Terra's one turn took 2,300 s, Opus answered at 4:20 AM
- Who built which lane, from the agents' own session records
- Lane A Opus 5.5 · Lane B GPT-6.1 Terrain all three trials; Trial 3 added lane C, Opus directing six Terra workers
- Trial 2 design reviews, Kimi K3.1 Pro
- Opus 96/100 · Terra 85/100two separate reviews with different category weights; a third Kimi read both and picked Opus's page
- Trial 2, my blind pick
- Opus's pagepicked with the builders' names hidden: "winner of trial 2"
- Trial 2, Opus's page on my GPU, its own measurement
- 44–48% CPU at 59–60 fpswhole-browser CPU at 1396×1153; the FABLE Grid reference page it was asked to beat ran at 97–98% at 37–38 fps
- Trial 2 time to finish
- Terra 21 min, one pass · Opus 3 h 24 minOpus's run included two messages from me: a yes to installing tools, then my real window size and a yes to a fix and polish pass
- Trial 3 time to finish, from the same start
- Terra 36 min · team 38 min · Opus 2 h 52 min1:02 PM ET start; the team was Opus directing six Terra workers in their own terminals
- Trial 3 fit in my 1688×1100 window
- 3 of 3 · no page scroll · no page errorsHabitat and List views checked for every entry; all three kept StarNet's two locked files byte for byte
- Trial 3, my blind pick
- Opus soloafter looking through all three previews with the builders' names hidden: "this one lol"
Verdict
Code: Terra, narrowly · Designs: Opus, picked blind
Tier list Linked from the Claude Opus 5.5 and GPT-6.1 Terra entries on my tier list. It moves no agent; Terra's B placement is my own verdict, not a score from these trials.
danmanreal.com/tests/terra-vs-opus-three-blind-trials
What was tested
On the tier list's bench
- Claude Opus 5.5 · 5 receipts on file
Shorts
Related
- AI Coding Agents Tier List
- GIT_REAL: free safety checks for AI agents
- How I test AI tools
- Steps → Reduced: opening a terminal agent went from 98 seconds to 11
Independent testing. Sponsorships are always disclosed, and verdicts are never for sale.
Disagree? Bring your receiptsAll reports
- One inbox for my AI coding agents: triage took 34% less time, resuming work did not improveHead-to-headNov 20
- Mythos 5.2: research and security benchmark source reviewPublic benchmarkNov 17
- Fable 5.2: agentic coding benchmark source reviewPublic benchmarkNov 17
- Claude Opus 5.5: public coding benchmark receiptsPublic benchmarkNov 17
- Astra 6.2: new model and public task-evaluation evidencePublic benchmarkNov 17
- Sol 6.2: new model and public coding benchmark evidencePublic benchmarkNov 17
- Gemini 4 Carbon: benchmark source and usage profilePublic benchmarkNov 17
- Qwen 4: Qwen4-Max launch and benchmark sourcesPublic benchmarkNov 17
- Kimi K3.1 Pro: public coding benchmark receiptsPublic benchmarkNov 17
- GLM-5.3: public coding benchmark receiptsPublic benchmarkNov 17
- Gemini 4 Argon: public coding benchmark receiptsPublic benchmarkNov 17
- Claude Sonnet 5.5: public coding benchmark receiptsPublic benchmarkNov 17
- DeepSeek V4.1 Flash: public coding benchmark receiptsPublic benchmarkNov 17
- Muse Wonder Full 1.0: public coding benchmark receiptsPublic benchmarkNov 17
- Grok 4.7: separate release and public coding resultsPublic benchmarkNov 17
- ai-design-check on two of my FABLE screens: 10 warnings, 4 of them false alarmsField testNov 2
- Vercel Agent Desktop used a sample Trip app in 6 actions without moving the mouseField testNov 1
- Cerberus Guard on sample data: one crawler name, two verdicts, and a block earned by behaviorField testOct 31
- GPT-6.1 Terra vs Opus 5.5: three real projects, and the designs I picked blindHead-to-headOct 29
- Steps → Reduced: opening a terminal agent went from 98 seconds to 11AuditOct 28
- One ChatGPT chat directed Opus and Codex at once through DMR_REMOTEField testOct 28
- Claude Duet ran my channel for a week and caught 269 fake AI crawlersField testOct 27
- Kimi K3.1 Pro built my animation family, after a quota stop and 21 minutes of failed requestsField testOct 24
- ChatGPT + Codex built 33 new DMR_REMOTE toolsField testOct 23
- Gemini 4 Argon reviewed a finished episode: 14 clips, one flag, one inconsistencyField testOct 21
- Qwen 3.8 Flash Next on my RTX 5080: 81 tokens a second with 202K tokens readAuditOct 21
- Does an extra AI review step pay off? A small DeepSeek trial said noHead-to-headOct 19
- My real DeepSeek bill: 12.31 billion tokens for $122.28Cost receiptOct 17
- Gemini 4 Argon kept hanging in Gemini CLI. The cause was a 60-second timeoutIncidentOct 14
- I asked 10 AI coding agents for their biggest weaknessSelf-reportOct 14
- Jarvis 27B on my own PC: 30 tokens a second, and it still recalls at 251KAuditOct 10
- My agents' first game benchmark was wrong. Here's how it was caughtIncidentOct 10
- Pixal3D + Blender turned my 2D art into animated 3D charactersField testSep 22
- DeepSeek Flash answered 200 OK and sent nothing for two hoursIncidentSep 14
- Grok, Codex and DeepSeek played WorldOS for Rank 1. Grok won with 98Head-to-headAug 20
- DeepSeek V4 Flash audited 18 of my real coding sessions. Two said 'done' too earlyAuditAug 4