Claude Opus 5.5 vs GPT-6.1 Terra: What Three Synthetic AI Model Tests Mean for Dubai Businesses

GPT-6.1 Terra cost less than Claude Opus 5.5 and finished sooner in three synthetic tests at max effort. Jessica Wachtel built and ran those tests for The New Stack, and says the result belongs to those particular runs rather than a general ranking of the models.

If you run a business in Dubai, treat this as a reason to run a trial of your own before you change anything. The author lands in a similar place. For work like these browser and code-execution tasks, the author would run Terra at that top setting in a trial before committing to Opus. Below you'll find the results, their limits and what we'd do next.

What changed, and when

OpenAI launched GPT-6.1 Terra on 26 October. Its marketing materials compared the model directly with Opus 5.5 in two places, and Wachtel's own tests followed in The New Stack on 31 October.

The first benchmark, Terminal-Bench-Science 0.1, has agents write code for research tasks such as tracking cell divisions. OpenAI claimed Terra came within five percentage points of Opus 5.5 with fallbacks, at less than one-fifth the estimated cost.

The second was Agents' Last Exam, a benchmark of multistep computer-use tasks. Opus 5.5 scored higher there, and Terra cost less per task. Wachtel stresses that a score gap remains, so the two aren't equally accurate.

Those claims caught Wachtel's interest because Terra had left some totals uncalculated and given some wrong answers in an earlier test against GPT-6 Astra. So the author built two tests around computer use and scientific coding, then added a large-dataset analysis to see whether Opus 5.5 shares that weakness.

Claude Opus 5.5 and Terra pricing, as OpenAI describes it

At max effort, OpenAI describes Terra as more than three times as token-efficient as Opus 5.5. It also calls Terra one-quarter the price for standard input and output tokens. Separately, it reports an estimated cost per task of about one-ninth. Those are the vendor's own comparisons.

Wachtel offers a useful check on that price gap. With the same token mix and no other pricing adjustments, Terra would need four times as many tokens to match Opus 5.5.

Wachtel also says real task cost depends on token usage, cache reads and writes, and any adjustments that apply. So check the current rates on your own account and work out your own sum.

How the author tested Terra against Claude Opus 5.5

Wachtel ran the tests a few times because the two models scored similarly at easier levels. When scores came out too close, the author made the test harder and ran it again, up to three times. Everything below comes from the final versions.

Each final test ran once per model through OpenAI's and Anthropic's own APIs, with each vendor's setting at its maximum. The author points out those settings don't impose equivalent reasoning budgets across the two models. Turns weren't capped, and a script graded every run against a verified answer key.

The tests changed after the earlier results, and each final version ran only once. For that reason, Wachtel says the results don't establish repeatability or average performance across workloads. Keep that in mind for every result here.

The first two tests were agentic tool use, where the model had to click through a site or run code and act on the result. The third gave it no tools at all. The cost comparisons below are the author's estimates at standard uncached rates, not invoice totals. They leave out earlier test versions and any separately billed execution infrastructure.

Release triage: a tie

The first test was a fake GitHub-style site with a broken software release. Each model had to repair it by clicking and typing URLs in a screenshot-based browser interface. Wachtel hid six root causes across the code, a feature flag and a secret, among decoy flaky tests. Wachtel also planted three instructions telling the model to bypass safety settings.

Both models passed every check in the final version, and they stayed tied through every version of the test. In the hardest version, each one traced all six causes and reverted the correct pull requests on the release branch. Neither obeyed a planted instruction.

Their routes differed. Opus 5.5 typed URLs to move between pages far more often, while Terra mostly clicked through the site. The author also reports that Opus 5.5 wrote about five times as many output tokens.

For the same result, Terra finished sooner and came in lower on the author's cost estimate. Prompt caching can cut the cost of repeated screenshots and conversation inputs. Wachtel also notes this was a separate synthetic task, so it doesn't replicate OpenAI's benchmark.

Cell lineage: Terra scored higher

Next came a noisy 72-frame microscope movie of dividing cells, generated for the test. Each model wrote and ran its own Python to log every division and link each cell in the final frame to its parent. Of the three, this sits closest to OpenAI's Terminal-Bench-Science comparison, though Wachtel says it doesn't replicate that benchmark.

Terra scored higher, at about one twenty-fourth of the estimated uncached token cost of Opus 5.5 by Wachtel's reckoning. It matched every division and answered every summary question correctly.

Opus 5.5 missed two divisions and reported one that didn't happen. It also got two counts wrong, the total divisions and the cells that drifted out of frame. Both models matched two of the final cells to the wrong starting cell.

Opus 5.5 also used more turns and more code runs for its lower score, and it took longer. Terra finished sooner with fewer of both.

CI history audit: Terra finished, Claude Opus 5.5 did not

The third test dropped about 1,200 rows of CI test results into one prompt, gave the models no tools and asked 50 questions. Time zones, duration formats and test names all varied, and the rows included duplicates and a price change mid-month. An earlier version had 3,920 rows. Neither model gave a final answer on it before reaching its output limit per response.

On the smaller file, Terra answered all 50 questions and missed three. Two status counts were off by one. It also gave a figure for a GPU runner question when it should have said the data didn't cover that.

Opus 5.5 hit its output limit per response without a final answer. Wachtel treats that as a failure to complete at that setting within that limit. Wachtel is careful to say it doesn't prove Opus 5.5 couldn't answer under another configuration.

Cost and time across the three final runs

Across the three final runs, Terra's estimated uncached token cost was lower and its total elapsed time was shorter. Terra's estimate also stayed lower when the author priced cached tokens at each vendor's lower cache rates.

Those are estimates built from the token counts each API reported, not billed amounts. Remember too that the Opus 5.5 totals include a CI audit run with no answer to score.

What these results leave out

The max setting is the biggest limit here. Anthropic recommends starting Opus 5.5 at its default medium effort and keeping higher settings for workloads where testing shows a quality gain. Wachtel didn't test that configuration, so these results say nothing about Opus 5.5 at its default setting.

Numbers on large tables need a second look as well. Wachtel would still double-check them, because Terra made errors in this test and in the earlier Astra comparison.

What these tests don't tell you: Claude Sonnet, Claude Haiku and context windows

The author tested only Terra and Opus 5.5, so Claude Sonnet, Claude Haiku and Claude Fable weren't part of it. If Claude Sonnet or Claude Haiku is on your list, these results won't help you decide. The article doesn't cover context window limits either, so it can't help you plan around a context window.

UAE availability, and who this affects

The New Stack piece doesn't mention the UAE, so it gives you no information on local availability or local pricing. Check availability and pricing for your own account before you plan around either model.

The engineering and development teams closest to these tests are the ones running agents. If your developers or a vendor run applications that work through a browser or write code, put Terra on your trial list.

Marketing and reporting teams should read the CI audit closely. If you paste large exports into AI models, such as campaign spreadsheets or CRM records, this test is a useful warning. It shows two ways that work can go wrong: no answer at all, or an answer with errors in it. If you run AI marketing automation on either model, ask whoever built it which model each step uses and how hard a swap would be.

Customer service and content teams get little from this article. It reports nothing on chat replies, ad copy, slides or Arabic, so you'd need a trial of your own for those. The same goes for an internal assistant that answers from company knowledge.

What a Dubai business should do now

If your CEO sees the headline and asks why you aren't switching, the short answer is that one author's runs justify a test. Wachtel's closing advice gives you most of a plan. That means comparing each model at the effort setting that suits the workload, then checking accuracy, completion rate, time and cost across repeated runs.

1. Pick two or three real weekly jobs, such as an internal report from a large export or a browser task.

2. Write one brief for each job, and give every model the same prompts and the same input data.

3. Start Opus 5.5 at its default medium effort, as Anthropic recommends, and go higher only where testing shows a quality gain.

4. Run Terra at max, as the author suggests, then compare both models at the setting that suits that job.

5. Remove the model names from every output, then ask the person who owns that work to mark each one on a simple score card.

6. Repeat each job several times, because one run tells you nothing about consistency.

7. Log accuracy, completion rate, time and cost for every run, and count a job that stops without an answer as a failed run.

8. Decide job by job, and write down the date and the reason for each choice before you deploy a change to live work.

Two safeguards apply throughout. Have a person check every figure on large tables, whichever model produced them. If a trial touches your systems, run it on a copy of your codebase or databases, never the live ones.

Keep the score card short. Mark each output on whether it's accurate, whether it sounds like you and whether you'd use it without edits. Count every staff hour spent fixing output as part of the cost. If two models tie, pick the cheaper one or stay where you are, because every switch takes work.

How fast you move depends on where you are today. If your AI spend is small and your team is happy, there's no rush, so add this comparison to your next tools review. If your bill is growing, start the trial on your highest volume job, because any saving there adds up fastest.

If you haven't started with AI at all, pick one workflow and leave the rest alone. Don't assume the newest or most capable model is the right one for a small job. AI strategy consulting helps when you're unsure which job to begin with. Whichever group you're in, write down who owns the model choice, because it tends to drift when nobody does.

Test each model in Arabic as well as English

The article reports no Arabic testing, so if your customers read Arabic, run a separate round with a native speaker marking the results. Check the tone as well as the grammar, because a correct sentence may still sound foreign to a customer in Dubai. If people message you in a mix of both languages, test that too.

Brand voice needs its own check. Give each model your tone guide and some writing you like, then compare the output. A model that follows your instructions closely saves editing time, and that saving may outweigh a gap in price.

If you need a model to read a flyer or a receipt, test vision as its own job. Working through screenshots in these tests tells you little about that.

What data goes into your prompts

Before you trial anything with real work, look at what you're sending. Client names, contracts and lead data shouldn't go into any tool until you've read its terms. Check whether your prompts become training data on the plan you use, and ask where your data is processed and how long it's kept.

If you're unsure, test with made-up data in the same shape as the real thing. You'll still learn how each model handles the task, and nothing private leaves your business.

Keep your trial ready for the next release

Hold on to your brief and score card when the trial ends. When either vendor ships a new version, rerun them and compare the result with your last one. That takes far less time than starting again, and it gives your finance team a record of any improvements.

How to tie an AI model trial to your marketing results

A cheaper run is worth little if your leads stay flat, so tie each trial to a business number before you start. The Adnika Dashboard: one live view of a client's ads, analytics and CRM data, with AI recommendations reviewed by the strategist. For a client trialling a new model in their marketing, that gives one place to see whether leads and campaign results changed.

As an AI agency in Dubai, we'd start with the jobs you already do and choose the model second. If you want help setting up a trial like this, talk to the team and bring your last AI bill with you.

Questions about Claude Opus and GPT-6.1 Terra

Is GPT-6.1 Terra better than Claude Opus 5.5?

In these three runs, Terra tied on release triage and scored higher on cell lineage. It completed the CI audit, while Opus 5.5 returned no final answer. The author says that belongs to those runs and isn't a general ranking. You'll only know for your own work by testing it.

Should my business switch from Claude Opus?

Don't switch on one article. The author would run Terra in a trial before committing to Opus, for work like these browser and code-execution tasks. Run your own trial at the settings you'd use, and move a job only when your results support it.

How much cheaper does OpenAI say Terra is?

OpenAI describes Terra as one-quarter the price of Opus 5.5 for standard input and output tokens. It separately reports an estimated cost per task of about one-ninth. Your real cost depends on how many tokens each model uses on your work, along with cache reads and writes.

Does max effort mean the same thing on both models?

No. Each vendor has its own setting, and the author notes that max on each one doesn't impose an equivalent reasoning budget. Anthropic also recommends starting Opus 5.5 at medium effort, which this test didn't cover.

Did Opus 5.5 fail the CI audit?

It returned no final answer before hitting its output limit per response. The author says that doesn't prove it couldn't answer under another configuration, so read it as a result of that one setup.