One AI task cost 20 cents or $32.08. Which are you buying?

The top spot only applies to one task

A high-ranking model sounds like an easy choice. But what does “good” mean when your staff need it to draft proposals, review code or answer customer enquiries? A test of interactive 3D scenes will not automatically answer those questions.

Artificial Analysis introduced its AA-3D benchmark on 12 October 2026. Models receive written instructions for interactive scenes and work in a closed test environment. Claude Sonnet 5.5 leads with an Elo score of 1,194. According to the source, the top four models are statistically tied. That points to strong performance in this particular test, not a general ranking for office work.

Comparisons with other benchmarks show why a single leaderboard is of limited use. FrontierCode tests whether AI agents can make changes to real open-source projects that the maintainers would accept. It also counts whether a model changes too much or works beyond the relevant scope. For a software team, that is closer to day-to-day work than a 3D test. For a medical practice or a trades business, it is still only one part of the decision.

The price gap can be huge

The cost of a model test can shift surprisingly quickly. In the AA-3D benchmark, a scene cost between 20 cents and $32.08, depending on the model. GPT-6 Luna cost 20 cents per scene and ranked eighth. Kimi K3 cost 68 times as much for a similar Elo score.

That is an eye-catching comparison, but it is not a forecast of your next monthly bill. These figures are for scenes in this benchmark. Your costs will depend on the task you set and how often you run it. Among the five highest-ranked AA-3D models, GPT-6 Sol was the cheapest at $3.46 per scene. Claude Sonnet 5.5 cost 4.4 times as much in this test, while Claude Fable 5.1 cost 9.3 times as much.

The picture changes for coding. In the FrontierCode test, GPT-6.1 Sol at maximum settings scored 49 per cent at 81 cents per task. Claude Opus 5.5 with fallback scored 52 per cent, but cost more than 12 times as much per task. Three percentage points may be valuable for a sensitive change. Across thousands of routine tasks, the extra cost may be harder to justify. Your work, not the price table, determines which option matters.

Open weights do not answer every question

A model you can run or adapt yourself sounds like it gives you more control. With open-weight models, businesses can access the trained model weights. That can open up options that are not the same with an online-only service. But it does not tell you whether the model can handle your tasks well enough or what conditions apply to your use.

In the AA-3D test, Kimi K3, with 982 Elo, and MiMo-V2.6-Pro, with 977 Elo, were the highest-ranked open-weight models, in 11th and 12th place. In the FrontierCode test, MiMo-V2.6-Pro led the open-weight models at 45 per cent. DeepSeek V4.1 Flash and Kimi K3 each scored 44 per cent. These are useful points of comparison, but being in an open-weight category is no guarantee about how a model will work in your own business.

Muse Spark Open was released as an open-weight model by Meta Superintelligence Labs on 15 October 2026. Its licence allows use, modification, distribution and sale under certain conditions. These include visibly labelling products or services that use the model or its derivatives with “Built with Muse Spark Open”. For Model-as-a-Service offerings, the licence sets out further obligations to Meta if aggregate gross revenue exceeds $20 million in any rolling 12-month period.

Pay attention to model versions too. The AA-3D results place Muse Spark 1.3 seventh. That is not automatically a test result for Muse Spark Open. The names are similar, but the versions tested are not interchangeable. For a procurement discussion, check the exact version and its licence, not just the family resemblance.

A strong result is not a reliable process

AI can shine in a test and still fail on the next real task. FrontierCode illustrates the gap: none of the models passed nine out of every 100 tasks in any attempt. Not a single task was passed by every model on every attempt. A high average score does not mean every individual change will work safely.

For a small or medium-sized business, that means an AI suggestion for a change to your stock management system should not go straight into live use without checks. For a sole trader, the same principle applies to a customer proposal. Check that the figures add up, that the services promised are actually available, and that personal data does not belong in the prompt. Responsibility does not disappear because a machine writes fluently.

OpenAI also describes Intelligent UI as a development in which answers can include graphics, buttons, forms or interactive content. That could change how tasks work, as a customer may be able to take the next step within an answer. It also raises practical questions: who checks the content, who takes responsibility if a process goes wrong, and where does the information go afterwards? Include these questions in your test before mistaking a polished prototype for a finished workflow.

What it means for your business

Pick a recurring task from your calendar. It could be the first reply to a common customer question, a summary of an internal conversation or a product description draft. Decide in advance what a useful result looks like. For a customer reply, that might mean accurate facts, an appropriate tone and a clear next step.

Test two or three models against the same examples. Use content you are allowed to share under your data protection rules and internal policies. Record how often someone has to make substantial edits, how long checking takes and what the test actually costs. In a larger business, the relevant teams, IT and data protection should agree which information can be used. In a small business, you may take on those roles yourself, but you should still write down the criteria.

In the FrontierCode test, the cost ranged from nine cents per task for GPT-6 Luna to $12 for Claude Fable 5.1. Those figures are not a general price list for your use. They show why a pilot should track both cost per completed task and quality. Factor in the time spent on corrections too. A cheaper model can cost more in practice if someone has to rewrite every other sentence.

Decide when a person should take over as well. For an unclear request, an unusual customer case or a decision with significant consequences, your team should know where to escalate it. That is what makes next Monday look different: a colleague opens the draft, uses the same checks to see what works, and knows when to write the reply themselves. That routine determines whether a model makes the working day easier.

• Choose a clearly defined use case.
• Test several models on identical tasks.
• Assess quality, editing time, cost and licence together.
• Assign an owner and set a clear handover point.

The next model change is coming anyway

In autumn 2026, choosing the right model is a question for more than developer teams. OpenAI announced a new GPT-6 model for ChatGPT on 7 October, describing it as an offering for more than 1.2 billion weekly users. At DevDay, the company also outlined a shared interface for people and agents, through which developers could make their own applications available.

For businesses, that mainly means more choice can bring more work to coordinate. Reassessing every workflow each time a new name appears takes time away from customer service, the workshop or sales. A repeatable assessment process is more useful. You can test a new version against the same tasks and criteria instead of being led by a leaderboard.

The first step remains the same: decide which task in your business needs reliable support.

Frequently asked questions

More articles

• Use casesMultilingual guests, one wrong answer: the AI trap for tourism
• Use casesAI sends visitors. Your website loses the lead
• Law & securityYour team is using secret AI. Your customer data may go with it