I benchmarked 15 leading AI models on controlled project-controls tasks, and the spread between them is large enough that it should change how we talk about "using AI" in scheduling.

Each model answered the same unaided schedule-analysis problems: critical path, float, DCMA-14 logic, statusing, and earned value. No calculators. No external tools. No schedule engine access during the answer.

The important part: every answer was checked against objective, arithmetic results from a deterministic scheduling engine.

These were calculable outputs, critical path, float, dates, variances, and earned-value numbers, not subjective opinions about whether an answer "sounded right," and not what we hoped the model understood. It was model output compared directly against computed ground truth.

Takeaway 1: The model you choose materially changes reliability.

Chart 1 shows the main schedule-analysis leaderboard.

• The top score was Fable 5 at 0.95. I have confirmed that result across three internal runs using the same objectively graded battery and the result was stable across those runs.
• Among the established frontier models, the spread is also meaningful. ChatGPT-5.5 scored 0.83, while Claude Opus 4.8 scored 0.59 on the same benchmark.
• Several lighter models landed below 0.25.

That matters because "we use AI" tells you almost nothing in project controls.

Which model you use is the real decision.

Takeaway 2: The best schedule reasoner is not always the best optimizer.

Chart 2 compares unaided schedule reasoning against tool aided (single shot) resource-constrained optimization / leveling (RCPSP) performance. (This would be similar to when you use ChatGPT or Claude chat capability - Models usually code for complex tasks!)

This is where the result gets more interesting.

• Claude Opus 4.8 was mid-pack on the schedule-reasoning leaderboard, but ranked first on the resource-leveling test, hitting the proven-optimal answer.
• ChatGPT-5.5 showed the opposite pattern: very strong on reasoning, but much weaker on resource leveling.
• Fable 5 was the only model that stayed near the top across both views.

That is the practical lesson "best AI model" is not a universal label. It depends on the job.

The model I would use to draft or challenge a delay narrative is not automatically the model I would trust to reason through an optimization problem.

And even then, resource leveling is exactly the kind of work where I would still want engine-backed verification before treating the answer as reliable.

My conclusion

My takeaway is neither "AI can do schedules" nor "AI cannot do schedules."

It is this:

AI is useful for schedule-analysis support, but model choice and verification are not optional, and it comes almost always with an additional cost.

• I would use top-tier AI for first-pass reasoning, narrative drafting, explaining a DCMA flag, and initial schedule logic review (on small, clean schedules).
• I would not treat a general chatbot as the system of record for schedule, earned-value, or resource-leveling numbers without a trusted engine-backed verification.
• A float calculation, SPI(t), or leveling result can look polished in a report and still be wrong.

Simply "Using AI" is not a strategy.

Choosing the right model for the right task, knowing where it fails, and verifying the arithmetic is the strategy.

Independent research; views my own.

Zine, great research and interesting topic.

I as well agree model quality makes a huge difference, and find great results with frontier models as you also note, across many scheduling related use cases.  Especially 5.5 lately.  I just tried Fable yesterday afternoon for something unrelated to scheduling and was very impressed.  Was going to try it out for some my scheduling workflows this weekend and then poof, what a crazy last 18 hours in AI land with the Fable ban.  Hope this gets sorted quickly and concerned at the precedent that was set..

You may find my paper that will be presented at the conference in a couple weeks aligned to the path you're exploring and helpful.  If you're in Vegas for the conference, you're welcome to attend on Monday afternoon and we can chat.  The next step I wanted to take is to start A/B testing individual models and glad to see you are taking that step and sharing.

But overall, especially since the new year and the frontier models since, I am getting very promising results utilizing AI as a partner for scheduling tasks.  From building schedules, reviewing, planning, etc.  I believe there will be a paper also presented on the CDR track for using AI in claims/delays, and very interested in seeing their uses and results as well.

Thank you for sharing!

Attachment(s)

DSAA-4875 Beyond Metrics - AI Agents for Schedule Quality...pdf   999 KB 1 version