---
title: "Which AI model plans a project best? 14 models, three projects and one change halfway through"
excerpt: "14 AI models planned the same projects and got a change halfway through. Three deliver plans you can build on. But how well a model plans says little about how well it replans."
publishedAt: "2026-11-14"
updatedAt: "2026-11-15"
readTime: "14 min read"
author: "Simon Schwer"
tags:
  [
    "ProjectBench",
    "Benchmark",
    "AI Project Management",
    "AI Models",
    "Scheduling",
    "Project Management",
    "Context-Driven Project Management",
  ]
featuredImage: "/images/blog/projectbench-one-change-half-the-plan-hero.webp"
featuredImageAlt: "Illustration: a message hits a plan made of task tiles. Half the tiles have slipped and left dashed outlines. On the right are four panels for tasks, milestones, project end date and decision log; one of them has stayed grey."
featuredImagePosition: "center"
sources:
  - title: "Model list with prices per million tokens (retrieved 15 November 2026)"
    url: "https://openrouter.ai/models"
    publisher: "OpenRouter"
---

The shell contractor will start three weeks late. That is a single sentence, and anyone who keeps a project plan knows what follows: move the dates, check the milestones, reset the end date and write down who decided it.

You can hand that work to an AI now. The plan then depends on the language model behind it, and no spec sheet says how well a given model does project work. So we had 14 current models plan the same projects and told each of them about a change like that halfway through.

We build TensorPM, software in which an AI assistant keeps project plans and where you choose the model. Two AI models scored the plans without knowing who had made them. By their scores, three of the 14 models deliver plans you can build on, and the most expensive of them is only just ahead of one that costs a fifth as much. How well a model draws up a plan, however, says little about how well it adjusts one. One of the three applied the three weeks to a single task.

## What did the models have to do?

There were three projects, each given as a four-sentence brief of the kind an owner or a project lead would write. All 14 models worked on the house and the cloud migration, though one run does not count (more on that below). Nine worked on the housing project. We did not run five models there, to limit the cost of the test.

| Project | Frame | The change halfway through |
| --- | --- | --- |
| Single-family house near Fürth, Germany | 150 m², EUR 480,000 cap, move-in after 14 months | The bank lends EUR 25,000 less and the shell contractor starts three weeks late |
| Moving an order-management application to the cloud | 200 users, EUR 175,000, cutover on one weekend, six hours of downtime at most | A load test shows response times that are too slow, and the dedicated network link that fixes it arrives after the cutover weekend |
| Housing project in Augsburg, Germany | 96 flats in three buildings, EUR 29.5 million construction cost, staged handover until mid-2029 | The shell contractor is insolvent, the next bid is 11% higher, and the bank wants more pre-sales |

The procedure was the same for every model. It creates the project from the brief, with goal, scope, milestones and risks. It enters the people involved, derives tasks with dates, budgets and owners, and links them. Then the owner reports the change in the chat and asks for the plan to be adjusted. After that the model writes a status report for a named reader: the owner's partner for the house, the bank and the investors for the housing project. Last, an email with new facts comes in, and the model turns it into proposals for the plan. Questions from the models are answered by a fixed AI model in the role of the owner, briefly and without new requirements.

All models worked in the same software with the same instructions, and none was told it was being tested. TensorPM is the test bench here. Our test program, which we call ProjectBench, reads out the complete plan after every step, so what changed can be counted.

Two AI models acted as jurors and did not know which model had made which plan. Their scores produce a plan value between 0 and 100. From about 78 a plan is something you can build on, between 62 and 78 it is usable with gaps, and below that it is a rough draft someone has to rework. How the jury works and where it disagreed comes further down.

## Which model plans best?

The ranking rests on the house and the cloud migration, the two projects every model worked on. The housing project is shown next to it.

![Bar chart: plan value per model as the mean of the single-family house and the cloud migration. GPT-6 Astra 84.1, Kimi K3.1 Pro 81.0, GPT-6.1 Sol 79.8, Grok 4.7 72.3, Qwen 3.8 Max 69.2, Mistral Large 4 65.7, GPT-6.1 Terra 65.5, DeepSeek V4.1 Flash 64.1, Claude Sonnet 5.5 61.5, Gemini 4 Hydrogen 60.4, Claude Haiku 5.5 57.5, GPT-6 Luna 55.0, MiniMax M3.1 Flash 51.5, GLM 5.3 44.0. Dots show the single projects.](/images/blog/projectbench-nov26-ranking-en.svg)

| Rank | Model | Vendor | Plan value | House | Cloud | Housing project | Cost per plan |
| --- | --- | --- | --- | --- | --- | --- | --- |
| 1 | GPT-6 Astra | OpenAI | 84.1 | 83.5 | 84.7 | 76.3 | about $15 |
| 2 | Kimi K3.1 Pro | Moonshot AI | 81.0 | 76.4 | 85.6 | | $3.19 |
| 3 | GPT-6.1 Sol | OpenAI | 79.8 | 74.8 | 84.8 | 81.4 | $2.96 |
| 4 | Grok 4.7 | xAI | 72.3 | 68.7 | 76.0 | | $1.79 |
| 5 | Qwen 3.8 Max | Alibaba | 69.2 | 70.1 | 68.3 | | $4.16 |
| 6 | Mistral Large 4 | Mistral | 65.7 | 63.0 | 68.4 | 62.1 | $0.77 |
| 7 | GPT-6.1 Terra | OpenAI | 65.5 | 58.3 | 72.7 | 64.8 | $0.80 |
| 8 | DeepSeek V4.1 Flash | DeepSeek | 64.1 | 68.4 | 59.8 | 27.7 | $0.07 |
| 9 | Claude Sonnet 5.5 | Anthropic | 61.5 (house only) | 61.5 | not scored | | $1.80 |
| 10 | Gemini 4 Hydrogen | Google | 60.4 | 57.2 | 63.5 | | $0.74 |
| 11 | Claude Haiku 5.5 | Anthropic | 57.5 | 54.6 | 60.5 | 47.9 | $0.20 |
| 12 | GPT-6 Luna | OpenAI | 55.0 | 49.1 | 61.0 | 51.7 | $0.07 |
| 13 | MiniMax M3.1 Flash | MiniMax | 51.5 | 45.0 | 58.0 | 58.4 | $0.27 |
| 14 | GLM 5.3 | Z.ai | 44.0 | 54.8 | 33.2 | 38.1 | $0.34 |

Three models deliver plans you can build on: GPT-6 Astra, Kimi K3.1 Pro and GPT-6.1 Sol. Both jurors put them in front, and the gap between the three is smaller than the gap between the jurors. Behind them is a wide middle group from Grok to DeepSeek whose plans are usable and have gaps. The six models after that deliver a rough draft or less.

Claude Sonnet 5.5 is listed with the house only. On the second project our credit with the service we use to reach the models did not last for the whole run. So we do not count it.

GLM 5.3 is last because the project it created stayed almost empty in the cloud migration and in the housing project: a description, but no milestones, no risks and no time frame. The tasks it derived were passable, but they then lack the frame they would be measured against.

## What does a plan cost?

The right-hand column gives what one complete run costs through the vendor's API, which is tokens used (the text units vendors bill by) times list price, averaged over the house and the cloud migration. Price plays no part in the ranking.

![Scatter plot: plan value against cost per plan on a logarithmic axis. DeepSeek V4.1 Flash and GPT-6 Luna at 7 cents, Kimi K3.1 Pro at $3.19 with a plan value of 81, GPT-6.1 Sol at $2.96 with 79.8, GPT-6 Astra estimated at about $15 with 84.1.](/images/blog/projectbench-nov26-cost-en.svg)

Between DeepSeek V4.1 Flash and GPT-6 Astra the price differs by a factor of 200 and the plan value by 20 points. Astra would cost about $15 per plan and leads Kimi K3.1 Pro, which plans for a little over $3, by three points. Mistral Large 4 and GPT-6.1 Terra stay under a dollar and sit 15 points behind Kimi. DeepSeek gets almost the same plan value as those two for seven cents.

The figure for Astra is an estimate. That model ran on a subscription that reports no token counts. We used its list price and the tokens GPT-6.1 Sol needed for the same project.

Run time varies as well. For the house, Gemini 4 Hydrogen took four minutes and Qwen 3.8 Max took 53. Most models needed between five and 15 minutes.

## What happens when the contractor starts three weeks late?

Back to the sentence from the beginning. In the house project the owner writes to the assistant, in essence:

> The bank confirmed financing, but EUR 25,000 lower than planned, so the interior-finishes budget is squeezed. At the same time the shell contractor moved the construction start three weeks later because their previous site overran. I still need to stay within the original budget cap. Can you update the plan for this?

This is what it looks like in one of the 14 plans, the one by Gemini 4 Hydrogen. It has 22 tasks. Design and permitting stay where they are. Everything that hangs on the construction start moves with it.

![Schedule of one plan before and after the change: 9 early tasks unchanged, 13 tasks now finish about three weeks later, their earlier dates drawn as thin lines.](/images/blog/projectbench-nov26-change-example-en.svg)

Across all 14 plans, 243 of 630 tasks got a new date. The median per plan is 50%. If you change only the line that says “shell” after a message like this, you very likely end up with a plan full of dates that no longer hold.

![Column chart: share of tasks with a new date per plan, sorted ascending. 0 percent in 2 plans, median 50 percent, maximum 64 percent.](/images/blog/projectbench-nov26-change-share-en.svg)

Task dates are one of four places where the three weeks have to show up. The others are the milestone list, the project end date and the decision log. Every change in the respective place is counted, not whether it is right.

![Bar chart: decision logged in 12 of 14 plans, task dates moved in 12, project end date changed in 5, milestones changed in 12, all four places in 5.](/images/blog/projectbench-nov26-change-landing-en.svg)

The change reached all four places in the plans by Astra, Kimi, Qwen, Grok and Gemini. Five of 14 does not mean nine plans are wrong, because a plan may hold the end date if it recovers the three weeks somewhere else. Sonnet, Haiku, Mistral and DeepSeek held the end date and compressed later work to do so, which the jurors partly rated as unverified. In the plans by GLM and MiniMax, however, move-in now falls three weeks after the project end date, which still stands unchanged.

The two plans without a single moved date come from GPT-6.1 Terra and GPT-6 Luna. Terra rewrote 24 task descriptions along the lines of “existing dates stand until the delay is confirmed” and noted the delay as a risk. Luna logged two decisions that name a construction start in autumn 2026, which neither the brief nor the message mentions and which contradicts its own plan. For the replanning the jury gave 35.5 and 21.5 points out of 100.

With the replanning taken out of the score, GPT-6.1 Sol had the best-scored plan for the house. It cut the interior budget by EUR 25,000, but wrote the delay only into the start of a single task and renamed the milestones “unverified targets”. Gemini 4 Hydrogen went into the change with one of the thinnest first plans and carried it through all four places.

Qwen 3.8 Max gives an example of a largely consistent adjustment: exactly EUR 25,000 off the interior tasks, 36 tasks moved, and the new move-in date entered the same way everywhere. The one week it recovers along the way, however, comes from overlaps that break existing dependencies. The jurors gave 90 and 64 points for it.

## What do models do when the change gives no new date?

The message in the cloud migration leaves something out. It says the required network link arrives after the planned weekend, but not when. No new date follows from it.

Many models saw that. They put the cutover on hold or marked it as conditional, and some of them created tasks to obtain a firm delivery date. Kimi K3.1 Pro did so most completely: 21 tasks blocked, all later milestones marked as suspended, three new linked tasks, and no replacement date assumed for the cutover. Only the due dates of two of the three new tasks are placeholders. That earned 86 points, the highest score for any replanning in the comparison.

Claude Haiku 5.5 went the other way. It moved the cutover to the weekend around 31 January and carried 13 tasks and six milestones along. The plan looks tidy. It rests on a delivery date for the link that the model assumed itself. GPT-6 Astra also names a new date in its reply that depends on an assumed activation, but keeps it open in the plan.

The same pattern appears in the housing project. Mistral Large 4, MiniMax M3.1 Flash and DeepSeek V4.1 Flash pushed the shell work back by months that nobody had stated, and Astra entered two provisional start dates for buildings B and C that nobody had named either. A plan with invented dates is harder to spot than a plan in which nothing happened, because it looks complete.

A third group held the date and built a detour, for example an interim solution without the new link. This is where the jurors agreed least. One read it as careful planning, the other as contradicting the message, which calls the link necessary. For Qwen the two scores are 35 points apart.

## What changes when the project gets big?

The housing project is the hardest of the three projects: three buildings, 16 people, three years, and a change that hits cost, schedule and financing at once.

The gaps widen. GPT-6.1 Sol reaches 81.4 and GPT-6 Astra 76.3, each with 100 tasks. The best of the cheaper models, GPT-6.1 Terra, is almost 17 points behind at 64.8. DeepSeek V4.1 Flash drops to 27.7. Like GLM, it created an almost empty project here. Whether that is down to size cannot be said from a single run.

Hardly any model gets the replanning right here. Sol scores 78.5, all the others between 36 and 55. Sol worked the 11% extra cost into all five budgets for shell and garage, kept the total under the cap, logged a decision and scheduled four tasks for the re-award as predecessors of the awards, without inventing a date.

Astra, with the second-best plan of the project, scores 51 for replanning. It rewrote 34 task texts but put no number on the extra cost, and its reply reports changed links that the stored plan does not contain. In the plans by Sol, Haiku, MiniMax and GLM, tasks for the shell or the re-award are still assigned to the insolvent contractor. In GLM's plan the shell sum also appears twice after the change, so the tasks add up to EUR 44 million against a cost target of EUR 29.5 million.

## Where do good models struggle too?

Some weaknesses turn up in all projects.

On the budget the models are far apart. In the house, Claude Sonnet 5.5 and Grok 4.7 put only 9% of the EUR 480,000 on tasks, GPT-6 Luna less than 1%. In the cloud migration Luna planned 135% of the budget, GPT-6.1 Terra and GLM 5.3 nothing at all.

Links and dates often do not fit together. Already in the first plan for the house, 51 successors in Kimi's plan and 49 in MiniMax's started before their predecessor ended, and the replanning cleared that up only in part (37 and 44). In the plans by Haiku and Sonnet there was no such contradiction, neither before nor after.

What a model says about its own work is not always true. Several replies report a cut or a link that the stored plan does not contain. So after an AI has changed a plan, check the plan and not the reply.

## What can you check yourself?

The four places from the house project apply to any plan, in construction as in IT or plant engineering, and whether an AI keeps it or a person. The next time a date slips on your project, look a week later whether it has arrived everywhere:

1. Are the successors on the new date?
2. Are the milestones in the last status report still right?
3. Has the end date been adjusted or deliberately held?
4. Is it written down who decided that, and when?

If an AI adjusted the plan, there is a fifth question: where does each new date come from? If it appears in no message and no minutes, the model assumed it.

For choosing a model this means: Kimi K3.1 Pro and GPT-6.1 Sol deliver results on a par with the most expensive model for about three dollars per plan, Sol with its weak replanning in the house project, though. Test a model on a change, because the first plan tells you little about that. Give it a project you know, and then the change you received most recently.

How to turn the four questions into a weekly routine on a construction project is covered in [Spotting schedule delays in construction projects early](/blog/terminverzug-bauprojekt-frueh-erkennen). What happens when the plan trails reality is described in [What Is the Biggest Problem in Project Management That Nobody Talks About?](/blog/stale-context-silent-killer).

## Has anything changed since the summer?

In June we had the same house project planned once before, with the same instructions, by 15 models of the previous generation, among them GPT-5.5, Claude Opus 4.8 and Gemini 3.1 Pro. The two jurors have now scored those stored plans again by the same standard.

| House project | June 2026 | November 2026 |
| --- | --- | --- |
| Plans | 15 | 14 |
| Best plan value | 69.3 (GPT-5.5) | 83.5 (GPT-6 Astra) |
| Median | 48.2 | 62.3 |
| Plans at 62 points or more | 4 | 7 |
| Plans below 45 points | 6 | 0 |
| Median replanning | 36 | 63 |
| Median schedule | 35 | 54 |
| Median risks | 70 | 71 |

Most of the gain is in replanning and in the schedule. Naming risks and setting up a project is something the models could already do in June.

The comparison has limits. Since June both the models and TensorPM have changed, so the difference cannot be put down to the models alone. And four of the June runs broke off in one of the steps, which pulls the median down.

The June scores also differ from those we published in the summer in [our first ranking](/blog/projectbench-plan-value). Back then a provisional jury of a single model did the scoring, and it was far more lenient than today's two jurors.

## How far can you trust these numbers?

Each model worked on a project only once. That is the main limitation. We had to repeat a number of runs, eight of them because of a gap in the measurement, and where repeat and first attempt could be compared, they looked different. Kimi, for instance, put the budget on categories once and on single tasks the other time. Differences of a few points therefore mean nothing. What holds are the groups: the three in front, the middle, the rough drafts behind.

The jury consists of two models, Claude Fable 5.2 by Anthropic and GPT-6 Astra by OpenAI. Both scored every plan on seven criteria without knowing who made it. In the plan value, replanning counts 28%, schedule and tasks 20% each, risks 17% and project setup 15%. The status report and the proposals from the incoming email do not count towards the plan value. A plan that exceeds the budget by more than 15% gets 77 points at most, and 61 at most beyond 35%. That applies to GPT-6 Luna in the cloud migration.

![Scatter plot: plan value by Claude Fable 5.2 against plan value by GPT-6 Astra for 36 plans. Most points lie above the diagonal, so Fable usually scores higher.](/images/blog/projectbench-nov26-jurors-en.svg)

The two jurors largely agree on the order, with a rank correlation (1 would mean the same order) of 0.86 for the house, 0.80 for the cloud migration and 0.95 for the housing project. Astra is the stricter juror and sits just under seven points below Fable on average.

Astra is also a contestant, and on OpenAI plans its verdict stands out. Astra scores plans by OpenAI models only 1.9 points lower than Fable on average, and plans by all other vendors 9.3 points lower. Whether Astra is lenient with its own family or Fable is harder on it cannot be settled with two jurors.

At the top it decides the places: Fable alone puts Kimi ahead of Astra and Sol, Astra alone puts Astra ahead of Sol and Kimi. The same three models lead either way. On the four plans by the two Claude models Fable is 7.6 points above Astra, on the plans by vendors other than OpenAI and Anthropic 9.7. No preference for its own family shows there, but four plans are too few for a verdict.

## What did the measurement miss?

We found gaps in the measurement along the way. They affect single scores.

- The software lets a model put budgets on categories, that is on groups of tasks, instead of on every single task. For the first hours our measurement did not record those amounts. We repeated eight runs because of that. For five plans of the house it could no longer be established whether category budgets had been set. There the jury neither credited such figures nor counted them as missing. Two of those plans, by GLM and MiniMax, may come off too badly in replanning as a result.
- Some models deliver the status report as a PDF. At first we did not keep its content. Those reports are not scored.
- In the cloud migration the incoming email only repeats what the brief already says. Nearly all models rightly propose nothing. That score is left out of the analysis there.
- The briefs name fixed calendar dates that were already in the past when we ran the test. Some models planned from the old start date, some from the day of the run. The jury accepted both.
- Claude Sonnet 5.5 rejects a technical setting the software uses to force structured answers. We built a fallback for that case which leaves the other models untouched.
- GPT-6 Astra ran on a subscription everywhere, GPT-6.1 Sol did so in the housing project, and all other runs went through the OpenRouter service. On the subscription route the software asks for the answer format by instruction instead of by a technical setting, and short helper calls run on GPT-6 Luna. The two top places in the housing project and first place overall were therefore earned under slightly different conditions.
- Every model ran with the software's default reasoning setting, which is low. With more reasoning the results may differ.
- In one case the simulated owner stated something that is not in the brief.

Faults that lay in our measurement were fixed or taken out of the scoring. They are listed here so you can judge what the numbers will bear.

## What do we take from this for TensorPM?

We measure in order to decide which models we recommend and where the software should take work off the model.

After a project is created, TensorPM should check whether time frame, budget, milestones and risks are filled in and ask for what is missing. GLM and DeepSeek showed that this cannot be taken for granted.

A milestone needs a state for “date open”. The models that correctly paused the cloud cutover had to leave the old date in place and write “suspended” into the description.

The fallback for models that reject forced answer formats shipped in version 1.46.

The differences could be counted because a plan in TensorPM consists of structured data that can be read out and compared after every step. In the Gantt chart you have been able to save snapshots of the plan, that is baselines, and compare them since version 1.40. The comparison shows for every task how many working days it has moved since the saved snapshot, and so also when nothing has moved. If a date contradicts its dependency, the app offers to align the successors and shows beforehand what will move.

In the end the contractor's three weeks stood in all four places in five of the 14 plans. For any other plan, the four questions further up are the quickest way to find out.
