Qwen 3.8 Max vs Grok 4.7
for project risk registers

The bottom line
Grok 4.7 leads by a small margin

Grok 4.7 has a small edge at finding risks tied to the plan itself and a tentative one at holding its ranking when challenged. Qwen 3.8 Max is the stronger pick when the plan pack is very large or long-context cost matters.

The evidence is indirect. On AA-Briefcase, a test that checks whether a model finds requirements hidden in source files and uses the correct evidence, Grok 4.7 scored 1,644 against 1,621 for Qwen 3.8 Max¹. It also led 1,715 to 1,671 on GDPval-AA¹. Neither is a project risk test, and both gaps are small.

For ranking under challenge the lean is weaker. No public test measures whether either model keeps a justified likelihood and impact ranking after someone pushes back. xAI says Grok 4.7 was trained to verify its work more carefully². That is a vendor claim, so treat it as something to test.

A staged workflow follows from this. For a normal plan, let Grok find, score, challenge and finish the register. For a very large plan pack, let Qwen pull out plan facts and candidate risks, then let Grok challenge the shortlist and the final ranking. This split is a judgment call and not a measured rule.

Who this is for
Which delivery roles this fits

What we compared
The models through their API

This page compares the two models through their API in one neutral setup, not one model inside one app against the other inside another.

A risk register is a ranked list of things that could go wrong with a project. Each entry says how likely the problem is, how much damage it would do and what to do about it. The parts that matter here are finding risks that come from facts in the plan, scoring likelihood and impact the same way each time, ranking the results, writing mitigations that really change the exposure, and holding up when someone challenges the order. Official docs come first, then the Artificial Analysis comparison of the two exact models and the methods behind its benchmarks^1, 10.

We left tools out of the specs table on purpose. File upload, spreadsheet export and similar features belong to the app around the model, so the same model can behave very differently in a chat product, in the API or inside a workspace. Judging those here would compare wrappers and say little about risk work. One more detail: Alibaba's unversioned qwen3.8-max name has pointed to the September 2, 2026 qwen3.8-max-0902 snapshot since September 5, 2026, and the benchmark figures here are for that snapshot³.

Specs at a glance
Context and price for long plans

The model facts that shape a risk register job. Tool and file features are left out, since those depend on the app around the model.

Head to head
Small edges that depend on plan size

The answer changes by job and by plan size. This is the main analysis: which model has the edge on each part of building a register, and what backs it up.

How to test
A fair test on your own plans

A useful test is repetitive. Same plan, same prompt, same reasoning level, same output format, then judge each register on what a reviewer would do with it.

What the evidence shows
Close scores and no direct test

No public benchmark gives both exact models the same project plans, so the evidence is a mix of related tests and vendor statements. Here is what each source helps judge.

How to prompt each one
Different shapes for each model

Both prompts ask for evidence behind every risk and a fixed scoring rubric. They differ in how they handle plan size, reasoning effort and the challenge pass.

Qwen 3.8 Max can take a very large plan, so tell it to extract the plan facts first and then name only the risks those facts cause. Independent testing describes it as unusually verbose, so cap the output and ask for structured JSON, which cuts commentary and unstable columns^14, 8. Use medium effort for routine extraction and xhigh for the final challenge pass, since xhigh is its documented default⁵.

Grok 4.7 works well when the adversarial review is spelled out. Ask it to argue that each top risk is overrated or underrated, then publish the final ranking with a short change log. Use high effort for the first pass and xhigh for disputed, high-value decisions, and pair it with a schema so the register comes back in fixed fields^7, 13, 16.

A Qwen 3.8 Max prompt: plan facts first and a JSON register

Read the plan in full. Extract plan facts first, then identify
only risks caused by those facts.

Return JSON with: risk, evidence_quote_or_section, cause,
consequence, likelihood_1_to_5, impact_1_to_5, score, mitigation,
trigger and owner.

Exclude generic risks unless the plan contains evidence for them.
Sort by score. Then challenge the top five and revise only when
the evidence warrants it.

A Grok 4.7 prompt: a skeptical review board pass

Produce a ranked risk register from this plan. Every risk must
cite a plan-specific fact. Apply the supplied likelihood and
impact rubric exactly.

After the first ranking, act as a skeptical review board. State
the strongest argument that each top risk is overrated or
underrated, then publish the final ranking with a short change log.

Return the specified JSON schema only.

Weak spots
Where each model needs guardrails

Neither model is safe to trust blindly on a risk register. The useful question is where each one adds cleanup work, and what to change in the prompt or the workflow.

Which one to choose
Start from the size of the plan

One question first: how many tokens is the full plan pack. Then follow the branch that matches how big it is and how much is at stake.

Recommendations
Pick by plan size and stakes

If the whole plan pack fits below 200,000 tokens and finding risks that belong to the plan is the priority, start with Grok 4.7^1, 6. If you also need a ranking you can defend, run separate generation and challenge passes.

If the pack falls between 200,000 and 500,000 tokens, compare the total cost. Choose Qwen 3.8 Max when you expect repeated review passes or cached input, since its cached rate is $0.25 per million tokens against $0.50 for Grok and Grok's token rates double from 200,000 tokens^4, 6. Otherwise pilot both. If the pack passes 500,000 tokens, choose Qwen 3.8 Max or cut the plan down first.

If the register feeds software, either model works with a strict schema and validation^8, 13. If the domain is high-stakes or highly specialised, make no automatic choice. Run a blind test with domain experts and require human approval.

If you only ever run one of these models through its own API and review plans alone, Playgram is not the right buy. Its value is in comparing models and sharing context across a team.

Bottom line
Grok suits most pasted plans

Grok 4.7 is the better default for a normal pasted plan, with a small and indirect lead on source-grounded knowledge work. Qwen 3.8 Max is the better choice for very large plan packs and long-context economics.

The limits are real. Benchmarks use different harnesses and scoring, vendor claims are not independent evidence, and no public test runs both models on project risk registers or on defending a ranking. Grok's edge on ranking under challenge is an inference that nobody has measured. Model aliases can change and prices move quickly, so recheck both before you commit³.

The safest final step is to test the shape of your own plans with your own scoring rules. A fair test needs the same setup for both models: the same plan text, the same prompt and the same place to run them, so the result reflects the models themselves. In practice that is harder than it sounds, since most teams end up running one model in one app and the other in a different one, on two separate subscriptions, which tilts the comparison before the first register comes back. The cleaner the setup, the more the difference you see is really Qwen 3.8 Max vs Grok 4.7, and which one happened to be easier to reach that day matters less.

Frequently asked
questions

Related comparisons