
On 13 November I called a winner in Copilot Cowork: Opus 5.5 edges Fable 5.2. That head-to-head was Opus 5.5 at 8.9 against Fable 5.2 at 8.6. I said then that I had not re-run Astra, so it was not a new overall series winner, and that I would test the next model when it turned up.
The next round is in. I have now put two new GPT 6.2 reports through the same kind of evaluation, alongside the Opus 5.5 report from September. The full evaluation is here on GitHub. The winner is not Opus 5.5 this time.
What was compared
The evaluation assessed three reports, all responding to the same underlying prompt: is the traditional MSP and TSP business model becoming obsolete between 2029 and 2036? The three documents are:
- Document A: the GPT 6.2 Astra report from Cowork, dated 18 November 2026, written as an Australian owner and investor decision report for CIAOPS.
- Document B: the Opus 5.5 report from Cowork, dated 23 September 2026, a strategic research report for MSP owners, Microsoft partners, boards and private equity investors.
- Document C: the GPT 6.2 Sol report from Cowork, dated 18 November 2026, a board and investment briefing for CIAOPS.
Scoring uses four weighted criteria: Argument 35%, Evidence 25%, Detail 20% and Presentation 20%. The evaluation is relative and substance-focused.
The result
| Rank | Document | Overall | Argument | Evidence | Detail | Presentation |
|---|---|---|---|---|---|---|
| 1 | A: Astra / GPT 6.2 | 8.93 | 9.2 | 9.0 | 8.5 | 8.8 |
| 2 | C: Sol / GPT 6.2 | 8.78 | 9.0 | 8.9 | 8.7 | 8.5 |
| 3 | B: Opus 5.5 | 8.69 | 8.5 | 8.7 | 9.2 | 8.5 |
Document A, the GPT 6.2 Astra report, comes first. The evaluation calls it the tightest on epistemic discipline, with a fully reconciled decision model and the most practical capital gates for an owner or investor audience. The Sol report is a very close second. Opus 5.5 is third, but it still wins on Detail, 9.2 against 8.5 and 8.7, because it has the broadest coverage of market, security and history.
The evaluation says A and C are both clearly ahead of B on the two heaviest criteria, Argument and Evidence. Its stated reason is that Opus 5.5 was more assertive on magnitude and probability claims than its evidence strictly supports, including the 30 to 60 per cent labour-removal figure. It also says the earlier research cut-off meant Opus 5.5 missed some later November Microsoft detail that A and C used.
Where Astra and Sol differ
The gap between first and second is small. According to the evaluation, A is tighter on capital gates, inference language and the executive reading path. C is stronger on cash-trough quantification, pricing-headroom sensitivity and residual multi-vendor service, and uses a larger reference firm. The file notes that C’s title says the traditional model is becoming obsolete, while its body concludes it remains viable. It lists that as the only notable polish issue, and says C would pass A with a title that matches the body and tighter tranche limits.
How this compares with 13 November
In Opus 5.5 edges Fable 5.2, Opus 5.5 beat Fable 5.2 on evidence (9.2 against 8.3) and detail (9.3 against 8.4), while Fable 5.2 was the more polished read. The same pattern shows up again here. Opus 5.5 is still the most detailed report and still loses points for how tightly it handles claims. What has changed is the competition: two GPT 6.2 reports, Astra and Sol, are now ahead of it on overall score.
A few cautions. This is a different evaluation from the one in the 13 November post, so the Opus 5.5 score of 8.69 here should not be set against the 8.9 there as if it were a drop. Fable 5.2 was not part of this evaluation, so I cannot tell you where it would land among these three. The file does not include credit costs, so there is no price comparison here. And this is one prompt on one topic, so it is a data point, not a verdict on every task.
I have not checked which of these models are available, or in which Microsoft 365 Copilot surfaces, for your tenant. This post says nothing about availability. Check Microsoft’s official documentation for that.
What it means for choosing models in Copilot
For an MSP, I think the useful lesson is about what you are choosing for. The evaluation rewarded the reports that were careful about what they could and could not claim, and that turned analysis into reversible, bounded decisions. That is the same quality I want in a client-facing report. A longer, broader report lost to tighter ones on the criteria that carried the most weight.
So I would not pick a model on the scoreboard alone. Run your own representative prompt, such as a client QBR, a security review summary or a pricing paper. Then judge the outputs on argument and evidence first and detail second. If you need breadth, Opus 5.5 still earned the best Detail score here. If you need something a board or an owner can act on, the GPT 6.2 reports did better in this test.
Model names matter here: Astra and Sol are both GPT 6.2, and Opus 5.5 is a different model again. Keep a record of exactly which model produced each output when you compare, because they change quickly and the winner changes with them.
As soon as the next model shows up in Cowork, I will run it through the same test.