I've added Apple Private Cloud Compute to Cumbersome. You can now choose Apple alongside your other AI providers without adding an API key. I wanted to know how it compares with Gemini, especially Flash-Lite.
PCC got twelve more answers right than Flash-Lite. All five services received the same 72 questions, with moderate reasoning for PCC and medium reasoning for Gemini.
PCC beat Flash-Lite, mostly on arithmetic
I gave each model 72 questions from BIG-Bench-Hard (opens in new tab): twelve each on logic, dates, arithmetic, and three other reasoning tasks. I picked the questions before running the models and checked their responses against the published answers.
PCC got 58 right. Gemini 3.5 Flash-Lite got 46. Arithmetic accounted for eight of those twelve extra correct answers: PCC scored 11/12, while Flash-Lite scored 3/12. Flash-Lite did better on Boolean expressions, getting all twelve right where PCC missed two.
| Task, twelve questions each | Apple PCC | Flash-Lite 3.5 | Pro 3.1 | Flash 3.8 | Argon 4 |
|---|---|---|---|---|---|
Boolean expressions | 10 | 12 | 12 | 12 | 12 |
Date calculations | 11 | 11 | 12 | 12 | 11 |
Formal fallacies | 9 | 8 | 12 | 12 | 12 |
Ordering seven objects from clues | 10 | 8 | 12 | 12 | 12 |
Arithmetic with several operations | 11 | 3 | 12 | 12 | 12 |
Tracking seven shuffled objects | 7 | 4 | 12 | 12 | 12 |
Total | 58 | 46 | 72 | 72 | 71 |
PCC struggled most with tracking objects through a series of swaps, missing five questions. Flash-Lite missed eight. Gemini 3.1 Pro Preview and Gemini 3.8 Flash got everything right; Gemini 4 Argon missed one date question. This set separates PCC from those three, but it's too easy for them to tell me much about their differences.
Apple's cloud AI is now in Cumbersome
Open Settings > AI > Connect Provider > Apple Private Cloud Compute. Cumbersome checks whether your device can use it, then asks you to agree to send your messages and conversation context to Apple. After setup, you can select it from the model picker.
You'll need iOS 27, iPadOS 27, or macOS 27 or later, Apple Intelligence enabled on a supported device, a supported region, and internet access. Apple sets the daily usage allowance.
Apple calls the developer API model AFM 3 Cloud (opens in new tab). Its model family announcement (opens in new tab) describes a collaboration with Google, but doesn't equate AFM 3 Cloud with a named Gemini model. The tests here measure the Apple service available to developers. I ran them on my Mac.
PCC's arithmetic fell apart on the repeat
I also repeated two prompts, each containing six questions. I'd chosen these repeats in advance: one batch of formal fallacies and one batch of arithmetic.
PCC got all six arithmetic answers right the first time and all six wrong the second time. Same prompt, fresh conversation, same settings. For example:
((-5 + -8 - -6 * 7) + (9 * -7 - -5 - -4)) =
The answer is -25: the two parentheses evaluate to 29 and -54. PCC answered -25 first, then -14 on the repeat.
It also corrected two of its earlier formal fallacy mistakes. Across the twelve repeated questions, PCC changed eight answers and Flash-Lite changed four. Pro, Flash, and Argon kept all twelve answers unchanged and correct.
That arithmetic result bothers me more than the total score. PCC's first pass looked good enough that I could easily have trusted the next calculation.
My results alongside published benchmarks
My 72-question results sit beside the Artificial Analysis Intelligence Index (opens in new tab) below:
| Model | My test (correct) | Published index |
|---|---|---|
Apple PCC (AFM 3 Cloud) | 58/72 | Not listed |
Gemini 3.5 Flash-Lite | 46/72 | 22 |
Gemini 3.1 Pro Preview | 72/72 | 30 |
Gemini 3.8 Flash | 72/72 | 40 |
Gemini 4 Argon | 71/72 | 53 |
The middle column counts correct answers out of 72. The published index uses a separate scale. Its Flash score uses medium reasoning and Argon uses high; my test used medium for every Gemini model. The published Pro and Flash-Lite labels don't specify a reasoning level.
The published scores show much bigger differences between Pro, Flash, and Argon than my questions could reveal. I couldn't find a matching AFM 3 Cloud result on Artificial Analysis, Arena (opens in new tab), or LiveBench (opens in new tab). Apple's own published results compare it with earlier Apple models using private human assessments.
For the questions I ran, PCC sits above Flash-Lite and below the other three.
Why try PCC in Cumbersome?
PCC chat is free in Cumbersome. You don't need Cumbersome Pro, an API key, or a separate AI subscription. Apple covers the cloud API cost (opens in new tab) for eligible developers, so there's no token bill to pass on. You get a daily allowance, with higher limits through iCloud+ (opens in new tab).
That changes how I judge the result. PCC's 58/72 is respectable, even if the stronger Gemini models beat it. I'd try it for drafts, summaries, and everyday questions. I'd check PCC's arithmetic and use a stronger model for demanding reasoning.
Privacy is the other reason I added it. According to Apple's security design (opens in new tab), your requests are encrypted to verified PCC servers, their contents aren't accessible to Apple staff, and they're deleted after the response. Researchers can inspect the server software (opens in new tab) that your device checks before sending a request.
The model runs in the cloud, so there's no model to download or fit into your phone's memory. You need an internet connection. Cumbersome's saved conversations and iCloud sync still follow your settings; use Incognito when you don't want to save the chat.
What I ran and what it cost
On October 24, 2026, each service answered twelve batches of six questions, followed by two repeated batches. The repeats don't count toward the 72-question score. I requested JSON answers and accepted Markdown code fences around them when present.
Every request started with fresh history, no custom system instructions, no tools, and no web access. I used default sampling, an 8,192 output token cap, and moderate reasoning on PCC versus medium on Gemini. Those settings don't guarantee equal computation. All completed responses fit within the cap.
Gemini ran through OpenRouter's Google Vertex endpoints with provider fallbacks disabled. All 56 completed requests, including repeats, cost $0.31. PCC reported about 23,900 tokens across fourteen completed requests and never showed a quota warning.
This is a selected, batched BBH test, not a full benchmark score. The questions are public and could have appeared in training. Apple's reported server metadata varied during the run, although it matched across the two arithmetic attempts. These results cover the six reasoning tasks above.
Bless up! 🙏✨