xAI Grok (4.6 / 4.3 / Grok Build): the cheapest US flagship on paper, with a content-safety gate before student use

In procurement meetings, IT coordinators often line Grok's $2 / $6 rate against GPT-6 Astra's $10 / $50 and call it "five times cheaper" — yet inside Edor.ai it is reachable only through Poe, and it is tied to X platform content more tightly than any other US vendor, so student use is not a straight swap for a cheaper ChatGPT.

xAI was founded by Elon Musk in 2023 and is the youngest of the four major American providers. For a Hong Kong school the appeal is list price well below OpenAI and Anthropic at comparable capability; the cost is that response style, X search tools and a non-native integration path all need a documented assessment the platform cannot sign on the school's behalf.

What the company is actually building

xAI moves fast. Grok 4 in July 2025 put the brand into frontier-model contention, and since then releases have landed roughly quarterly, with Grok 4.6 in August 2026 already a third-generation flagship. Two constants shape the product: price below OpenAI and Anthropic, and deep embedding in the X platform — X search as a first-party tool and a response style presented as less restricted.

The second point is a selling line for developers and a compliance question for schools. Hong Kong primary and secondary schools answer to minors, parents and the question "why this vendor?" — none of which is settled by "try it and see". An IT coordinator and panel heads should sign a content-safety assessment first. The Grok Build coding agent (grok-build-0.1) is a separate line aimed at ICT programming, not general classroom chat.

Model line-up at a glance

The Grok family visible in September 2026 splits cleanly by reasoning strength and context length:

The xhigh level on Grok 4.6 sits above low, medium and high for long-chain reasoning on Citizenship and Social Development or History marking. Once a request reaches 200K tokens, the whole request bills at $4 / $12 — the same class of long-context trap as on OpenAI's flagship.

Where the money actually goes

At $2 / $6, Grok list price is about a fifth of GPT-6 Astra's $10 / $50 — the line that dominates a procurement slide. Three layers matter on a school platform:

First: inside Edor.ai you pay Poe, not xAI. An administrator adds a Poe key and selects Grok from the supported list; there is no separate xAI key field. Monthly spend follows the Poe plan, not xAI's pricing page multiplied by usage.

Second: 200K tokens is a hidden switch. A 500K or 1M window sounds generous, but crossing 200K steps the whole request up a price tier. Pasting a term of material at once is often twice as expensive as retrieval over the relevant passages.

Third: tool calls bill separately. X search and other server-side tools are itemised on xAI's pricing page, along with batch discounts. An IT coordinator's estimate should include them, not only "cheaper than OpenAI".

A list price five times lower buys an extra Poe hop, a content-safety assessment and a separate OpenAI key for knowledge base and speech — the cheapest flagship is not the simplest school platform.

What role it plays inside Edor.ai

xAI is not a native provider. The native routes are OpenAI, Azure OpenAI, Anthropic, Poe and local Ollama; Grok arrives only through Poe. Before enabling it, complete a content-safety assessment signed by the IT coordinator and relevant panel heads. A sound sequence is teacher tools only for a term, then a management discussion about students.

For teachers

During a Wednesday break, a Secondary 2 History teacher tells the panel head in the staff room: "I ran twenty data-response drafts through Grok — the unit cost is real, but a couple came back too blunt and I still had to edit each one." Bulk question generation, English writing feedback, marking that needs point-by-point argument, and long-document comparison are where the low unit price pays off. All output still needs teacher review, especially Traditional Chinese and Hong Kong curriculum wording — public classroom evidence is far thinner than for OpenAI or Claude, so Chinese subjects deserve a small comparison first.

For students and audit

Our advice is keep student conversations on an already assessed provider until assessment and written management agreement are in place. If a school proceeds afterwards, keep four-layer gating, teacher monitoring, and a more conservative tutor-mode hint ceiling than the default. Every call still lands in ai_audit_logs; privacy assessment must cover Poe as an additional third party.

Poe carries no embedding or speech models — the school knowledge base and oral practice still need an OpenAI or Azure OpenAI key, as with other Poe-served providers.

Where it fits and where it does not

Fits: high-volume, repeatable, teacher-reviewed work; Citizenship or History marking that benefits from xhigh reasoning; long-document tasks kept under 200K tokens; adult users only after a documented content-safety assessment.

Does not fit: student access before assessment; treating Poe pricing as xAI list pricing; pasting a whole term and triggering doubled billing; expecting Grok alone to run knowledge base and Cantonese oral practice; default Chinese-subject marking without local testing.

Copy these into an Edor.ai teacher tool to test:

You are a Hong Kong secondary Citizenship and Social Development teacher. Below is a Secondary 4 student's data-response answer on population ageing in Hong Kong.
Output: first identify which claims are supported by the source material and which are bare assertions; then score Understanding of Sources, Argument and Expression out of 5 each with two comments per criterion; finish with three directions for improvement.
Constraints: write in Hong Kong English conventions; do not supply arguments on the student's behalf; do not introduce statistics beyond the source material provided; if you want to cite any figure whose origin you are unsure of, write "teacher to verify".
[paste sources and student answer]

You are a Hong Kong primary English teacher. Design a 20-minute reading comprehension worksheet for Primary 4 on the theme "a day in the life of a community cleaner".
Output: a passage of about 180 words, five comprehension questions (two literal retrieval, two inferential, one personal response), and a teacher marking guide.
Constraints: use the English conventions common in Hong Kong schools; keep the vocabulary at Primary 4 level; avoid proper nouns pupils would need a dictionary for; if you are unsure of the official English name of a Hong Kong occupation or department, write "teacher to verify" rather than inventing one.

To compare the other US vendors, read the four-provider comparison. For native integration covering embedding and speech, see OpenAI or Anthropic Claude.

The trade-off for schools

• If you want the lowest list price at comparable capability and only adult, reviewable tasks, limit Grok to teacher tools and complete a content-safety assessment before enabling it.
• If you plan student access, obtain written management agreement first, keep four-layer gating and audit logging in place, and do not assume Poe routing equals direct coverage under xAI account terms.
• If you need knowledge-base retrieval and oral practice together, keep an OpenAI or Azure OpenAI key and use RAG to control request length so you do not cross the 200K-token doubling threshold.