The first three cards were the “big three” closed labs (OpenAI, Anthropic, Google). xAI is the next frontier provider engineers actually see on shortlists: Grok models, a developer API, and explicit reasoning SKUs. Vol. 11.5 did not give Grok its own family page—treat this lecture as the curriculum’s Grok home, still using the same closed-API logic as closed-source models.
xAI is tightly associated with the X (Twitter) consumer surface and with xAI’s API console. That distribution story matters for product design (realtime/social context, brand tone) but does not authorize unverified claims about training data or secret MoE diagrams. After Grok, Module 22.1 shifts toward open-weight-heavy vendors starting with Meta Llama.
Learning Objectives
By the end of this lesson, students should be able to:
- Identify xAI as a closed frontier lab and Grok as its model/product family.
- Separate Grok-in-X (consumer) from the xAI developer API.
- Describe reasoning SKUs qualitatively (extra test-time compute / tokens) without fake bench scores.
- Fill a Pricing / Strengths / Weaknesses card vs OpenAI, Anthropic, and Google.
- Sketch an OpenAI-compatible-style xAI chat call and know to re-read current base URLs and model IDs.
- State brand, ToS, and data-path risks unique to X-adjacent distribution.
xAI is the lab behind the Grok family of large language (and multimodal, depending on SKU) models, delivered via consumer Grok apps / X integration and via the xAI API. Reasoning SKUs (names such as Grok reasoning / “think” modes—confirm current IDs) spend additional test-time compute to improve multi-step answers; they are still closed weights. This page catalogs public product posture. It does not assert unpublished architecture or training-set composition.
Product Surfaces
| Surface | Job | Engineering note |
|---|---|---|
| Grok on X / grok.com apps | Consumer chat, search-flavored answers, image gen (SKU-dependent) | Great for brand/UX spikes; not your multi-tenant backend |
| xAI API | Metered chat/completions-style access for apps | Historically OpenAI-compatible enough to reuse client patterns—verify current docs |
| Reasoning SKUs | Harder math/code/planning with extra thinking tokens | Budget latency + token spend; eval on your tasks (Vol. 19) |
| Realtime / X context (product claim) | Answers that can incorporate recent public X posts when enabled | Treat as a grounding mode with ToS + hallucination risk—not ground truth |
Pricing / Strengths / Weaknesses
| Axis | Pricing (qualitative) | Strengths | Weaknesses |
|---|---|---|---|
| Fast / standard Grok SKUs | Token-metered API; cheaper/faster SKUs vs flagship. Consumer Grok / X Premium-class plans are seats, not API invoices. | Competitive closed alternative; often OpenAI-shaped API ergonomics; strong public “witty / less lecturing” product tone some apps want. | Smaller enterprise ecosystem than OpenAI/Google/Azure; closed weights; brand association with X may be a procurement plus or minus. |
| Reasoning SKUs | Higher effective cost per answer: thinking tokens + wall-clock. Do not assume they are cheaper than GPT/Claude reasoning modes without measuring. | Useful when your eval shows multi-step gains worth the latency. | Easy to overspend if every chatbot turn uses reasoning; not a substitute for tools + RAG. |
| X-grounded / realtime modes | May include extra retrieval or tool units—read the live price page. | Differentiator vs labs without a social firehose product. | Social posts are not citations; Vol. 19 hallucination tests still apply; moderation/ToS scrutiny. |
| Multimodal add-ons | Image understanding/generation SKUs billed separately or via the same meter—confirm, do not invent per-image dollars. | One more closed multimodal option. | Thinner catalog than Google (Veo/Imagen/NotebookLM) or OpenAI (Whisper/Sora). |
Pick xAI when
- You want a fourth closed frontier for failover/eval
- Product tone or X-adjacent features are a real requirement
- Reasoning SKU wins your internal hard-task eval at acceptable latency
Look elsewhere when
- You need Azure/Vertex/Bedrock procurement paths first
- You need Whisper/Sora/Veo/Claude Code depth
- Brand/ToS risk around X is a blocker for the customer
Always still do
- Vol. 19 eval on your traces, not screenshots of leaderboards
- Vol. 20 data map: what leaves your VPC to xAI
- Pin model IDs; reasoning off by default for cheap intents
Why Grok appears on RFPs
- Another frontier closed model to break duopoly pricing
- Familiar chat API shape for Vol. 18.1-trained teams
- Reasoning SKUs as an explicit product line
Why it may fall off
- Less mature enterprise control story than GCP/Azure incumbents
- Closed: no self-host flagship
- Consumer Grok ≠ production API policy
Practical API Sketch
xAI has documented OpenAI-compatible chat endpoints. Confirm base URL, model IDs, and reasoning parameters on current docs before shipping.
Related Lectures
| Lecture | Role |
|---|---|
| Google / DeepMind | Previous closed estate |
| Closed-source models | How to think about any closed API lab |
| OpenAI SDK | Compatible client pattern many xAI spikes reuse |
| Hallucination tests | Required if you enable X/realtime grounding |
| Chatbots / AI search | Product types Grok often powers |
| Meta (Llama, SAM, Meta AI) | Next: open-weight center of gravity |
Grok on X is the same contract as the xAI API. Second: reasoning SKUs are always worth enabling. Third: realtime/X grounding is citation-grade evidence. Fourth: this lecture contains secret xAI training recipes. Fifth: Grok is open-weight like Llama. Sixth: brand tone (“less censored” marketing) replaces your Vol. 13 guardrails and Vol. 20 policy.
Knowledge Check
- Short Answer: What is xAI’s flagship model/product family called? Answer: Grok.
- True/False: Flagship Grok weights are generally open for download like Llama. Answer: False.
- Multiple Choice: Reasoning SKUs typically: (a) spend extra test-time compute/tokens, (b) are free and instant, (c) replace RAG forever. Answer: (a).
- Short Answer: Why separate Grok-on-X from the xAI API? Answer: Consumer/workspace product vs metered developer contract (billing, ToS, SLAs differ).
- True/False: Social posts used as grounding are automatically ground-truth citations. Answer: False.
- Multiple Choice: A common xAI spike uses: (a) an OpenAI-compatible client + xAI base URL, (b) Gemma weights from DeepMind, (c) Claude Artifacts only. Answer: (a).
- Short Answer: Name one qualitative weakness vs Google’s estate. Answer: Thinner multimodal/cloud catalog (no Veo/Imagen/NotebookLM/Vertex-depth)—or smaller enterprise ecosystem (either OK).
- True/False: You should memorize invented Grok dollar prices from this page. Answer: False.
- Multiple Choice: Enable reasoning by default on every chatbot turn: (a) usually wasteful; route hard tasks only, (b) required by law, (c) makes weights open. Answer: (a).
- Short Answer: Which vendor lecture comes next? Answer: Meta (Llama, Segment Anything, Meta AI).
Key Takeaways
- xAI = Grok consumer surfaces + developer API + explicit reasoning SKUs; closed weights.
- Qualitative pricing: seats vs tokens; reasoning costs extra latency and tokens.
- Differentiator: X-adjacent realtime story—treat as fallible grounding, not evidence.
- Compatible client spikes are fine; pin IDs; keep Vol. 19/20 controls.
- Next: Meta Llama / SAM / Meta AI opens the open-weight half of 22.1.
Lab: Vendor card: Grok fast vs reasoning SKU routing table (intent → SKU). One compatible chat spike if keys exist; otherwise a dry-run config (base URL, env var, pin ID). Write a one-paragraph ToS/brand risk note for a regulated Vol. 21 vertical (finance/health/education)—no X scraping how-tos.
Whiteboard: Closed four: OpenAI, Anthropic, Google, xAI. Then a divider: “open-weight labs start here” → Meta.
Recap: xAI’s Grok API and reasoning SKUs are a fourth closed frontier option. Next is Meta.