🎯 What We'll Cover
This is the practical pay-off of the week: a tool-by-tool tour with honest verdicts. For each category we ask three questions — what does it actually do well, what is its specific failure mode, and what does the free tier deliver for a student who cannot put a foreign credit card into a sign-up form? That last question is not a footnote in this course; for many UCT students it decides whether a tool exists at all.
The centrepiece is a comparison of Deep Research modes — the agentic-RAG tools from 10.4 — because the Week 10 activity (10.6) hinges on them. We then cover the Chinese free-tier options (often more generous than the Western ones, with one serious catch), the MCP connectors worth wiring into a research workflow, what is expensive and skippable, and a free-tier research pipeline you can actually run.
A hard warning before any number appears: every quota in this sub-lesson is a May 2026 snapshot and several will be wrong by the time you read this. Free tiers are the most volatile thing in the entire field — they change monthly, sometimes weekly. Treat the specific figures as “go and check”, not gospel. The framework for choosing is the durable part.
🌍 Why “Free” Is the Whole Game Here
A course written in California would treat the free-versus-paid question as a budgeting detail. From Cape Town it is structural. The UNDP's 2025 analysis of African AI talent (drawing on Zindi's network of some 11,000 data scientists) found that only about 5% have access to the computing power their work needs — roughly 1% with on-premise GPUs and 4% paying for cloud access. The continent holds an estimated 0.1% of the world's computing capacity, while 60% of the top supercomputers sit in just three countries. The lived effect is stark: a well-resourced researcher in a G7 country can retrain a model every thirty minutes, while an African peer may wait up to six days between iterations.
The compute gap is why this week is free-tier-first
When 95% of a continent's AI researchers depend on free or shared tools, “just pay for the Pro plan” is not advice — it is an assumption that excludes most of the people this course is for. So everything below is filtered through what is genuinely reachable on a free tier from a South African connection. This is the same equity argument the Esethu Framework (Week 4, expanded in Week 11) makes about infrastructure: access is not a detail, it is the precondition.
The practical access picture as of May 2026: Claude.ai, ChatGPT, Gemini, Perplexity, and Microsoft Copilot all work from South African networks without a VPN, and UCT's Microsoft 365 subscription means Copilot Chat is very likely already available to you through your university account. The Chinese services below also work from a Cape Town connection without a VPN. One thing unites all of them, Western and Chinese alike: they process your data on servers outside South Africa — a data-protection point we treat seriously below.
🔎 Deep Research Modes: The Comparison That Matters
Deep Research modes take a question, spend several minutes planning, searching, and reading, and return a cited report — the consumer face of the agentic RAG from 10.4. They are also what the 10.6 activity asks you to test. Here is the free-tier landscape, with the staleness warning in force.
⚠️ Quotas below are a May 2026 snapshot — verify before relying
These numbers move constantly. While this was being written, sources already disagreed about Perplexity's free Deep Research allowance (some said five a day, one said none) — which tells you exactly how much weight to put on any single figure. Check the current limit yourself before building the activity around a tool.
| Tool | Free-tier Deep Research (May 2026) | Notes |
|---|---|---|
| Perplexity | ~5 / day | The most generous Western free tier; the recommended default for the activity. Comet browser also free. |
| Kimi (kimi.com) | Usable, quota not published | Moonshot AI. The strongest free alternative to Perplexity — cited, visualised reports. |
| ChatGPT | ~5 / month, then a lightweight fallback | Hard monthly cap on the full version; the fallback is shallower. |
| Gemini | ~5 reports / month | On the Gemini Apps free tier; 1M-token context on the Pro models. |
| DeepSeek (chat.deepseek.com) | No one-button mode; approximate it manually | DeepThink reasoning + web search + file upload, chained by hand, gets close. Fully free, English UI. |
For the 10.6 activity, the recommended free Deep Research tool is Perplexity; if its quota is exhausted or sign-up is blocked, Kimi is the strongest substitute — provided you read the data-protection note next, and disclose your choice as the activity requires.
🌐 Chinese Free Tiers: Often More Generous
The Western free-tier list is West-Coast-heavy, and for a student who cannot pay, that matters — the major Chinese services are often more generous, accept email or Google sign-up, and (as of May 2026) work from a Cape Town connection without a VPN. They belong in any honest free-tier map. The honest ranking for the UCT student who cannot pay:
1. DeepSeek — chat.deepseek.com
The most capable fully-free reasoning chat in May 2026: V4 plus a DeepThink toggle, web search, and file upload, with no message cap and no Chinese phone number required. English UI, email or Google sign-up. Throttles to “Server Busy” at peak times.
2. Kimi — kimi.com
Moonshot AI. The closest free analogue to Perplexity Deep Research, plus document and slide generation and a very long (256K+) context window. Google sign-in, English UI. Free Deep Research quota is not published — expect it to tighten.
3. Qwen — chat.qwen.ai
Alibaba. The widest modality coverage on a free tier: hybrid thinking, web search, image/video understanding, image generation. Note: Qwen's free coding-agent CLI was discontinued on 15 April 2026 — the chat surface stays free, the coding CLI now needs payment.
4. Z.ai / Zhipu — chat.z.ai
Zhipu AI. The most open-weights-aligned option: free chat on GLM-4.5/4.6 (MIT-licensed and self-hostable) with an agent mode and document generation. Free agent runs are slower and shorter than DeepSeek's or Kimi's.
⚠️ Sending data abroad: a rule for all these tools, not just the Chinese ones
It is tempting to attach the data-protection worry only to the Chinese services, but that would be both unfair and wrong. Almost every free AI tool in this sub-lesson processes your data outside South Africa — the US-based services (ChatGPT, Claude, Gemini, Perplexity) on infrastructure outside South Africa, typically in the United States; the Chinese services (DeepSeek, Kimi, Qwen, Z.ai) on servers in mainland China. Under South Africa's Protection of Personal Information Act (POPIA, section 72), the relevant question is not which country, but whether sending personal information there is permitted at all: a transfer abroad is restricted unless the recipient is bound by data-protection rules substantially similar to POPIA's, or the data subject has consented (among a few narrow conditions). POPIA has no “approved-country” whitelist — the test is the recipient's actual protection, which for any consumer AI service hosting your prompts abroad you have no basis to assume is met, wherever its servers sit.
In plain terms, applied equally to the US-hosted and the China-hosted tools: do not put identifiable personal data, participant information, unpublished results, or third-party confidential material into any of them. For public, non-personal research questions they are all excellent and free. For anything involving real people's data, the obligation is the same regardless of vendor — and you must record which tool you used and where its data went. The 10.6 activity makes this disclosure a required, graded section, for every tool, not an afterthought.
🔌 MCP Connectors Worth Wiring In
From 10.3: MCP lets an agent reach your tools through one standard. A handful of connectors are genuinely useful for research — reference managers (community-maintained Zotero MCP servers exist; search for the current one), GitHub, Google Drive, and Notion. On Claude.ai a set of connectors is available even on the free plan (free users get one custom connector). Used well, an agent can search your own library or read your repository through a single interface.
⚠️ Every connector is two risks at once
Each connector you add does two things beyond its convenience. First, it widens the prompt-injection surface (10.2): an agent that can read your Drive and browse the web can be steered by hostile text in a document or page. Second, it opens a POPIA-relevant data flow: connecting your inbox or institutional Drive sends that content to the model provider. Connect deliberately, one at a time, with the lowest permissions that do the job — and never connect a personal-data source to a tool whose servers you would not trust with that data.
💰 What's Expensive — and Probably Skippable
For a postgraduate on a stipend, most of the heavily-marketed paid agents are not worth it:
- Enterprise agents (Sierra, Decagon, Cognition's Devin) — priced for companies; not aimed at, or affordable for, an individual researcher.
- Manus — the “general agent” from 10.3. Access is mixed and its capability claims are company-reported; not something to build a workflow on now.
- Cursor Pro — genuinely good if you code every day; if you do not, the free coding tools cover it.
- Paid Deep Research tiers — faster and higher-quota, but the free tiers are enough to do real work and to complete this course. The activity is deliberately built to be done entirely free.
🌍 A Week 3 reminder: agents are expensive in tokens, not just money
An agentic run — planning, many tool calls, reading long results, retrying — burns far more tokens than a single chat message, sometimes by orders of magnitude. That is why free Deep Research is rationed, and it is also a Week 3 environmental point: a Deep Research report has a much larger compute (and therefore energy and water) footprint than a one-shot answer. Use the heavy tools when the task warrants them, not reflexively.
🧾 A Free-Tier Research Pipeline
Putting the week together, here is a five-step pipeline a UCT student can run entirely on free tools — and which deliberately keeps you in the loop at every verification point:
- Explore — run your question through a free Deep Research mode (Perplexity, or Kimi for non-personal topics) to map the landscape and surface sources. (Week 10.4)
- Verify — do not trust the report. Run every citation through the Week 5 Five-Point Citation Check and the Week 9 “which model, which date, retested?” check. Expect some sources to be wrong or invented.
- Synthesise — organise what survived verification into your own structure, in your own words. The work of deciding what matters and how it fits together is yours; a chat tool can react to your structure, not replace it. (Week 6)
- Write — write your own first draft. Then use AI to refine it — tightening, questioning, suggesting — checking each change as you go. This is the Week 6 order, and it matters: the thinking and the first words should be yours, with AI sharpening what you wrote, not generating it for you. Drafting by prompt is how you end up with fluent text you do not actually understand or stand behind.
- Audit — before anything leaves your hands, check it as you would AI-generated code: claims against sources, numbers against primaries, and a record of which tool did what. (Weeks 7 and 9)
Every step here has a free tool behind it, and every verification step is yours, not the agent's. That is the whole argument of the week in one workflow: agents change what the tools can do; they do not change who is responsible for the result.
📖 Sources & Further Reading
- UNDP, “Only five percent of Africa's AI talent has the compute power it needs” — the compute-equity figures.
- South Africa Information Regulator (POPIA) · Michalsons, “How POPIA affects AI” — for the cross-border data-flow rules behind the Chinese-tools caveat.
- Deep Research free tiers (verify current): Perplexity · OpenAI Deep Research FAQ · Gemini.
- Chinese free tiers: DeepSeek · Kimi · Qwen · Z.ai.
- Esethu Framework (Rajab et al., 2025). arXiv:2502.15916 — the equity-and-sovereignty argument, expanded in Week 11.
👉 What Comes Next
Sub-Lesson 10.6 — Hands-On Activities and Assessment. Now you put it to work. The headline activity, “Same Task, Three Ways”, has you run a research question from your own field through plain chat, chat-with-tools, and a Deep Research mode, then judge the results with the Week 9 failure taxonomy and verify them with the Week 5 citation checks — using free tools only, and disclosing where your data went. It is the whole week, made concrete on a question you actually care about.