AI models are the trained software systems that read a prompt and produce text, code, images or answers, and ChatGPT, Claude and Gemini are the apps built on three families of them: OpenAI's GPT, Anthropic's Claude and Google's Gemini. Which model sits behind a tool decides what you pay, how fast it replies and where your data travels, and the same families sit behind many of the chat and answer tools covered in our guide to AI search optimisation.

This post is a map as at 13 November 2026: each vendor's current lineup from its own documentation, what a standard job costs in Rand on every tier, and a decision table for choosing one. Model names change every few months, so it teaches the structure first and the names second. For a head-to-head between two of the families, see Claude vs ChatGPT.

Quick Answer

AI models are the engines; ChatGPT, Claude and Gemini are the products built on them. In November 2026 OpenAI and Anthropic each sell a top tier, a mid tier and a small, cheap tier, and the price gap between top and small is 100 times at OpenAI and up to 80 times at Anthropic; Google's range is narrower. As a cost-first working rule, try a South African business task on a cheaper tier and move up when it fails on your own examples. Check POPIA section 72 before putting customer personal information into any of them.

Wondering whether answer engines can find your business?

Send us your website and the five questions your customers ask most, and we will tell you where your visibility in AI answers looks weakest.

Request a visibility review

What Are AI Models and How Do They Differ From ChatGPT or Gemini?

An AI model is the trained system that does the work, while a product such as ChatGPT or Gemini is the app, account and billing wrapped around it. The same model can reach you in three ways: through a chat subscription, through an API billed by usage, or built into other software such as Microsoft Copilot.

The distinction matters because the three routes are priced differently. A subscription is a flat monthly fee for a chat interface; our post on ChatGPT Plus in South Africa prices that route in Rand. An API charges per token, a token being a chunk of text rather than a word. Anthropic's documentation puts 1M tokens at roughly 555k words on its current tokenizer, and other vendors count differently, so treat any token-to-word conversion as approximate. Our OpenAI API pricing post covers that second route for one vendor.

A model also carries a version name, and vendors retire and reprice versions on their own dates, which is why every name in this post is dated and why the last section covers what to watch.

Which Vendors' Lineups Can a South African Business Use in November 2026?

A South African business can choose between three closed-model vendors, OpenAI, Anthropic and Google, plus open-weight models it can download. These are the main types of AI models on offer, and the table keeps the AI models compared on one basis: published list price per million tokens. OpenAI and Anthropic each sell a top, a middle and a small tier; Google's Gemini 4 range is narrower, with two reasoning tiers at the same list price and an everyday tier. The grouping of each vendor's published descriptions into tiers is ours.

Vendor and modelTierPublished price per million tokens, input / outputVendor's own description
OpenAI GPT-6 AstraTop$10 / $50"Our most capable model for the most demanding work."
OpenAI GPT-6.1 SolMiddle$2 / $10"Near-Astra performance for complex work at a lower cost."
OpenAI GPT-6.1 TerraMiddle, lighter$1 / $5"A fast, efficient model for visual reasoning, precise computer use, and clear, natural writing."
OpenAI GPT-6 LunaSmall$0.1 / $0.5"Our most efficient model for focused, high-volume tasks."
Anthropic Claude Fable 5.2Top$8 / $40For demanding reasoning and long-horizon agentic work
Anthropic Claude Opus 5.5Upper middle (Anthropic's suggested starting point)$4 / $20For long-running agentic coding and knowledge work
Anthropic Claude Sonnet 5.5Middle$2 / $10The best combination of speed and intelligence
Anthropic Claude Haiku 5.5SmallFrom $0.10 / from $0.50For high-volume, latency-sensitive tasks such as classification, extraction, and routing
Google Gemini 4 CarbonTop$4 / $20 from 1 December 2026 ($2 / $10 until 30 November)Strongest reasoning model, built for complex software engineering and multimodal knowledge work
Google Gemini 4 ArgonMiddle$4 / $20 from 1 December 2026 ($2 / $10 until 30 November)Efficient frontier reasoning model with faster responses and lower token spend per turn than Carbon
Google Gemini 4 HydrogenEveryday (no priced small tier in this table)$1.50 / $7.50; the API has a free tierSpeed and efficiency for everyday coding and agentic workflows

These come from OpenAI's model list, Anthropic's models overview and Google's Gemini pricing page, all read on 13 November 2026. Prices are in US dollars, and all three vendors list a context window of about a million tokens on their current flagship models.

Open-weight models such as Meta's Llama, DeepSeek, Alibaba's Qwen and Mistral's releases sit outside the table. Their weights can be downloaded and run on infrastructure you control, which matters when data must stay in your own environment. Licences differ by family and sometimes by release, so read the licence before building a product on one.

Local usage data on AI models in South Africa comes from larger firms. In the South African Generative AI Roadmap 2025, run by World Wide Worx with Dell Technologies and Intel among more than 100 mid-sized and large enterprises, ITWeb reports generative adoption rose from 45% in 2024 to 67% in 2025.

In that 2025 survey, ChatGPT was used by 93% of those enterprises, Microsoft Copilot by 74% and Google Gemini by 41%, down from 56% in 2024. The findings ITWeb reports do not mention Claude, and the survey covers large enterprises, not small businesses. See our post on Microsoft Copilot in South Africa for that product.

Quick Answer: Which Models Exist?

OpenAI, Anthropic and Google each sell a family of models in tiers, and open-weight families such as Llama, Qwen and Mistral are the route when data has to stay on your own infrastructure. In the one South African enterprise survey we found, large and mid-sized firms reported using ChatGPT (93%), Microsoft Copilot (74%) and Gemini (41%) in 2025; Claude was not mentioned in the findings reported.

How Much Do the Tiers Cost Per Job in Rand?

A tier's real cost is the tokens you send in plus the tokens you get back, priced separately, so a per-million-token price means little until it is applied to a job. The worked example below uses one assumption you should replace with your own: a job of 2,000 tokens in and 1,000 tokens out, a mid-sized drafting or summarising task, run 1,000 times.

The arithmetic, using OpenAI's top and smallest tiers

1,000 jobs is 2 million tokens in and 1 million tokens out. On GPT-6 Astra that is 2 x $10 + 1 x $50 = $70, which at R16.42 to the US dollar is R1,149.40. On GPT-6 Luna it is 2 x $0.1 + 1 x $0.5 = $0.70, or R11.49. The top tier costs 100 times the small one for the same job. R16.42 is the exchange-rate assumption used in the SA Digital Cost Index (August 2026), not a live rate, so rerun it at today's rate.

ModelCost of 1,000 jobs in US$Cost of 1,000 jobs in RandBest for
OpenAI GPT-6 Astra$70R1,149.40Rare, hard problems where a wrong answer is expensive
OpenAI GPT-6.1 Sol$14R229.88Complex work at a lower cost
OpenAI GPT-6.1 Terra$7R114.94Fast writing and visual tasks
OpenAI GPT-6 Luna$0.70R11.49Focused, high-volume tasks
Anthropic Claude Fable 5.2$56R919.52Long, multi-step agentic work
Anthropic Claude Opus 5.5$28R459.76Long-running knowledge work and agentic coding
Anthropic Claude Sonnet 5.5$14R229.88Everyday business tasks
Anthropic Claude Haiku 5.5, from-price$0.70R11.49Sorting, extraction, routing at volume
Google Gemini 4 Carbon, list price from 1 December 2026$28R459.76Complex reasoning and multimodal work
Google Gemini 4 Argon, list price from 1 December 2026$28R459.76Complex reasoning with lower token spend per turn than Carbon
Google Gemini 4 Hydrogen$10.50R172.41Everyday coding and workflows; has a free API tier

Anthropic's overview says Batch API requests are 50% off, and OpenAI's pricing page lists a 50% reduction for batch mode, so work that can wait cuts these figures further. Cached input is discounted as well, by about 90% on OpenAI's page. Subscriptions are flat-fee and not comparable to these API figures, and everything here is billed in dollars, so a weaker Rand raises your bill with no change in usage.

Quick Answer: What Does the Gap Cost?

On a 2,000-token-in, 1,000-token-out job run 1,000 times, the smallest tiers cost about R11.49 and the mid tiers such as GPT-6.1 Sol and Claude Sonnet 5.5 cost R229.88, at the SA Digital Cost Index's R16.42 assumption. The top tiers run from R459.76 to R1,149.40. The gap between top and small is 100 times at OpenAI and up to 80 times at Anthropic, measured against Haiku's from-price.

Which Model Tier Should You Pick for a Given Job?

Pick the cheapest tier that passes your own examples, because the right tier is set by the job and the cost of a wrong answer, not by the vendor's brand. For AI models for business use, the table is our judgement keyed to the vendors' own descriptions, not a benchmark result.

Job or constraintStart atMove up whenBest for
Tagging enquiries, extracting fields from invoices, routing tickets at volumeSmall tier (GPT-6 Luna, Claude Haiku 5.5)It mislabels cases you check by handRepeatable, checkable work
Drafting customer replies, product descriptions, summariesMiddle tier (Claude Sonnet 5.5, GPT-6.1 Sol, Gemini 4 Argon) or Google's everyday Gemini 4 HydrogenEditing takes longer than writing it yourselfMost day-to-day business writing
Multi-step analysis, long agent workflows, hard codeTop tier (GPT-6 Astra, Claude Fable 5.2, Gemini 4 Carbon); Claude Opus 5.5 sits just belowOnly when the middle tier fails repeatedly on the same taskLow volume, high cost of error
Prompts that contain customer personal informationAny tier, after the POPIA check belowNot a tier question: it is a data questionDeciding where data may go
No developer and a small teamA chat subscription, not the API; see ChatGPT alternativesUsage grows to the point a per-job bill is cheaperGetting started without building anything

As a working rule of thumb, not a statistical standard, run about twenty of your own real tasks through two adjacent tiers and compare the outputs side by side before committing. No public leaderboard knows your invoices, your tone of voice or your suppliers. Note the clash with the vendor: Anthropic's own advice is to start with Claude Opus 5.5 for most workloads, while our rule is cost-first, so decide on your own test results. For how the families compare on marketing work, see Gemini vs ChatGPT.

A model name is not something a business optimises for: the one official statement we checked, Google's documentation for its AI search features, says there are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary. Its guidance is to meet the technical requirements for Google Search and to create helpful, reliable, people-first content, and performance is reported in Search Console like other search traffic.

That page covers Google only and does not say which model powers those features, so it is not evidence about how ChatGPT or Claude choose what to cite. Treat claims that a specific model rewards a specific trick as practitioner hypothesis until a vendor documents it. For how the main answer engines differ in practice, see ChatGPT vs Google vs Perplexity, and for the brand-visibility side, LLM optimisation for South African brands.

What Does POPIA Say About Putting Customer Data Into a Chatbot?

POPIA section 72 restricts sending personal information to a third party in a foreign country unless one of its listed conditions is met. Whether a prompt containing a customer's name or order, sent to a vendor that processes it abroad, falls within that section is a legal question for your attorney. The conditions in section 72(1), as set out on popia.co.za, are that:

  • the recipient is subject to a law, binding corporate rules or a binding agreement giving an adequate level of protection;
  • the data subject consents;
  • the transfer is necessary for a contract with the data subject, or for pre-contractual measures;
  • the transfer is necessary for a contract in the data subject's interest between you and a third party; or
  • it benefits the data subject, consent is impractical to obtain, and they would likely consent if asked.

Which of those a given vendor setup satisfies is a legal question for your attorney, not a feature of any model, and this summary rests on a published text of the Act, not the Gazette copy.

While you wait for that answer, a cautious default is to keep personal information out of prompts. The other options to discuss with your adviser are a vendor agreement they have reviewed, or an open-weight model on your own infrastructure.

Governance is thin even in larger firms. The World Wide Worx findings reported by ITWeb found only 14% of the surveyed enterprises with a formal company-wide generative strategy and 13% with governance frameworks, so a one-page rule on what staff may paste is worth writing at any size.

Quick Answer: Can Customer Data Go Into a Hosted Model?

POPIA section 72 allows a transfer of personal information abroad only when a listed condition applies, such as adequate protection under a binding agreement or the data subject's consent. Whether a vendor's hosted service triggers it, and whether its terms meet a condition, is a question for your attorney.

How Often Do Vendors Retire or Reprice Their Lineups, and What Breaks?

Vendors retire and reprice models on their own published dates rather than a fixed cycle, so a workflow built on one model name needs a calendar entry for its retirement date. Anthropic's overview commits to keeping its current models available into late 2027: Claude Opus 5.5 not sooner than 22 September 2027, Sonnet 5.5 not sooner than 28 September 2027 and Fable 5.2 not sooner than 9 November 2027.

The same page lists Opus 4.1 and Opus 4 as retired on its own API. Google's model page shows Gemini 2.5 Flash and 2.5 Pro as deprecated, routing to newer models after 1 December 2026.

Prices move too: Google lists Gemini 4 Carbon and Argon at $2 / $10 through 30 November 2026 and $4 / $20 after, which is why the cost table uses the later price. Anthropic's model IDs are pinned snapshots, so a named ID does not change underneath you. Whichever vendor you use, rerun your twenty tasks before swapping a model, and budget on the post-promotion price.

Want a second opinion before you commit to a plan?

Send us your site and goals and we will audit where AI search fits and where it does not, with no obligation.

Request a free audit

Why South African Businesses Choose Growth Pulse Media

Growth Pulse Media is a Johannesburg agency that judges every channel by what it returns. Founder Dirk van Greuning built and scaled a large South African ecommerce business before starting the agency, so the advice comes from the operator's side of the desk. All work is executed in-house and the client load is limited, which keeps senior attention on each account.

For AI search specifically, our AEO agency page sets out what we do and for whom. Choosing a model is a tool decision; being found in the answers those tools give is a search-visibility problem, and that second problem is the one the service page covers.

Who This Is NOT For

You want a ranked "best model" leaderboard. This post compares published prices and vendor descriptions, not benchmark scores. Which model wins depends on your tasks, and only your own examples can settle it.

You need a legal opinion on data transfers. The POPIA section here summarises the conditions in section 72(1). Whether your vendor setup meets one is a question for your attorney.

You plan to train or self-host a model. We cover open-weight models only as an option. Hardware, licences and engineering belong with a technical specialist.

You want a promised placement in AI answers. Google says there are no special optimizations for its AI features, and nothing we read documents a way to buy a mention. Be wary of anyone who promises a result.

Frequently Asked Questions

What are AI models?

AI models are trained software systems that take a prompt and generate text, code, images or answers. ChatGPT, Claude and Gemini are apps built on model families from OpenAI, Anthropic and Google, and the same model can be reached through a subscription, an API or another product.

What is the difference between an AI model and ChatGPT?

The model is the trained system, such as GPT-6 Astra, while ChatGPT is the product that wraps a model with an interface, an account and billing. Vendors retire and reprice model versions on their own published dates.

How much do AI models cost in rand?

On a worked assumption of 2,000 tokens in and 1,000 out, 1,000 jobs cost about R11.49 on the smallest tiers, R229.88 on mid tiers such as GPT-6.1 Sol and Claude Sonnet 5.5, and R1,149.40 on GPT-6 Astra, at the index assumption of R16.42 to the US dollar. Prices are set in dollars, so the Rand cost moves with the exchange rate.

Which AI model is best for a small business in South Africa?

No single model is best; the cheapest tier that passes your own examples is. As a cost-first working rule, try a small tier for sorting and extraction and a mid tier for writing, and test about twenty real tasks before moving up. Anthropic itself suggests starting with Claude Opus 5.5 for most workloads, so let your own results decide.

Can I put customer data into AI models under POPIA?

POPIA section 72 allows a transfer of personal information to a third party in a foreign country only when a listed condition is met, such as adequate protection under a binding agreement or consent. Whether a vendor's hosted service triggers it, and whether its terms satisfy a condition, is a question for your attorney; a cautious default is to keep personal information out of prompts until it is answered.

Do I need to optimise my website for each AI model?

Google says there are no additional requirements to appear in AI Overviews or AI Mode, beyond meeting Search's technical requirements and publishing helpful, reliable content. That page covers Google only, so treat model-specific tricks for other assistants as unproven until a vendor documents them.

Get found where your customers now ask questions

Send us your website and your top customer questions, and we will tell you where answer-engine visibility is weakest. Our work covers Google Ads, Meta Ads, GA4 and Search Console from our Johannesburg base. No obligation: we will get back to you within 24 hours.

Talk to us

Dirk van Greuning — Founder, Growth Pulse Media
Dirk van Greuning Founder, Growth Pulse Media

Founder of Growth Pulse Media and a specialist in South African search dominance. Dirk translates his experience in scaling South African businesses into high-velocity digital strategies for B2B and retail leaders. He writes about SEO, lead generation, and paid media from an operator's perspective — prioritising pipeline value over impressions.

Connect on LinkedIn