- Capability is our combined score. Epoch AI and Artificial Analysis each boil dozens of tests (maths, coding, science, reasoning and real work tasks) down to one index. We put both on Epoch's scale and average them. Higher is better. For the same model the two usually land within about 2.3 points of each other, so a smaller gap than that is a tie; ten points is a clear difference.
- One source so far marks a model so new that only one of the two has tested it. Artificial Analysis scores new models within days; Epoch takes a few weeks. The model is ranked on the score it has, and the other is averaged in as soon as it is published.
- Facts right is accuracy on Artificial Analysis's knowledge test: 6,000 hard questions across business, law, health, science and engineering, software and the humanities, answered from the model's own knowledge. It is the share it gets right.
- Makes things up comes from the same test. Of the questions a model did not get right, it is how often it gave a wrong answer instead of saying it did not know. Lower is better. A model can know a lot and still guess badly.
- Rand per million tokens is what the model costs through its API. A token is roughly three quarters of a word, so a million tokens is about 1,500 pages. “In” is what you send; “out” is what the model writes back.
- 1,000 messages is our worked example of a customer-service assistant: each message sends about 1 000 tokens (instructions, context and the question) and gets about 300back. Reasoning models can “think” before answering and bill that thinking as output, so hard tasks can cost several times more.
- Reads at once is the context window: how much text the model can take in at one time, in pages of about 500 words.
Best AI models · October 2026: Which AI model is best? The ZAIQ AI Benchmark
The best AI models on sale right now, ranked by two independent benchmarks combined into one score and turned into plain answers: which is the most capable, which makes things up least, what each costs in rand, and which to use for what. Updated every month.
The short answer
Most capable: Claude Opus 5.5 (166.7), with Claude Sonnet 5.5 and GPT-6 Astra within 2.3 points. That is inside the margin between our two sources, so they are effectively tied at the top.
Just released: Gemini 4 Argon, GPT-6.1 Sol, Grok 4.7, MiMo-V2.6-Pro and GPT-6 Luna are days or weeks old. Only one of our two sources has scored them so far, so they are ranked on that one and marked.
Best value: MiMo-V2.6-Pro scores within 10 points of the leader and costs about R11,41 to answer 1,000 customer messages, against R164 for Claude Opus 5.5.
Cheapest: DeepSeek V4.1 Flash, at about R3,39 per 1,000 messages. Free ChatGPT runs GPT-5.6 Luna.
Most reliable on facts: Claude Opus 5.5, with 66% of hard knowledge questions right. Least likely to make things up among the strong models: Qwen3.8 Max (0902). Best model you can run yourself: MiMo-V2.6-Pro.
- 166.7
- Top capability score, Claude Opus 5.5
- 121×
- Price gap between the cheapest and the most expensive model here
- 21
- Current models ranked, from two independent benchmark sources
The leaderboard
Ranked by capability: the average of two independent indexes, Epoch AI's and Artificial Analysis's, on one scale. Prices are what the model costs to use through an API, converted at US$1 = R16.39 (30 September 2026). Apps like ChatGPT and Claude charge a monthly plan instead.
- 1. Claude Opus 5.5Anthropic166.7Capability
- Facts right
- 66%
- Makes things up
- 59%
- 1,000 messages
- R164
- Reads at once
- about 1 500 pages
- Rand per million tokens (in / out)
- R65,56 / R327,80
- Scores behind it
- Epoch 167.3 · Artificial Analysis 57.6
- 2. Claude Sonnet 5.5Anthropic165.2Capability
- Facts right
- 54%
- Makes things up
- 47%
- 1,000 messages
- R82
- Reads at once
- about 1 500 pages
- Rand per million tokens (in / out)
- R32,78 / R163,90
- Scores behind it
- Epoch 165.2 · Artificial Analysis 56.0
- 3. GPT-6 AstraOpenAI165.0Capability
ChatGPT Pro (from R1,839 a month) uses it for Pro reasoning
- Facts right
- 63%
- Makes things up
- 51%
- 1,000 messages
- R410
- Reads at once
- about 1 550 pages
- Rand per million tokens (in / out)
- R163,90 / R819,50
- Scores behind it
- Epoch 166.5 · Artificial Analysis 52.7
- 4. Claude Fable 5.1Anthropic164.3Capability
- Facts right
- 67%
- Makes things up
- 73%
- 1,000 messages
- R410
- Reads at once
- about 1 500 pages
- Rand per million tokens (in / out)
- R163,90 / R819,50
- Scores behind it
- Epoch 164.8 · Artificial Analysis 53.4
- 5. Gemini 4 ArgonGoogle · one source so far163.3Capability
Not on general sale yet: Google has given it only to security partners in its Fairwind Program and has set no date for wider release
- Facts right
- 50%
- Makes things up
- 15%
- 1,000 messages
- no price
- Reads at once
- Rand per million tokens (in / out)
- no price
- Scores behind it
- Artificial Analysis 52.6 · Epoch not yet
- 6. GPT-6.1 SolOpenAI · one source so far162.9Capability
In ChatGPT it is in Work and Codex on paid plans, not yet in normal chat
- Facts right
- 62%
- Makes things up
- 54%
- 1,000 messages
- R82
- Reads at once
- about 1 550 pages
- Rand per million tokens (in / out)
- R32,78 / R163,90
- Scores behind it
- Artificial Analysis 51.8 · Epoch not yet
- 7. GPT-5.6 SolOpenAI161.1Capability
ChatGPT Plus and Pro run it for normal chat and thinking
- Facts right
- 59%
- Makes things up
- 92%
- 1,000 messages
- R82
- Reads at once
- about 1 550 pages
- Rand per million tokens (in / out)
- R32,78 / R163,90
- Scores behind it
- Epoch 161.8 · Artificial Analysis 47.0
- 8. Grok 4.7xAI · one source so far160.0Capability
- Facts right
- 48%
- Makes things up
- 29%
- 1,000 messages
- R62
- Reads at once
- about 750 pages
- Rand per million tokens (in / out)
- R32,78 / R98,34
- Scores behind it
- Artificial Analysis 46.4 · Epoch not yet
- 9. MiMo-V2.6-ProXiaomi · open weights · one source so far159.9Capability
- Facts right
- 35%
- Makes things up
- 41%
- 1,000 messages
- R11,41
- Reads at once
- about 1 550 pages
- Rand per million tokens (in / out)
- R7,13 / R14,26
- Scores behind it
- Artificial Analysis 46.3 · Epoch not yet
- 10. Muse Spark 1.3Meta158.9Capability
- Facts right
- 44%
- Makes things up
- 33%
- 1,000 messages
- R41
- Reads at once
- about 1 550 pages
- Rand per million tokens (in / out)
- R20,49 / R69,66
- Scores behind it
- Epoch 156.9 · Artificial Analysis 48.1
- 11. Kimi K3Moonshot AI · open weights158.0Capability
- Facts right
- 48%
- Makes things up
- 53%
- 1,000 messages
- R61
- Reads at once
- about 1 550 pages
- Rand per million tokens (in / out)
- R11,47 / R163,90
- Scores behind it
- Epoch 157.6 · Artificial Analysis 43.6
- 12. GLM-5.3Z.ai · open weights157.4Capability
- Facts right
- 34%
- Makes things up
- 30%
- 1,000 messages
- R27
- Reads at once
- about 1 550 pages
- Rand per million tokens (in / out)
- R4,92 / R72,12
- Scores behind it
- Epoch 155.8 · Artificial Analysis 44.8
- 13. Qwen3.8 Max (0902)Alibaba157.3Capability
- Facts right
- 32%
- Makes things up
- 29%
- 1,000 messages
- R62
- Reads at once
- about 1 500 pages
- Rand per million tokens (in / out)
- R32,78 / R98,34
- Scores behind it
- Epoch 155.2 · Artificial Analysis 45.4
- 14. Gemini 3.8 FlashGoogle157.0Capability
- Facts right
- 55%
- Makes things up
- 55%
- 1,000 messages
- R31
- Reads at once
- about 1 550 pages
- Rand per million tokens (in / out)
- R12,29 / R61,46
- Scores behind it
- Epoch 156.9 · Artificial Analysis 40.9
- 15. GPT-5.6 LunaOpenAI155.7Capability
ChatGPT Free runs it for unlimited everyday chats
- Facts right
- 43%
- Makes things up
- 93%
- 1,000 messages
- R9,18
- Reads at once
- about 1 550 pages
- Rand per million tokens (in / out)
- R3,28 / R19,67
- Scores behind it
- Epoch 156.4 · Artificial Analysis 37.3
- 16. DeepSeek V4.1 FlashDeepSeek · open weights155.6Capability
- Facts right
- 46%
- Makes things up
- 97%
- 1,000 messages
- R3,39
- Reads at once
- about 1 550 pages
- Rand per million tokens (in / out)
- R0,44 / R9,83
- Scores behind it
- Epoch 155.0 · Artificial Analysis 39.5
- 17. GPT-6 LunaOpenAI · one source so far155.5Capability
- Facts right
- 44%
- Makes things up
- 77%
- 1,000 messages
- R4,10
- Reads at once
- about 1 550 pages
- Rand per million tokens (in / out)
- R1,64 / R8,20
- Scores behind it
- Artificial Analysis 38.1 · Epoch not yet
- 18. DeepSeek V4 Pro 0813DeepSeek · open weights154.9Capability
- Facts right
- 49%
- Makes things up
- 95%
- 1,000 messages
- R41
- Reads at once
- about 1 550 pages
- Rand per million tokens (in / out)
- R21,63 / R64,90
- Scores behind it
- Epoch 155.4 · Artificial Analysis 36.0
- 19. Gemini 3.1 ProGoogle152.9Capability
- Facts right
- 55%
- Makes things up
- 51%
- 1,000 messages
- R92
- Reads at once
- about 1 550 pages
- Rand per million tokens (in / out)
- R32,78 / R196,68
- Scores behind it
- Epoch 154.8 · Artificial Analysis 29.7
- 20. Claude Haiku 4.5Anthropic143.2Capability
- Facts right
- 18%
- Makes things up
- 27%
- 1,000 messages
- R41
- Reads at once
- about 300 pages
- Rand per million tokens (in / out)
- R16,39 / R81,95
- Scores behind it
- Epoch 142.4 · Artificial Analysis 16.9
- 21. Mistral Medium 3.5Mistral AI · open weights141.9Capability
- Facts right
- 25%
- Makes things up
- 82%
- 1,000 messages
- R61
- Reads at once
- about 400 pages
- Rand per million tokens (in / out)
- R24,59 / R122,93
- Scores behind it
- Epoch 141.4 · Artificial Analysis 14.2
Download the full table: zaiq-ai-benchmark-2026-10.csv. Capability from Epoch AI (CC BY 4.0) and Artificial Analysis; the knowledge test from Artificial Analysis; prices from OpenRouter. All checked 1 October 2026. The download also has the models these replaced: Claude Opus 5, GPT-6 Sol, Grok 4.6 and Claude Sonnet 5.
What the numbers mean
Which model for which job
- The hardest work (complex analysis, long reports, difficult code): Claude Opus 5.5, Claude Sonnet 5.5 and GPT-6 Astra. They score within the margin of each other, so choose on price: Claude Sonnet 5.5 is the cheapest of them, at about R82 per 1,000 messages.
- A customer assistant or document Q&A: MiMo-V2.6-Prois the value pick. If a wrong answer is costly, choose on the “makes things up” column instead: Qwen3.8 Max (0902) is the most careful of the strong models. Either costs a fraction of the top models, which matters when thousands of messages come in.
- Keeping data on your own servers: an open-weights model such as MiMo-V2.6-Pro, or DeepSeek's V4 models, can run on infrastructure you control, which helps with POPIA's rules on sending personal information abroad. It takes engineering to run well.
- Everyday use for free: free ChatGPT runs GPT-5.6 Luna. See ChatGPT free vs paid in South Africa for what each plan costs in rand.
Picking the model is the easy part. The value comes from connecting it to your own data and processes, with a person checking anything that touches money or customers. That is what ZAIQ builds.
How the ZAIQ AI Benchmark is built
We do not run the tests ourselves. We combine two independent sources and show our working. Capability scores come from Epoch AI, a research group that publishes its results openly under a Creative Commons Attribution licence (Epoch AI, ‘Capabilities & benchmarking’), and from the Intelligence Index published by Artificial Analysis, which also runs the knowledge test behind “Facts right” and “Makes things up”. Prices and context windows come from OpenRouter's public model list, converted to rand at US$1 = R16.39. Everything was retrieved on 1 October 2026.
To put the two indexes on one scale, we fit a straight line through the 28models that both have scored and use it to convert Artificial Analysis's index to Epoch's scale. On those models the two agree closely: typically within 2.3 points, and by 5.5at most. A model's score is the average of the two. The newest models have one score so far and are marked: this month that is Gemini 4 Argon, GPT-6.1 Sol, Grok 4.7, MiMo-V2.6-Pro and GPT-6 Luna.
Our part is choosing the models on sale now, matching each one across the sources, working out what it costs in rand, and explaining it in plain English. When a newer version replaces a model, the old one moves to the download. The table is rebuilt every month from the same sources.
Questions about the benchmark
Which AI model is the best right now?
Claude Opus 5.5 scores highest on the ZAIQ AI Benchmark, 166.7. Claude Sonnet 5.5 and GPT-6 Astra are within 2.3 points of it, which is closer than our two sources usually agree, so treat them as tied at the top. Through its API it costs about R164 to answer 1,000 typical customer messages.
What is the best value AI model?
MiMo-V2.6-Pro. It scores 159.9, within 10 points of the leader, and costs about R11,41 per 1,000 customer messages at US$1 = R16.39. Only one of our two sources has scored it so far, so its score may move. For most business assistants that answer questions from your own documents, a model in this bracket is enough.
Which AI makes up the fewest facts?
Of the strongest models, Qwen3.8 Max (0902) is the least likely to invent an answer: when it does not know, it gives a wrong answer 29% of the time in Artificial Analysis's knowledge test. Claude Opus 5.5 has the best balance, with 66% of 6,000 hard questions right, but it still answers wrongly 59% of the time when it does not know. Every model makes things up, so connect it to your own documents or to search for anything that matters.
Is the free ChatGPT model any good?
Free ChatGPT runs GPT-5.6 Luna, which scores 155.7 against 166.7 for the leader. It is good enough for everyday writing and questions, but it guesses: when it does not know an answer it gives a wrong one 93% of the time. Through the API it costs about R9,18 per 1,000 messages. Paid plans unlock stronger models.
How much does it cost to run AI in a South African business?
For a customer-service assistant, the model itself is cheap: from about R3,39 to R410 per 1,000 messages on the models here, at US$1 = R16.39. The larger cost is building it properly into your systems. ZAIQ quotes that as one fixed price in rand.
Want the right model working in your business?
ZAIQ picks the model that fits the job and the budget, connects it to your own data and systems, and quotes one fixed price in rand before any work starts.
Tell us the job→