The AI benchmark · November 2026
Which AI model is best? Every leading model, ranked in rand
The models on sale right now, ranked by two independent benchmarks combined into one score, then turned into plain answers: which is most capable, which makes things up least, what each costs in rand, and which to use for what. Rebuilt every month.
- 173.5Top score, Claude Fable 5.2
- 160×Price gap between the cheapest and the dearest model here
- 22Current models ranked, from two independent sources
The leaderboard
Ranked by capability: the average of Epoch AI's and Artificial Analysis's indexes on one scale. Prices are what the model costs through its API, at US$1 = R16.56 (11 November 2026). Apps like ChatGPT and Claude charge a monthly plan instead.
-
1. Claude Fable 5.2
173.5Capability
- Facts right
- 79%
- Makes things up
- 59%
- 1,000 messages
- R331
- Reads at once
- about 1,500 pages
- Rand per million tokens (in / out)
- R132.48 / R662.40
- Scores behind it
- Epoch 174.2 · Artificial Analysis 64.6
-
2. Gemini 4 Carbon
170.6Capability
- Facts right
- 58%
- Makes things up
- 11%
- 1,000 messages
- R83
- Reads at once
- about 1,550 pages
- Rand per million tokens (in / out)
- R33.12 / R165.60
- Scores behind it
- Artificial Analysis 60.0 · Epoch not yet
-
3. Claude Opus 5.5
169.4Capability
- Facts right
- 66%
- Makes things up
- 50%
- 1,000 messages
- R166
- Reads at once
- about 1,500 pages
- Rand per million tokens (in / out)
- R66.24 / R331.20
- Scores behind it
- Epoch 168.7 · Artificial Analysis 59.1
-
4. GPT-6 Astra
169.1Capability
In ChatGPT it comes with the Pro 500 plan (R9,199 a month)
- Facts right
- 62%
- Makes things up
- 50%
- 1,000 messages
- R414
- Reads at once
- about 1,550 pages
- Rand per million tokens (in / out)
- R165.60 / R828.00
- Scores behind it
- Epoch 169.2 · Artificial Analysis 56.9
-
5. GPT-6.1 Sol
168.0Capability
In ChatGPT on the Plus and Pro plans; not on Go
- Facts right
- 62%
- Makes things up
- 52%
- 1,000 messages
- R83
- Reads at once
- about 1,550 pages
- Rand per million tokens (in / out)
- R33.12 / R165.60
- Scores behind it
- Artificial Analysis 54.6 · Epoch not yet
-
6. GPT-6.1 Terra
167.3Capability
- Facts right
- 48%
- Makes things up
- 35%
- 1,000 messages
- R41
- Reads at once
- about 1,550 pages
- Rand per million tokens (in / out)
- R16.56 / R82.80
- Scores behind it
- Epoch 168.1 · Artificial Analysis 51.3
-
7. Claude Sonnet 5.5
167.2Capability
- Facts right
- 52%
- Makes things up
- 43%
- 1,000 messages
- R83
- Reads at once
- about 1,500 pages
- Rand per million tokens (in / out)
- R33.12 / R165.60
- Scores behind it
- Epoch 166.3 · Artificial Analysis 54.5
-
8. Gemini 4 Argon
166.0Capability
On Gemini's paid plans
- Facts right
- 48%
- Makes things up
- 4%
- 1,000 messages
- R83
- Reads at once
- about 1,550 pages
- Rand per million tokens (in / out)
- R33.12 / R165.60
- Scores behind it
- Artificial Analysis 50.4 · Epoch not yet
-
9. Kimi K3.1 Pro
163.1Capability
- Facts right
- 49%
- Makes things up
- 37%
- 1,000 messages
- R83
- Reads at once
- about 1,550 pages
- Rand per million tokens (in / out)
- R33.12 / R165.60
- Scores behind it
- Epoch 162.0 · Artificial Analysis 46.5
-
10. Muse Wonder
162.4Capability
- Facts right
- 38%
- Makes things up
- 21%
- 1,000 messages
- R166
- Reads at once
- about 1,550 pages
- Rand per million tokens (in / out)
- R66.24 / R331.20
- Scores behind it
- Epoch 162.0 · Artificial Analysis 43.6
-
11. Claude Haiku 5.5
159.8Capability
- Facts right
- 35%
- Makes things up
- 34%
- 1,000 messages
- R4.14
- Reads at once
- about 1,500 pages
- Rand per million tokens (in / out)
- R1.66 / R8.28
- Scores behind it
- Epoch 160.3 · Artificial Analysis 36.5
-
12. Grok 4.7
158.9Capability
- Facts right
- 43%
- Makes things up
- 20%
- 1,000 messages
- R63
- Reads at once
- about 750 pages
- Rand per million tokens (in / out)
- R33.12 / R99.36
- Scores behind it
- Artificial Analysis 35.5 · Epoch not yet
-
13. DeepSeek V4.1 Flash 1017
158.4Capability
- Facts right
- 40%
- Makes things up
- 64%
- 1,000 messages
- R2.58
- Reads at once
- about 1,550 pages
- Rand per million tokens (in / out)
- R0.10 / R8.28
- Scores behind it
- Epoch 159.3 · Artificial Analysis 32.6
-
14. Gemini 3.8 Flash
157.7Capability
- Facts right
- 51%
- Makes things up
- 55%
- 1,000 messages
- R34
- Reads at once
- about 1,550 pages
- Rand per million tokens (in / out)
- R13.66 / R68.31
- Scores behind it
- Epoch 158.7 · Artificial Analysis 30.8
-
15. GLM-5.3
157.0Capability
- Facts right
- 32%
- Makes things up
- 39%
- 1,000 messages
- R8.11
- Reads at once
- about 1,550 pages
- Rand per million tokens (in / out)
- R1.16 / R23.18
- Scores behind it
- Epoch 156.9 · Artificial Analysis 31.7
-
16. Qwen3.8 Max (0902)
156.9Capability
- Facts right
- 30%
- Makes things up
- 27%
- 1,000 messages
- R63
- Reads at once
- about 1,500 pages
- Rand per million tokens (in / out)
- R33.12 / R99.36
- Scores behind it
- Epoch 155.8 · Artificial Analysis 33.7
-
17. MiMo-V2.6-Pro
156.2Capability
- Facts right
- 35%
- Makes things up
- 60%
- 1,000 messages
- R11.53
- Reads at once
- about 1,550 pages
- Rand per million tokens (in / out)
- R7.20 / R14.41
- Scores behind it
- Artificial Analysis 29.9 · Epoch not yet
-
18. GPT-6 Luna
155.2Capability
ChatGPT Free runs it for unlimited everyday chats
- Facts right
- 40%
- Makes things up
- 79%
- 1,000 messages
- R4.14
- Reads at once
- about 1,550 pages
- Rand per million tokens (in / out)
- R1.66 / R8.28
- Scores behind it
- Artificial Analysis 27.9 · Epoch not yet
-
19. Muse Spark Open
154.6Capability
- Facts right
- 42%
- Makes things up
- 28%
- 1,000 messages
- R42
- Reads at once
- about 1,550 pages
- Rand per million tokens (in / out)
- R20.70 / R70.38
- Scores behind it
- Epoch 155.4 · Artificial Analysis 24.9
-
20. Mistral Large 4
154.5Capability
- Facts right
- 26%
- Makes things up
- 28%
- 1,000 messages
- R22
- Reads at once
- about 1,550 pages
- Rand per million tokens (in / out)
- R11.59 / R34.78
- Scores behind it
- Epoch 154.6 · Artificial Analysis 26.1
-
21. DeepSeek V4 Pro 0813
154.4Capability
- Facts right
- 45%
- Makes things up
- 94%
- 1,000 messages
- R24
- Reads at once
- about 1,550 pages
- Rand per million tokens (in / out)
- R1.84 / R73.53
- Scores behind it
- Epoch 155.9 · Artificial Analysis 23.1
-
22. Mistral Medium 3.5
142.9Capability
- Facts right
- 23%
- Makes things up
- 80%
- 1,000 messages
- R62
- Reads at once
- about 400 pages
- Rand per million tokens (in / out)
- R24.84 / R124.20
- Scores behind it
- Epoch 141.6 · Artificial Analysis 4.6
| # | Model | Capability | Facts right | Makes things up | Rand per million tokens (in / out) | 1,000 messages | Reads at once |
|---|---|---|---|---|---|---|---|
| 1 | Claude Fable 5.2 Anthropic | 173.5 Epoch 174.2 · Artificial Analysis 64.6 | 79% | 59% | R132.48 / R662.40 | R331 | about 1,500 pages |
| 2 | Gemini 4 Carbon Google · one source so far | 170.6 Artificial Analysis 60.0 · Epoch not yet | 58% | 11% | R33.12 / R165.60 | R83 | about 1,550 pages |
| 3 | Claude Opus 5.5 Anthropic | 169.4 Epoch 168.7 · Artificial Analysis 59.1 | 66% | 50% | R66.24 / R331.20 | R166 | about 1,500 pages |
| 4 | GPT-6 Astra OpenAI In ChatGPT it comes with the Pro 500 plan (R9,199 a month) | 169.1 Epoch 169.2 · Artificial Analysis 56.9 | 62% | 50% | R165.60 / R828.00 | R414 | about 1,550 pages |
| 5 | GPT-6.1 Sol OpenAI · one source so far In ChatGPT on the Plus and Pro plans; not on Go | 168.0 Artificial Analysis 54.6 · Epoch not yet | 62% | 52% | R33.12 / R165.60 | R83 | about 1,550 pages |
| 6 | GPT-6.1 Terra OpenAI | 167.3 Epoch 168.1 · Artificial Analysis 51.3 | 48% | 35% | R16.56 / R82.80 | R41 | about 1,550 pages |
| 7 | Claude Sonnet 5.5 Anthropic | 167.2 Epoch 166.3 · Artificial Analysis 54.5 | 52% | 43% | R33.12 / R165.60 | R83 | about 1,500 pages |
| 8 | Gemini 4 Argon Google · one source so far On Gemini's paid plans | 166.0 Artificial Analysis 50.4 · Epoch not yet | 48% | 4% | R33.12 / R165.60 | R83 | about 1,550 pages |
| 9 | Kimi K3.1 Pro Moonshot AI · open weights | 163.1 Epoch 162.0 · Artificial Analysis 46.5 | 49% | 37% | R33.12 / R165.60 | R83 | about 1,550 pages |
| 10 | Muse Wonder Meta | 162.4 Epoch 162.0 · Artificial Analysis 43.6 | 38% | 21% | R66.24 / R331.20 | R166 | about 1,550 pages |
| 11 | Claude Haiku 5.5 Anthropic | 159.8 Epoch 160.3 · Artificial Analysis 36.5 | 35% | 34% | R1.66 / R8.28 | R4.14 | about 1,500 pages |
| 12 | Grok 4.7 SpaceXAI · one source so far | 158.9 Artificial Analysis 35.5 · Epoch not yet | 43% | 20% | R33.12 / R99.36 | R63 | about 750 pages |
| 13 | DeepSeek V4.1 Flash 1017 DeepSeek · open weights | 158.4 Epoch 159.3 · Artificial Analysis 32.6 | 40% | 64% | R0.10 / R8.28 | R2.58 | about 1,550 pages |
| 14 | Gemini 3.8 Flash Google | 157.7 Epoch 158.7 · Artificial Analysis 30.8 | 51% | 55% | R13.66 / R68.31 | R34 | about 1,550 pages |
| 15 | GLM-5.3 Z.ai · open weights | 157.0 Epoch 156.9 · Artificial Analysis 31.7 | 32% | 39% | R1.16 / R23.18 | R8.11 | about 1,550 pages |
| 16 | Qwen3.8 Max (0902) Alibaba | 156.9 Epoch 155.8 · Artificial Analysis 33.7 | 30% | 27% | R33.12 / R99.36 | R63 | about 1,500 pages |
| 17 | MiMo-V2.6-Pro Xiaomi · open weights · one source so far | 156.2 Artificial Analysis 29.9 · Epoch not yet | 35% | 60% | R7.20 / R14.41 | R11.53 | about 1,550 pages |
| 18 | GPT-6 Luna OpenAI · one source so far ChatGPT Free runs it for unlimited everyday chats | 155.2 Artificial Analysis 27.9 · Epoch not yet | 40% | 79% | R1.66 / R8.28 | R4.14 | about 1,550 pages |
| 19 | Muse Spark Open Meta · open weights | 154.6 Epoch 155.4 · Artificial Analysis 24.9 | 42% | 28% | R20.70 / R70.38 | R42 | about 1,550 pages |
| 20 | Mistral Large 4 Mistral AI · open weights | 154.5 Epoch 154.6 · Artificial Analysis 26.1 | 26% | 28% | R11.59 / R34.78 | R22 | about 1,550 pages |
| 21 | DeepSeek V4 Pro 0813 DeepSeek · open weights | 154.4 Epoch 155.9 · Artificial Analysis 23.1 | 45% | 94% | R1.84 / R73.53 | R24 | about 1,550 pages |
| 22 | Mistral Medium 3.5 Mistral AI · open weights | 142.9 Epoch 141.6 · Artificial Analysis 4.6 | 23% | 80% | R24.84 / R124.20 | R62 | about 400 pages |
Download the full table: ai-guide-ai-benchmark-2026-11.csv. Capability from Epoch AI (CC BY 4.0) and Artificial Analysis; the knowledge test from Artificial Analysis; prices from OpenRouter. All checked 11 November 2026. The download also has the models these replaced: Claude Fable 5.1, Claude Opus 5, GPT-5.6 Sol, Kimi K3, Muse Spark 1.3, Grok 4.6, GPT-5.6 Luna, Claude Sonnet 5, DeepSeek V4.1 Flash, Gemini 3.1 Pro and Claude Haiku 4.5.
The short version
- Most capable: Claude Fable 5.2 (173.5), clearly ahead.
- Just released: Gemini 4 Carbon, GPT-6.1 Sol, Gemini 4 Argon, Grok 4.7, MiMo-V2.6-Pro and GPT-6 Luna have been scored by one of the two sources so far, so they are ranked on that score and marked.
- Best value: GPT-6.1 Terra scores within 10 points of the leader and costs about R41 per 1,000 customer messages, against R331 for Claude Fable 5.2.
- Cheapest: DeepSeek V4.1 Flash 1017, at about R2.58 per 1,000 messages. Free ChatGPT runs GPT-6 Luna.
- Most reliable on facts: Claude Fable 5.2, with 79% of hard knowledge questions right. Least likely to make things up among the strong models: Gemini 4 Argon.
- Best model you can run yourself: Kimi K3.1 Pro, an open-weights model.
Which model for which job
- The hardest work (complex analysis, long reports, difficult code): Claude Fable 5.2. It costs more than most, so use it where quality pays for itself.
- A customer assistant or document Q&A: GPT-6.1 Terra is the value pick. If a wrong answer is costly, choose on the "makes things up" column instead: Gemini 4 Argon is the most careful of the strong models.
- Keeping data on your own servers: an open-weights model such as Kimi K3.1 Pro can run on infrastructure you control, which can help with POPIA's rules on sending personal information abroad. It takes engineering to run well. See can you put customer data into ChatGPT.
- Everyday use for free: free ChatGPT runs GPT-6 Luna. See ChatGPT free vs paid in South Africa for every plan in rand, and Claude pricing in South Africa.
What the numbers mean
- Capability is the combined score. Epoch AI and Artificial Analysis each boil dozens of tests (maths, coding, science, reasoning and real work tasks) down to one index. Both are put on Epoch's scale and averaged. Higher is better. For the same model the two usually land within about 1.7 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 scored it. It is ranked on that score until the other is published.
- Facts right is accuracy on Artificial Analysis's knowledge test: thousands of hard questions across business, law, health, science and engineering, software and the humanities, answered from the model's own knowledge.
- 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.
- Rand per million tokens is what the model costs through its API. A token is about three quarters of a word, so a million tokens is roughly 1,500 pages. "In" is what you send; "out" is what the model writes back.
- 1,000 messages is a worked example of a customer-service assistant: each message sends about 1,000 tokens and gets about 300 back. Reasoning models can bill their thinking as output, so hard tasks can cost several times more.
- Reads at once is the context window: how much text the model takes in at one time, in pages of about 500 words.
How this ranking is built
AI Guide does not run the tests. It combines two independent sources and shows the working. Capability scores come from Epoch AI, which publishes its results 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 at US$1 = R16.56. Everything was retrieved on 11 November 2026.
To put the two indexes on one scale, a straight line is fitted through the 17 models that both have scored, and used to convert Artificial Analysis's index to Epoch's scale. On those models the two agree closely: typically within 1.7 points, and by 2.9 at most. A model's score is the average of the two.
When a newer version replaces a model, the old one moves to the download. Until November 2026 this ranking was published on zaiq.ai as the Zaiq AI Benchmark; earlier editions are kept: October 2026 and September 2026. For what the individual tests measure and where they mislead, read AI benchmarks explained.
Questions about the ranking
Which AI model is the best right now?
Claude Fable 5.2 scores highest, 173.5 on the combined index. It leads clearly: the next model is more than the usual 1.7-point gap between the two sources behind it. The ranking changes every few weeks, so check the date.
What is the best value AI model?
GPT-6.1 Terra. It scores 167.3, within 10 points of the leader, and costs about R41 per 1,000 customer messages at US$1 = R16.56, against R331 for Claude Fable 5.2.
Which AI makes up the fewest facts?
Of the strong models, Gemini 4 Argon: when it does not know an answer, it gives a wrong one 4% of the time in Artificial Analysis's knowledge test. Claude Fable 5.2 gets the most hard knowledge questions right, 79%.
Is the free ChatGPT model any good?
Free ChatGPT runs GPT-6 Luna, which scores 155.2 against 173.5 for the leader. It is fine for everyday writing and questions, but when it does not know an answer it gives a wrong one 79% of the time, so check anything important.
How much does it cost to run AI in a South African business?
The model itself is cheap. For a customer-service assistant it runs from about R2.58 to R414 per 1,000 messages on the models here, at US$1 = R16.56. The bigger cost is building it properly into your own systems.
How often is this ranking updated?
Every month, from the same public sources. This edition was built on 11 November 2026. New models are added as soon as either source scores them.