LLMs 101

Monthly model tracker

How AI models rank when
real humans compare them

Benchmark scores don't tell the full story. This tracker translates human preference data and real-world usage into plain-English tiers — updated every month.

Updated 3rd October 2026
🏆
What is human preference ranking?
Real users are shown two anonymous AI responses to the same prompt and pick the one they prefer. Thousands of these battles produce an Elo-style ranking — the same system used to rank chess players.
💰
What does "cost vibe" mean?
Instead of listing raw dollar-per-token figures (which change constantly), we use plain-English cost tiers: Free, Low, Standard, and Premium — reflecting how expensive it is to run each model at real production scale.
📊
Why not show benchmark scores?
Benchmarks like MMLU or HumanEval measure narrow academic tasks. A score of "86.2%" means nothing in practice. Human preference rankings and real use-case guidance are far more useful for most readers.
🔄
How often is this updated?
Monthly. The AI landscape moves fast — a model released last month can jump several tiers. Check back each month, or read the Trends section for major shifts as they happen.
Show
How to read this table: Rankings reflect our monthly editorial review of current AI benchmarks and independent testing, combined with qualitative assessment of real-world professional use. Tier 1 = consistently preferred across that review. Tiers are relative — all listed models are genuinely excellent by any historical standard.
1
Anthropic
Claude Fable 5.1
Anthropic's most capable public model
🏆 Tier 1 - Top (near tie with GPT-6 Astra)
Best for
Long, complex agentic coding, research and knowledge work where getting it right matters more than cost. More than most people need for everyday chat.
Premium Closed API
2
OpenAI
GPT-6 Astra
OpenAI's most capable model
🥈 Tier 1 - Top (near tie with Fable 5.1)
Best for
The hardest reasoning, computer-use and document work, like building slides, spreadsheets and apps from start to finish. Expensive. In ChatGPT it powers GPT-6 Pro on higher plans.
Premium Closed API
3
Anthropic
Claude Opus 5.5
Near-Fable smarts at lower cost
🥉 Tier 1 - Best all-round flagship
Best for
The default pick for serious work: coding, analysis and long documents. Anthropic says it matches Fable 5.1 on most tasks for much less money.
Standard Closed API
4
OpenAI
GPT-6.1 Sol
Near-Astra performance, 1/5 the cost
Tier 2 - Best value OpenAI
Best for
Everyday professional and coding work in ChatGPT Work, Codex or the API when you want close to top quality without paying Astra prices. Not in regular ChatGPT chat yet.
Standard Closed API
5
Moonshot AI
Kimi K3
Largest open-weight frontier model
Tier 3 - Top open-weight
Best for
An open-weight alternative to US flagships for long coding, research and front-end work. At 2.8T parameters you use the hosted API, not a laptop. The license limits big resellers.
Standard Open weights
6
Anthropic
Claude Sonnet 5.5
Fast, cheaper everyday Claude
Tier 2 - Best value Claude
Best for
Clearly defined daily tasks: fixing bugs, drafting, and making documents, slides and spreadsheets. Over 30% faster than Sonnet 5 at half the token price of Opus 5.5.
Standard Closed API
7
Z.ai
GLM-5.3
Open-weight agentic coding leader
Tier 3 - Open-weight coding value
Best for
Agentic coding and long tasks that use many tools, at a fraction of US flagship prices. Text only and always reasons. The weights use a custom license, so check the terms.
Low Open weights
8
xAI
Grok 4.7
xAI flagship for code and agents
Tier 1 - Value-priced frontier
Best for
Coding agents and knowledge work for less than Claude or GPT flagships cost, with strong Cursor support. Independent tests show smaller gains than xAI claims.
Standard Closed API
9
Google DeepMind
Gemini 3.1 Pro
Google's top public model, now aging
Tier 1 - Google's best available (aging)
Best for
Multimodal reasoning over long documents, video and audio in Google's apps. Still a February preview. Gemini 4 Argon is out, but only for selected security partners.
Standard Closed API
10
Google DeepMind
Gemini 3.8 Flash
Google's smartest Flash model
Tier 2 - Best value Gemini
Best for
The best-value Google option for coding, agents and multi-step work, with a free tier in AI Studio. The introductory API price doubles on January 1, 2027.
Low Closed API
11
DeepSeek
DeepSeek V4.1 Flash
MIT-licensed, very cheap, multimodal
Tier 3 - Open-weight price disruptor
Best for
High-volume coding and agent work on a tight budget, or hosting it yourself under a permissive MIT license. Off-peak API prices are among the lowest anywhere.
Ultra-low Open weights
12
Alibaba Qwen
Qwen3.8-Max
Alibaba's 2.4T flagship
Tier 3 - Open-weight contender
Best for
Long coding and office tasks through Alibaba's low-cost API. The downloadable 2.4T version is text-only with a custom license. Qwen 4 is announced but not out yet.
Low Open weights
13
Meta
Muse Spark 1.3
Meta's closed agentic coder
Tier 3 - Emerging disruptor
Best for
Low-cost coding and agent work through Meta Model API or Muse Code, with a 1M-token context. Weights are closed, it is still labeled a public preview, and EU access is unconfirmed.
Low Closed API
14
Mistral AI
Mistral Medium 3.5
Europe's open-weight flagship
Tier 3 - EU open-weight pick
Best for
Teams that want an EU-made model they can host themselves on about four GPUs for coding and agent work. Benchmarks put it a step below the frontier leaders.
Low Open weights
15
OpenAI
GPT-6 Luna
OpenAI's cheapest, fastest GPT-6
Tier 4 - Budget OpenAI
Best for
High-volume jobs with a clear goal: summaries, data extraction, quick answers. Very cheap through the API. Free users get it in the desktop app; ChatGPT chat still uses GPT-5.6 Luna.
Ultra-low Closed API
16
Z.ai
GLM-5.3-Flash
Cheap open multimodal workhorse
Tier 4 - Budget open-weight
Best for
Budget agent and coding loops that take image and video input, with MIT-licensed weights you can host yourself. Great value, but independent tests have not confirmed the vendor's benchmarks.
Ultra-low Open weights
17
Google DeepMind
Gemini 3.5 Flash-Lite
Google's lowest-cost current model
Tier 4 - Budget Gemini
Best for
Sorting, data extraction, tagging, translation and small helper tasks at scale. Much better than 3.1 Flash-Lite, but not built for hard reasoning.
Ultra-low Closed API
18
Cohere
Command A+
Enterprise open model, 48 languages
Tier 3 - Enterprise specialist
Best for
Regulated or sovereign deployments that need multilingual search over your own documents, with citations, on hardware you control (fits on 2 H100 GPUs). Not a frontier coder.
Low Open weights
19
Alibaba Qwen
Qwen3.8-27B
Open model that fits on one GPU
Tier 3 - Best small local model
Best for
Running a capable model that handles images locally or privately on a single GPU under Apache 2.0, for coding and agent tasks. Well behind frontier models on hard work.
Free* Open weights
20
Meta
Muse Glimmer 30B
Meta's on-device open agent model
Tier 3 - Local and private
Best for
Private, offline agent tasks on a Mac or PC with one consumer GPU, under Apache 2.0. Good for local experiments, not for frontier-level reasoning.
Free* Open weights

* Llama models are free to download and run but require your own hardware or cloud compute. Self-hosting costs (electricity, GPU rental) vary. · Rankings reflect monthly editorial review of current AI benchmarks and independent testing, combined with editorial assessment — see our Methodology & Glossary page for full detail. · Cost tiers are illustrative and based on that same monthly review — actual pricing changes frequently; check each provider's current pricing page for exact figures. · This tracker focuses on general-purpose chat and reasoning models. Specialised models (image generation, audio, video) are not included.