Three families of AI. Pick by the job, not the headline.
Closed frontier models, Chinese and open-weight models, and models you run on your own hardware, compared on capability, price, licence, data and Thai. Every figure links to the page we read it on, on 3 October 2026. Prices and rankings change fast; check before you buy.
Rent the best, run something open, or keep it on your desk.
Family 1
Frontier, closed
The most capable models today, reached by API or through the big clouds. You never hold the weights; you rent the answers, under the vendor's data terms.
ForLow cost at scale; open weights mean you can self-host and keep data in-house.
AgainstHosted APIs fall under Chinese law; the biggest open models need serious servers.
Family 3
Local, on your hardware
Smaller open models that run on a laptop, a Mac or a single GPU. No per-token bill and nothing leaves the machine, at the cost of capability and upkeep.
Google · Gemma 4 (E2B to 31B)4-bit 12B needs about 6.7 GB; 4-bit 31B about 17.5 GB. Gemma 4 license allows commercial use.Gemma 4 docs, opened 3 Oct 2026 ↗
ForPrivate, works offline, no per-token cost, nobody can change the model on you.
AgainstHardware up front; well below frontier on hard tasks; you own updates and security.
Listed in no order of merit. Prices are list prices in US dollars, input / output per million tokens, standard tier, before batch or cache discounts. The lines blur: some US makers publish open models too, and Chinese open models are hosted by many providers outside China.
Side by side
Six questions, three honest answers.
Family
Capability tier
Price per 1M tokens (in / out)
Licence
Data control
Thai support
Hardware needed
Frontier, closed
Capability tierHighest. All of the top 12 on Artificial Analysis and the top 5 on LMArena this week.
Price per 1M tokens (in / out)About $2 / $10 to $10 / $50 for flagships; cheaper small tiers exist.
LicenceProprietary; API terms.
Data controlVendor's cloud. Some offer regional processing for 1.1× (Anthropic, US) or +10% (OpenAI).
Thai supportMultilingual; quality on Thai varies, so test with your own documents.
Hardware neededNone: a browser or an API key.
Chinese and open-weight
Capability tierHigh. Best open-licence models rank 27th to 38th on LMArena, 27 to 51 points behind the leader.
Price per 1M tokens (in / out)About $0.30 / $1.20 to $3 / $15 on hosted APIs.
LicenceOften open: MIT (DeepSeek, GLM), Apache 2.0 (Qwen), own licence (Kimi).
Data controlHosted: provider's servers (DeepSeek states China). Self-hosted: yours.
Thai supportVaries by model; Qwen is a common base for Thai fine-tunes. Test it.
Hardware neededNone for the API; multi-GPU servers to self-host the largest.
Local, on your hardware
Capability tierModerate. Good for drafting, sorting and search over your files; not for the hardest reasoning.
Price per 1M tokens (in / out)No per-token fee; you pay for the hardware and electricity.
LicenceOpen weights; each has its own terms (Apache, MIT, Gemma, Llama).
Data controlComplete: nothing leaves the machine.
Thai supportBest of the three for self-hosted Thai: Typhoon, OpenThaiGPT and THaLLE are open Thai models.
Hardware needed16 GB RAM to start; more memory runs bigger models.
All sources opened 3 October 2026. Thai support is our summary, not a measured score: no leaderboard we could open ranks Thai specifically, so test on your own text.
Leaderboard snapshot
Closed models lead, open ones are close behind.
LMArena text leaderboard, 2 October 2026
#1 Gemini 4 Argon HighGoogle
1525
#2 Claude Opus 4.6 HighAnthropic
1505
#3 Claude Fable 5 HighAnthropic
1504
#4 Claude Opus 5.5 HighAnthropic
1504
#5 Claude Opus 4.7 HighAnthropic
1501
#27 GLM 5.3 MaxZ.ai · MIT
1478
#32 GLM 5.2 MaxZ.ai · MIT
1476
#38 DeepSeek V4.1 Flash MaxDeepSeek · MIT
1474
ClosedOpen licence
Method: people compare two anonymous answers and vote; scores are Elo-style ratings from 8.6 million votes across 413 models. Bars start at 1,450, not zero, so the visual gap is exaggerated: the leader and the best open model are 47 points apart. LMArena text leaderboard, 2 Oct 2026, opened 3 Oct 2026 ↗
Artificial Analysis Intelligence Index, top five models
Claude Opus 5.5 (max)Anthropic
58
Claude Sonnet 5.5 (max)Anthropic
56
GPT-6 Astra (max)OpenAI
53
Gemini 4 Argon (high)Google
53
GPT-6.1 Sol (max)OpenAI
52
Method: a combined score from the site's own set of evaluations; read on 3 October 2026. All of its top 12 entries were proprietary on that day. This bar starts at zero. Artificial Analysis leaderboard, opened 3 Oct 2026 ↗
What runs on what
Memory decides the model size. Quantisation stretches it.
Tier 1
A 16 GB laptop
2–12B
LM Studio recommends 16 GB. At 4-bit, Gemma 4 E4B needs about 4.5 GB and 12B about 6.7 GB. Phi-4 (14B) sits at the edge. Fine for drafting and summaries.
Apple Silicon shares memory between CPU and GPU, and llama.cpp and LM Studio use it through Metal. Gemma 4 31B at full 16-bit needs about 70 GB. Check our arithmetic against the model card.
The big open models are huge: DeepSeek V4.1 Flash has 552B backbone parameters, Kimi K3 2.8T. Self-hosting them is a data-centre project, or rent the GPUs.
Memory figures are for the model weights only; long conversations need more on top. Sizes are rough classes, not guarantees: speed and quality depend on the exact model and settings.
Pick by job
What are you trying to do? Tap a job.
Choose a job above. The matching family card is highlighted, with our reasoning here.
Stanford's AI Index found open-weight models closed the gap with closed ones from 8% to 1.7% on some benchmarks in one year, and US and Chinese models reached near parity on major benchmarks in 2024.Stanford HAI, AI Index 2025, opened 3 Oct 2026 ↗
Tokenisers differ: Anthropic notes its newer models produce about 30% more tokens for the same text. Thai text often costs more tokens than English. Compare cost per finished task, not per token.Anthropic pricing, opened 3 Oct 2026 ↗
Blocked when we checked: openai.com/api/pricing returned an error, so OpenAI prices come from its developer docs. No figure on this page was estimated.
Still weighing it up for your own work?
We're builders who use models from all three families, and we have no vendor to push. Tell us about the job and the data, and we'll think it through with you.
