Verdict
Frontier models insights
Model
Picker
Use insights that aggregate models intelligence/$ across many benchmarks, to pick the best model for your application/use case
Latest Data Snapshot: 2026-10-13
Decision workspace
Model Matchup
Compare outcome quality, the total cost to reach it, and end-to-end latency.
Your decision, simplified
Pick one model. Add a challenger if you need one.
Your baseline unlocks tailored frontier recommendations. The second field is optional and opens the full quality, cost, and latency breakdown.
01 Model I currently use / comparison baselineRequired
⌄
02 ChallengerOptional
⌄
Model Picker
Which LLM gives the most intelligence for the money — the live Pareto frontier.
Get told when it changes
Watch the model you use
Models get beaten within weeks. Run this once a day and you get told the moment the model you are using right now is beaten: 200 while it still holds the frontier, 409 when something gives more intelligence for the money.
Pick a baseline in the matchup above to watch your own model. Full cron script and status codes
For agents
Give your coding agent the skill
An agent skill that makes sure every time your agent chooses an LLM for your projects, it picks the best intelligence-per-dollar model. Save cost and get the most intelligence.
Current landscape
Filters & axes
Adjust the measures and cohort. The chart updates instantly.
Input modality:
Most attractive quadrant
Pareto line
Open-weights frontier
No models match the current chart filters.
The log axis is centered on the median cost of the displayed models, matching AA’s centering rule. Equal distances represent equal cost ratios. The green quadrant remains relative to the displayed cohort.
Frontier table
The 17 LLMs that give the most intelligence for the money (2026-10-13)
Of the 692 models in the dataset, 17 are on the intelligence-vs-cost Pareto frontier: no other model is both smarter and cheaper to run. The cheapest frontier model is GPT-6 Luna (Low) (intelligence 21.50 for $10.6 per benchmark run); the smartest is Claude Opus 5.5 (Max, Default Fallback) (intelligence 57.60, $8708). The smartest open-weights (self-hostable) model on its own frontier is MiMo-V2.6-Pro by Xiaomi (intelligence 46.30, $207). Every model name links to a page with its numbers, whether it holds the frontier, and what beats it.
Overall frontier: cheapest to smartest
| # | Model | Creator | Intelligence | Total cost | E2E total | Open | Released |
|---|---|---|---|---|---|---|---|
| 1 | GPT-6 Luna (Low) | OpenAI | 21.50 | $10.6 | 6.52s | 2026-09-22 | |
| 2 | GPT-6 Luna (Medium) | OpenAI | 29.90 | $31.2 | - | 2026-09-22 | |
| 3 | GPT-6 Luna (High) | OpenAI | 32.90 | $47.9 | 23.4s | 2026-09-22 | |
| 4 | Claude Haiku 5.5 (Medium) | Anthropic | 34.50 | $56.4 | 15.7s | 2026-10-07 | |
| 5 | GPT-6 Luna (Xhigh) | OpenAI | 34.60 | $66.8 | 45.6s | 2026-09-22 | |
| 6 | Claude Haiku 5.5 (High) | Anthropic | 37.80 | $94.1 | 27.4s | 2026-10-07 | |
| 7 | MiMo-V2.6-Flash | Xiaomi | 37.90 | $110 | 54.0s | 2026-09-21 | |
| 8 | GPT-6 Luna (Max) | OpenAI | 38.10 | $122 | 124.0s | 2026-09-22 | |
| 9 | Claude Haiku 5.5 (Xhigh) | Anthropic | 41.20 | $158 | 81.8s | 2026-10-07 | |
| 10 | MiMo-V2.6-Pro | Xiaomi | 46.30 | $207 | 55.9s | 2026-09-21 | |
| 11 | GPT-6.1 Sol (Medium) | OpenAI | 47.80 | $361 | 16.7s | 2026-09-29 | |
| 12 | GPT-6.1 Sol (High) | OpenAI | 50.20 | $521 | 68.4s | 2026-09-29 | |
| 13 | GPT-6.1 Sol (Xhigh) | OpenAI | 51.00 | $662 | 186.1s | 2026-09-29 | |
| 14 | GPT-6.1 Sol (Max) | OpenAI | 51.80 | $1082 | 338.3s | 2026-09-29 | |
| 15 | Claude Opus 5.5 (High, Default Fallback) | Anthropic | 53.60 | $2172 | 58.7s | 2026-09-22 | |
| 16 | Claude Opus 5.5 (Xhigh, Default Fallback) | Anthropic | 56.00 | $4057 | 165.5s | 2026-09-22 | |
| 17 | Claude Opus 5.5 (Max, Default Fallback) | Anthropic | 57.60 | $8708 | 722.0s | 2026-09-22 |
Open-weights frontier: the best self-hostable LLMs for the money
| # | Model | Creator | Intelligence | Total cost | E2E total | Open | Released |
|---|---|---|---|---|---|---|---|
| 1 | Ministral 3 3B | Mistral | 4.80 | $12.2 | 2.95s | 2025-12-02 | |
| 2 | Granite 4.2 3B | IBM | 9.10 | $14.9 | 12.4s | 2026-08-25 | |
| 3 | DeepSeek V3 0324 | DeepSeek | 9.70 | $19.9 | 7.59s | 2025-03-25 | |
| 4 | Mistral Small 4 (Reasoning) | Mistral | 11.30 | $42.5 | 16.6s | 2026-03-16 | |
| 5 | Nemotron 3.5 Lightning | NVIDIA | 12.90 | $105 | 8.90s | 2026-08-11 | |
| 6 | MiMo-V2.6-Flash | Xiaomi | 37.90 | $110 | 54.0s | 2026-09-21 | |
| 7 | MiMo-V2.6-Pro | Xiaomi | 46.30 | $207 | 55.9s | 2026-09-21 |
All 692 models compared How the frontier got here How fast frontier models get replaced This table as Markdown
How it works
How Model Picker ranks LLMs
What is the LLM Pareto frontier?
The set of models no other model beats on both axes at once. On the default chart that means no model is both more intelligent and cheaper to run. Everything else is dominated: something gives you more intelligence for the same money, or the same intelligence for less.
Why total cost to run the benchmark instead of price per token?
Per-token prices hide how many tokens a model spends, and reasoning models spend a lot. Model Picker uses the total cost Artificial Analysis reports for running its full Intelligence Index evaluation, so a cheap-per-token model that thinks for pages is priced honestly.
What does the intelligence index measure?
The Artificial Analysis Intelligence Index, a composite of reasoning, knowledge, math, coding, and agentic benchmarks. Model Picker takes the score as published and never re-scores models.
Which is the best open-weights LLM for the money?
The open-weights frontier in the chart's table tab lists the models with downloadable weights that nothing self-hostable beats on both intelligence and cost, so any row is a best-value choice at its price. Some open-weights licenses restrict commercial use.
Is the model I use still the best value?
Type it into the matchup above: the better-than guidance names the best frontier model at the same or lower cost and the next step up. Agents and cron jobs can ask the same question with one request to the API and get a 409 status the day the model is beaten.
How fresh is this data?
The dataset is refreshed every six hours from Artificial Analysis & other sources; this page was rendered from the 2026-10-13 snapshot. Metrics move between snapshots as models are re-measured, so a quoted number should carry its snapshot date.