Model watch
Which AI models and coding agents lead right now
> as of 2026-10-15 · updated every month · back to AI Frontier
These tables come from public leaderboards, not from my own tests. Each leaderboard measures something different, so a model can rank high on one and lower on another. Use them to pick two or three models to try on your own work.
Frontier models
| Model | Maker | Intelligence Index | Arena text rank | Price per 1M tokens (in / out) |
|---|---|---|---|---|
| Claude Opus 5.5 | Anthropic | 58 | 3 | $4 / $20 |
| Claude Sonnet 5.5 | Anthropic | 56 | 36 | $2 / $10 |
| Muse Wonder | Meta | 56 | 2 | $4 / $20 |
| Gemini 4 Argon | 53 | 1 | $2 / $10 | |
| GPT-6 Astra | OpenAI | 53 | 30 | n/a |
| Claude Fable 5.1 | Anthropic | 53 | 6 | $10 / $50 |
| GPT-6.1 Sol | OpenAI | 52 | 22 | $2 / $10 |
As of 2026-10-15. Intelligence Index: Artificial Analysis (tests of reasoning, knowledge, math and coding; higher is better). Arena text rank: Arena (people vote on which of two answers they prefer; 1 is best). Prices as listed on Arena.
How to read it. The Intelligence Index combines many hard tests (reasoning, knowledge, math, coding); it shows raw ability. The Arena rank comes from millions of people choosing the better of two answers; it shows which answers people like. A model with a high index and a low Arena rank (like Claude Sonnet 5.5) is strong but less chatty; the reverse means people like its answers more than its test scores suggest. Scores use each model’s highest “effort” setting, which is slower and costs more.
Coding agents
A coding agent gets a task in plain English, then writes, runs and fixes code on its own until the task is done. Terminal-Bench 4.0 gives agents real tasks in a computer terminal and counts how many they finish.
| Rank | Agent + model | Tasks solved | Cost of the full test |
|---|---|---|---|
| 1 | Claude Code + Claude Opus 5.5 | 64.8% | $4.7k |
| 2 | Claude Code + Claude Sonnet 5.5 | 61.8% | $7.3k |
| 3 | Codex + GPT-6 Astra | 58.2% | $3.3k |
| 3 | Codex + GPT-6.1 Sol | 58.2% | $0.6k |
| 5 | Claude Code + Claude Fable 5.1 | 57.9% | $6.2k |
| 9 | Claude Code + GLM-5.3 (open model) | 41.8% | $2.7k |
| 10 | Grok Build + Grok 4.7 | 37.6% | $3.7k |
Source: Terminal-Bench 4.0 leaderboard, read 2026-10-15. Scores have a margin of about ±3 points.
What stands out. Claude Code leads, but Codex with GPT-6.1 Sol matches GPT-6 Astra for about a fifth of the cost. For most people, cost per solved task matters more than the top score.
Open models
An open-weight model can be downloaded and run on your own computers. That keeps data private and can cut costs, but needs technical setup. The best open models now rank close to last year’s frontier.
| Model | Maker | Arena text rank | License | Price per 1M tokens (in / out) |
|---|---|---|---|---|
| Kimi K3 | Moonshot | 18 | Kimi K3 license | n/a |
| GLM-5.3 | Z.ai | 28 | MIT | $1.40 / $4.40 |
| DeepSeek V4.1 Flash | DeepSeek | 38 | MIT | $0.30 / $1.20 |
Source: Arena text leaderboard, read 2026-10-15. Prices are what hosted services charge; running the model yourself costs only your hardware.
Which one should I use?
- Hard thinking and long documents: Claude Opus 5.5 or Gemini 4 Argon.
- Everyday work at a low price: GPT-6.1 Sol or Claude Sonnet 5.5.
- Coding: Claude Code or Codex (see the table above).
- Private data, no cloud: an open model such as GLM-5.3 or DeepSeek V4.1.
These are starting points, not verdicts: try two models on the same real task and compare.
Rankings change every few weeks. Notes for learning. Nothing here is investment advice.