# AI

> Which model to use and with how much effort: four tables of intelligence, AA cost and AA time for Claude, OpenAI/Codex and the API.

*Updated September 10, 2026.*

Which model to use, and with how much effort. Four tables: a Claude subscription while Fable lasts, the same once Fable runs out, an OpenAI/Codex subscription, and paying by API. Intelligence, cost and time are Artificial Analysis figures.

## Claude subscription, Fable available

| Usage | Model | Effort | Intelligence | AA cost per task | AA time |
|---|---|---|---|---|---|
| Fast | Claude Opus 5 | low | 39.8 | $1.098 | 179 s |
| Normal | Claude Opus 5 | medium | 45.1 | $2.189 | 359 s |
| Complex | Claude Fable 5.1 | medium | 49.1 | $2.983 | 325 s |
| Hard | Claude Fable 5.1 | high | 51.2 | $3.913 | 432 s |
| Maximum | Claude Fable 5.1 | xhigh | 53.2 | $5.978 | 648 s |
| Absolute maximum | Claude Fable 5.1 | max | 53.4 | $7.630 | 739 s |

Sweet spot: xhigh. Max adds only 0.2 intelligence for $1.652 more and 91 seconds longer.

## Claude subscription, Fable used up

| Usage | Model | Effort | Intelligence | AA cost per task | AA time |
|---|---|---|---|---|---|
| Mechanical work / saving quota | Claude Sonnet 5 | low | 24.7 | $0.509 | 148 s |
| Fast | Claude Opus 5 | low | 39.8 | $1.098 | 179 s |
| Normal | Claude Opus 5 | medium | 45.1 | $2.189 | 359 s |
| Hard | Claude Opus 5 | high | 48.2 | $3.613 | 557 s |
| Very hard | Claude Opus 5 | xhigh | 49.7 | $4.878 | 741 s |
| Maximum | Claude Opus 5 | max | 50.7 | $5.858 | 851 s |

## OpenAI / Codex subscription

| Usage | Model | Effort | Intelligence | AA cost per task | AA time |
|---|---|---|---|---|---|
| Volume / saving quota | GPT-5.6 Luna | xhigh | 34.8 | $0.085 | 218 s |
| Fastest | GPT-5.6 Sol | low | 33.8 | $0.261 | 69 s |
| Normal | GPT-5.6 Sol | medium | 39.5 | $0.505 | 125 s |
| Serious | GPT-5.6 Sol | high | 42.5 | $0.808 | 224 s |
| Hard | GPT-5.6 Sol | xhigh | 44.1 | $1.184 | 314 s |
| Maximum | GPT-5.6 Sol | max | 47.1 | $1.988 | 435 s |

*AA cost is comparative: a subscription burns quota, not dollars.*

## Paying by API: best intelligence per dollar

| Usage | Model | Effort | Intelligence | AA cost per task | AA time |
|---|---|---|---|---|---|
| Cheapest | GPT-5.6 Luna | xhigh | 34.8 | $0.085 | 218 s |
| Cheap | GPT-5.6 Luna | max | 37.5 | $0.178 | 375 s |
| Balanced | GPT-5.6 Sol | medium | 39.5 | $0.505 | 125 s |
| More capable | GPT-5.6 Sol | high | 42.5 | $0.808 | 224 s |
| Hard | GPT-5.6 Sol | xhigh | 44.1 | $1.184 | 314 s |
| Very hard | GPT-5.6 Sol | max | 47.1 | $1.988 | 435 s |
| Premium | Claude Fable 5.1 | medium | 49.1 | $2.983 | 325 s |
| High intelligence | Claude Fable 5.1 | high | 51.2 | $3.913 | 432 s |
| Best near-maximum | Claude Fable 5.1 | xhigh | 53.2 | $5.978 | 648 s |
| Absolute maximum | Claude Fable 5.1 | max | 53.4 | $7.630 | 739 s |

## Sources

- [Artificial Analysis, model comparison](https://artificialanalysis.ai/models)
- [Artificial Analysis, methodology](https://artificialanalysis.ai/methodology)
- [OpenAI, plan pricing and limits](https://learn.chatgpt.com/docs/pricing)
- [Anthropic, Fable models on your plan](https://support.claude.com/en/articles/15424964-claude-fable-models-on-your-plan)

---

Daniel Gil · daniel@gil.so
Canonical: https://gil.so/ai/
All pages: https://gil.so/llms.txt