# Atlas

> Atlas is a business-specific AI system being developed by Novo Studio.

It is designed to turn what a business knows into repeatable work toward the results the business defines. This website is public.

## Visiting agents

- Read the same public catalogue as a person. The catalogue is read-only.
- Do not submit forms, register anyone, or request an Atlas. This site does not accept agent submissions.

There is one Atlas. The walkthrough shows Atlas set up for one example business, a residential HVAC, plumbing, and electrical shop. The shop is an illustrative fixture, not a customer, a testimonial, or a price.

## Read

- [Home](https://www.atlas8.ai/): product story
- [Service walkthrough](https://www.atlas8.ai/workflows/service-business): the default business fixture
- [Intelligence Lab](https://www.atlas8.ai/community): Building together. Online Labs on Tuesdays and Thursdays at 11:00 Pacific Time, and in-person Labs.
- [Agents](https://www.atlas8.ai/agents) and [agents.md](https://www.atlas8.ai/agents.md)
- [Models](https://www.atlas8.ai/models) and [models.md](https://www.atlas8.ai/models.md)

## Models

Which model is best at what, checked daily against public leaderboards.

Atlas is being built to choose a model for each task; this page tracks the public evidence for that kind of choice.

Rankings reflect public leaderboards; they do not state which models Atlas runs.

Checked: 2026-10-13 ~14:15 PT.

## Intelligence and execution

1. Claude Opus 5.5 (Anthropic) — 94.1
2. Claude Fable 5.1 (Anthropic) — 82.4
3. Gemini 4 Argon (Google) — 81.7
4. Claude Sonnet 5.5 (Anthropic) — 77.5
5. GPT-6 Astra (OpenAI) — 73.6
6. GPT-6.1 Sol (OpenAI) — 71.9
7. Muse Spark 1.3 (Meta (Meta Superintelligence Labs)) — 53.4
8. GLM-5.3 (Zhipu AI (Z.ai)) — 38.1
9. Grok 4.7 (SpaceXSI) — 35.1
10. GPT-5.6 Terra (OpenAI) — 35.1
11. Gemini 3.8 Flash (Google) — 33.9
12. DeepSeek V4.1 Flash (DeepSeek) — 30.5
13. Qwen3.8 Max (Alibaba (Qwen)) — 29.6
14. Kimi K3 (Moonshot AI) — 29.3
15. GLM-5.3-Flash (Zhipu AI (Z.ai)) — 27.5
16. GPT-6 Luna (OpenAI) — 20.1

## Cost efficiency

Each estimate counts the task's input and output tokens at the company's list price, so real costs usually run higher.
Prices: official vendor pricing pages, checked Oct 13, 2026.

### Read a 20-page PDF and convert it to structured Markdown

GLM-5.3-Flash: $0.0085
GPT-6 Luna: $0.007
DeepSeek V4.1 Flash: $0.0186
Muse Spark 1.3: $0.0715
Gemini 3.8 Flash: $0.0525
Gemini 4 Argon: $0.14
Claude Sonnet 5.5: $0.14
GPT-6.1 Sol: $0.14
GLM-5.3: $0.0772
Claude Opus 5.5: $0.28
Grok 4.7: $0.108
Qwen3.8 Max: $0.108
GPT-5.6 Terra: $0.156
Kimi K3: $0.21
Claude Fable 5.1: $0.7
GPT-6 Astra: $0.7

### Update one recurring document from new notes

GLM-5.3-Flash: $0.0037
GPT-6 Luna: $0.003
DeepSeek V4.1 Flash: $0.0081
Muse Spark 1.3: $0.0315
Gemini 3.8 Flash: $0.0225
Gemini 4 Argon: $0.06
Claude Sonnet 5.5: $0.06
GPT-6.1 Sol: $0.06
GLM-5.3: $0.0342
Claude Opus 5.5: $0.12
Grok 4.7: $0.048
Qwen3.8 Max: $0.048
GPT-5.6 Terra: $0.066
Kimi K3: $0.09
Claude Fable 5.1: $0.3
GPT-6 Astra: $0.3

### Fix one bug in a mid-size codebase

GLM-5.3-Flash: $0.012
GPT-6 Luna: $0.009
DeepSeek V4.1 Flash: $0.0252
Muse Spark 1.3: $0.1005
Gemini 3.8 Flash: $0.0675
Gemini 4 Argon: $0.18
Claude Sonnet 5.5: $0.18
GPT-6.1 Sol: $0.18
GLM-5.3: $0.1104
Claude Opus 5.5: $0.36
Grok 4.7: $0.156
Qwen3.8 Max: $0.156
GPT-5.6 Terra: $0.192
Kimi K3: $0.27
Claude Fable 5.1: $0.9
GPT-6 Astra: $0.9

## Open-weight only

4 of the 16 ranked models are open-weight: GLM-5.3, Kimi K3, DeepSeek V4.1 Flash and GLM-5.3-Flash.

8. GLM-5.3 (Zhipu AI (Z.ai)) — 38.1
12. DeepSeek V4.1 Flash (DeepSeek) — 30.5
14. Kimi K3 (Moonshot AI) — 29.3
15. GLM-5.3-Flash (Zhipu AI (Z.ai)) — 27.5

## Best at

Top 3 among ranked models, plus the best open-weight model.

Coding: #1 Claude Fable 5.1 (97.7); #2 Claude Opus 5.5 (96.3); #3 GPT-6.1 Sol (96.0); best open-weight DeepSeek V4.1 Flash (85.4)
Agentic: #1 Claude Opus 5.5 (85.0); #2 GPT-6 Astra (82.4); #3 Claude Fable 5.1 (80.8); best open-weight DeepSeek V4.1 Flash (43.8)
Reasoning: #1 Gemini 4 Argon (95.1); #2 GPT-6 Astra (92.7); #3 GPT-6.1 Sol (92.6); best open-weight Kimi K3 (90.7)
Browser Use: #1 Claude Fable 5.1 (100.0); #2 Gemini 4 Argon (92.0); #3 GPT-6.1 Sol (88.0); best open-weight GLM-5.3-Flash (84.0)
Data Analysis: #1 Gemini 4 Argon (85.2); #2 GPT-6 Astra (83.0); #3 GPT-6.1 Sol (82.7); best open-weight DeepSeek V4.1 Flash (79.3)
Research: #1 Grok 4.7 (86.4); #2 DeepSeek V4.1 Flash (85.3); #3 GPT-6 Astra (85.2); best open-weight DeepSeek V4.1 Flash (85.3)
Design: #1 Claude Opus 5.5 (86.9); #2 Claude Sonnet 5.5 (85.0); #3 Claude Fable 5.1 (83.4); best open-weight GLM-5.3 (77.3)

## Method

Composite score: Artificial Analysis 35%, LMArena 25%, LiveBench Coding 20%, LiveBench Agentic 20%.
Ties go to the model with the higher LMArena low-end score.
LiveBench counts as one source, and a model needs all three sources to be ranked.
OpenRouter shows usage share only, not quality.
Share of tokens among listed models (OpenRouter, week of Oct 12)
Reasoning models also bill hidden reasoning tokens, which these assumptions do not include. Cached-input discounts are not applied.

## Sources

- Artificial Analysis: https://artificialanalysis.ai/leaderboards/models (2026-10-13 ~14:00 PT)
- LMArena: https://lmarena.ai/leaderboard/text (2026-10-13 (page date))
- LiveBench: https://livebench.ai/ (2026-10-13 ~14:15 PT)
- OpenRouter: https://openrouter.ai/rankings (week view, bucket 2026-10-12, read ~14:10 PT)

## Mistral Large 4 (Mistral AI)

Mistral AI released Mistral Large 4 in public preview on October 6, 2026, a general-purpose multimodal model with a mixture-of-experts design and listed context 1M.

- Availability: pilot
- Verification: verified
- Source: Mistral docs: Mistral Large 4
- Verified at: 2026-10-18T17:04:00Z
- Valid until: 2026-10-25T17:04:00Z
- Limitation: Public preview, not general availability.
- Limitation: Field news about a vendor release. It does not mean Atlas uses or integrates this model.
- [Tools](https://www.atlas8.ai/tools) and [tools.md](https://www.atlas8.ai/tools.md)
- [What's new](https://www.atlas8.ai/updates) and [updates.md](https://www.atlas8.ai/updates.md)
- [Catalogue JSON](https://www.atlas8.ai/catalog.json): same facts, with serve-time staleness
- [Build your Atlas](https://www.atlas8.ai/blueprint): human form only

## In-person Labs

- Wednesday, October 21 · 4:00–8:00 pm Miami time (EDT). Miami · venue details by email.
- Wednesday, October 21 · 4:00–8:00 pm New York time (EDT). New York City. 1 Hotel Central Park, 1414 Sixth Avenue, New York, NY 10019.
- Wednesday, October 28 · 4:00–8:00 pm Nashville time (CDT). Nashville · venue details by email.

## Do not assume

- No prices, customer counts, or testimonials are published.
- Results as a Service: Atlas is paid only when the results you define happen.
- [Register](https://www.atlas8.ai/lab/signup) opens a short form. Submitting it saves the person's interest and does not reserve a place.
- Model and tool records that are unconfirmed or expired are not current capabilities.

## People

- Meet the Novo Studio team to discuss business partnerships: https://calendly.com/michael-novostudio/30min
- Novo Studio: https://novostudio.io/
- Atlas: https://www.atlas8.ai