Translation tables for AI tools
Every provider calls the same thing something else
Two translation tables. The first translates the concepts you work with in an AI tool. The second translates the model tiers underneath. Both contain names, not checkmarks. Where a cell stays empty, the provider lacks the concept.
Terms
What you do in the tool
Row 8 is where people talk past each other in meetings most often. Three providers call the preconfigured assistant from row 2 an agent, two use the word for the autonomous execution in row 8, and one does not use it at all.
| Concept | ChatGPT | Claude | Gemini | Copilot | Vibe | Langdock |
|---|---|---|---|---|---|---|
| 1Workspace for a project A separate space with its own files and its own context | Projects | Projects | Notebooks | Copilot Notebooks | Projects | Projects |
| 2Preconfigured assistant Fixed instructions, stored knowledge, shareable | GPTs → Plugins 1 | shared Project | Gems | Agents | Agents | Agents |
| 3Reusable capability Instructions the model loads by itself when needed | Skills | Skills | Spark only 2 | Skills Cowork only | Skills | Skills |
| 4Global preferences Details about yourself and how you work, applied in every chat | Custom Instructions | Instructions for Claude | Instructions for Gemini | Custom instructions | Custom instructions | Custom instructions 3 |
| 5Memory across chats Remembering beyond a single conversation | Memory | Memory | Memory | Copilot memory | Knowledge Base | Memory 3 |
| 6Connecting external systems Access to other services and data sources | Connected apps MCP | Connectors MCP | Connected Apps MCP | Copilot connectors MCP | Connectors MCP | Integrations MCP |
| 7Run on a schedule Something runs automatically on a schedule | Scheduled tasks | Scheduled tasks | Scheduled actions | Scheduled prompts | Scheduled tasks | Scheduled tasks |
| 8Autonomous execution Multi-step work, including the browser or the computer | ChatGPT Work | Claude Cowork | Gemini Spark 2 | Copilot Cowork | Vibe Work | Workflows 4 |
| 9Deep research with sources A long research job with citations | Deep research | Research | Deep Research | Researcher | Deep Research | Deep Research |
| 10Working document next to the chat A shared surface for writing and editing | Writing blocks | Artifacts | Canvas | Copilot Pages | Canvas | Document Editor |
| Coding agent A separate product alongside, not part of the chat | Codex | Claude Code | Antigravity | GitHub Copilot | Vibe Code | none |
- 1Custom GPTs are being replaced by Plugins, shutdown on 11 Dec 2026. An end date for creating new ones is only given for Enterprise workspaces: 26 Oct 2026.
- 2Gemini Spark, and with it Skills, is not available in the European Economic Area.
- 3In Langdock, Memory and Custom instructions do not apply in Projects and Agent chats.
- 4Langdock Workflows is a visual builder sold as a paid add-on, explicitly human-in-the-loop, with no browser or computer control.
Each column reflects the paid single-seat plan. The Microsoft column uses a work account. In companies, features are often available but locked by the administrator. Langdock has no single-seat plan, workspaces start at 25 euros per seat.
Model tiers
The model tiers underneath
Ten providers from three regions. Each has a fast tier, a middle tier and a top tier, and almost every one names them differently. A fourth tier above the top exists only in the USA so far. Reasoning is no longer a separate model anywhere, it is a toggle everywhere.
| USA | Europe | China | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Tier | OpenAI | Anthropic | Meta | Mistral | DeepSeek | Qwen | Kimi | GLM | MiniMax | |
| 1Fast, low-cost tier For volume and short answers | GPT-6 Luna | Haiku 4.5 | Gemini 3.5 Flash-Lite | Muse Glimmer 30B | Ministral 3 | DeepSeek-V4.1-Flash | Qwen3.8-Flash | none | GLM-5.3-Flash | M2.7-highspeed |
| 2Standard tier The workhorse in between | GPT-6 Sol | Sonnet 5 | Gemini 3.8 Flash | none | Mistral Small 4 | none | Qwen3.7-Plus | Kimi K2.6 | GLM-5.2 | MiniMax-M2.7 |
| 3Top tier What the provider sells as its best model | GPT-6 Astra 3 | Opus 5.5 4 | Gemini 3.1 Pro | Muse Spark 1.3 | Mistral Medium 3.5 | DeepSeek-V4-Pro | Qwen3.8-Max | Kimi K3 | GLM-5.3 | MiniMax-M3 |
| 4One tier above that So far only in the USA | none | Fable 5.1 plus Mythos 5.1 | Deep Think mode, Ultra only | |||||||
| 5Reasoning No longer a separate model anywhere | Toggle | Toggle | Toggle | Toggle | Toggle | Toggle | Toggle | Toggle | Toggle | Toggle |
| 6Open weights And under which license | side line only | no | side line only | small ones only Apache 2.0 | almost all Apache 2.0 | MIT unchanged | mixed 27B Apache 2.0 | own license | own license 5.2 was MIT | Community License M2.7 non-commercial |
| 7Hosting in the EU Offered by the provider itself | no | no | no | no | yes French provider | no | yes, Frankfurt 1 | no | no 2 via Mistral | no |
- 1Alibaba Model Studio has a Germany (Frankfurt) region, eu-central-1. The provider states: Request data is stored in that region.
- 2Mistral hosts GLM 5.2 and 5.3 itself, in its own words hosted by Mistral. Z.ai does not confirm this on any of its own pages.
- 3Until 22 Sep 2026, GPT-6 Astra sat one tier above the top. Since OpenAI filled the GPT-6 line below it with Sol and Luna (Terra is gone), Astra is the regular top tier.
- 4Opus 5.5 has replaced Opus 5 since 22 Sep 2026. According to Anthropic it matches Fable 5.1 on most tasks and costs 40 percent less to run.
Microsoft, Langdock and ByteDance are missing. The first two do not build their own models, they pass through other providers' models. ByteDance has no open weights and no international access route.
Word traps
Where the same word means different things
These six spots cause most of the misunderstandings. They are the real content of both tables.
Agent
At Microsoft, Mistral and Langdock, an agent is the preconfigured assistant, what OpenAI calls a GPT. At Google, and until recently at OpenAI, an agent is autonomous execution. Anthropic does not use the word as a product name at all, only as an adjective.
Cowork
Two providers picked the same name independently. Claude Cowork at Anthropic, Copilot Cowork at Microsoft. Same word, same row, different products.
Canvas
At Gemini and Vibe, Canvas is the working document next to the chat. OpenAI replaced it with Writing blocks, Langdock renamed it Document Editor. At OpenClaw the name stands for a display-only panel you cannot write in.
Skill
The term converging fastest, but with very different reach. At Google, Skills exist only in Spark, at Microsoft only in Cowork. Same name, a quarter of the reach.
Large
At Mistral, since Mistral 3, Large describes model size, not the top tier. Large 3 is bigger than Medium 3.5 and costs a fifth of it.
The higher version number
At Google, the tiers are decoupled. Flash is at 3.8, Pro at 3.1. Here the higher number belongs to the weaker tier.
Every cell checked individually on a provider page.
In three months, parts of these tables will be wrong. The concepts in the left column stay, the names in the cells move.