Part of AI and Research Practice.
What you can do, and where
What you can do with a model depends more on where you use it than on which company made it. Each row is a move from the opening talk. Each column is a place to make it.
What the columns mean
- Chat window. ChatGPT, Claude, Gemini, or Copilot in a browser, the desktop app, or your phone. The company wraps the model in memory, web search, and a system prompt you never see.
- Agent in the terminal. Claude Code, Codex, or Gemini CLI, run from the terminal. You give it a task in plain English, and the model works inside a folder on your computer. It reads and writes files, runs code, searches the web, and calls outside services until the task is done.
- Agent in the desktop app. The same agents in an ordinary window: the Code tab in the Claude desktop app, and Codex mode in the ChatGPT desktop app. They can do nearly everything the terminal can. The one thing only the terminal can do is run a prompt over and over by itself.
- API in your script. A Python or R script that sends a prompt to the model and saves what comes back. The demos were run this way.
- Model on your laptop. An open-weight model you download and run with Ollama or LM Studio. Nothing leaves your computer, the model never changes, and you can see the probabilities. These models are weaker than the companies’ current ones.
| Chat | Agents | You write the code | |||
|---|---|---|---|---|---|
| What you want to doWhere it comes up in the hub | Chat windowChatGPT, Claude, Gemini, Copilot | Agent in the terminalClaude Code, Codex CLI, Gemini CLI | Agent in the desktop appClaude app (Code tab), ChatGPT app (Codex mode) | API in your scriptPython or R calling OpenAI, Anthropic, or Google | Model on your laptopOpen-weight models in Ollama or LM Studio |
| How it works | |||||
| See the word probabilitiesFill in the blank | Hidden in every app | Only through a script it writes, so the same models as the API | Only through a script it writes, so the same models as the API | Only older OpenAI models like gpt-4o, or smaller ones with reasoning off. Never Claude or Gemini 3† | Ollama and LM Studio return them |
| Set the temperatureTemperature | No app has the setting | Only through a script it writes, with the API’s limits | Only through a script it writes, with the API’s limits | Only older models, or OpenAI’s smaller ones with reasoning off | Any value |
| Run one prompt 20 or 100 timesWhere would you apply it? | By hand, one new chat at a time | claude -p in a loop, each run a new session |
No repeat mode. It can write a script that loops | A loop in your code | A loop, with no cost per run |
| Tally the answersWhere would you apply it? | Paste them in and it runs Python | It writes and runs the code | It writes and runs the code | You write the code | You write the code |
| Context | |||||
| Attach a paperAttaching the source | Upload it | Reads it from your folder | Reads it from your folder | PDFs, at all three companies | Usually as extracted text† |
| Work across a whole project folderLab 3 | Uploaded copies in a Project. Desktop apps can connect a folder | Reads and writes your files in place | In place. The web versions work on a GitHub copy | Your code loads the files | Not unless you plug it into an agent |
| See and control memory and instructionsSame question, different models | On by default, set in Settings. Not on msu.edu Gemini | Plain-text files you can read and edit | The same files as the terminal | You write the system prompt. Nothing carries over | You write the system prompt |
| Search the webWhen the product is doing the work | Built in | Built in | Built in | A tool you turn on | Not unless you wire it up |
| Verification | |||||
| Check against a record like CrossrefA fabricated citation | By web search, one at a time | Calls Crossref or OpenAlex directly | Calls Crossref or OpenAlex directly | You write the code | You write the code |
| Re-run cold, in a fresh context“Are you sure?” | Only a temporary or incognito chat. A new chat still reads memory | claude -p --bare skips memory and instruction files |
A new session still loads memory and instruction files | Every call starts fresh | Every call starts fresh |
| Vary the modelSame question, different models | One company’s models | One company’s. Claude Code and Codex can run open models | One company’s models | Any company, with a key for each | Many open models, all weaker than the frontier |
| Red team with many agentsAgents | Deep Research is built in, not yours to design | Subagents defined in plain-text files | The same subagents. Agent teams are terminal only | Agent SDKs, with lots of code | Not unless you plug it into an agent |
| Record and rules | |||||
| Pin a dated model versionReproducibility and reporting | The model changes under the same name | claude --model pins. Codex and Gemini CLI partly |
Not checked† | Claude yes. OpenAI and Gemini partly | The weights you download never change |
| Licensed by MSU | Copilot and Gemini free with an MSU login. ChatGPT Edu bought by department. Claude not on MSU’s list | Codex comes with ChatGPT Edu, local use only. Claude Code and Gemini CLI not on MSU’s list | Claude not on MSU’s list. Codex mode unclear† | OpenAI, by separate request and billed to a department. Claude not on MSU’s list | No license needed |
| Data MSU allows in it | Licensed tools: up to confidential research, with permissions. Personal accounts and Claude: public data only† | Codex: same as licensed chat. Claude Code and Gemini CLI: public data only† | Claude: public data only. Codex mode unclear† | OpenAI through MSU: up to confidential research. Personal keys, including Anthropic: public data only† | Data stays on your laptop, but MSU hasn’t ruled on local models |
Licensed and allowed for your data are different questions. No tool, licensed or not, may take HIPAA, payment card, CUI, or export-controlled data. For confidential research data, MSU also requires anonymization, approval from MSU IT, and written consent.
Notes and sources
See the word probabilities
- Claude: no model returns them.
- Gemini: no 3.x model returns them (Google staff on the developer forum, August 5, 2026). Gemini 2.5 still does, for past users only.
- OpenAI: GPT-6 Astra and GPT-6.1 Sol never do. GPT-6 Sol, GPT-6 Luna, GPT-5.6, and GPT-5.5 do only with reasoning set to “none,” according to OpenAI’s documentation. gpt-4o and gpt-4.1 do.
- Ollama (since version 0.12.11) and LM Studio (since version 0.3.39) both return them.
Set the temperature
- Claude: models released after Opus 4.6 reject any value other than 1.0. Haiku 4.5 still accepts it.
- Gemini 3.x: deprecated July 21, 2026, and ignored on current Flash models.
- OpenAI: GPT-6 Astra, GPT-6.1 Sol, and the gpt-5 family reject it. The smaller models accept it only with reasoning set to “none.”
Run one prompt 20 or 100 times
claude -p,codex exec, andgemini -peach start a new session, so a loop of 100 runs works.- The desktop apps have no such mode. Claude’s documentation says the desktop app “is interactive only.”
Tally the answers
- ChatGPT, Claude, and Copilot run Python on text you paste in. Whether the Gemini app can is unclear.
Attach a paper
- All three APIs take PDFs. Local setups were not checked.
Work across a whole project folder
- Chat windows take uploaded copies in a Project (ChatGPT, Claude), a Gem, or Gemini Notebook (formerly NotebookLM).
- The Claude and ChatGPT desktop apps can connect a folder on your computer. No web version can.
- Claude Code on the web and Codex cloud work on a copy of a GitHub repository in a cloud machine, not on your folder.
See and control memory and instructions
- ChatGPT, Claude, and Copilot turn memory on by default for personal accounts, and you can switch it off. Claude Team and Enterprise start with it off.
- Gemini memory and temporary chats do not work on school accounts, including msu.edu.
- Claude Code reads CLAUDE.md and keeps an “auto memory,” on by default, in plain text. Codex reads AGENTS.md, and its memory store is off by default. Gemini CLI reads GEMINI.md.
Search the web
- Codex searches a cached index unless you ask for live results.
Check against a record like Crossref
- ChatGPT’s Python has no internet access. Claude’s reaches only software package sites on personal plans, and administrators can open more on Team and Enterprise. So from a chat window the only check is a web search, one record at a time.
Re-run cold, in a fresh context
- ChatGPT’s Temporary Chat now asks whether to personalize it. Unpersonalized is the default.
claude -ploads CLAUDE.md, memory, and plugins unless you add--bare, which needs an API key.
Vary the model
- Claude offers only Claude models, and Gemini only Gemini models.
- The paid Microsoft 365 Copilot offers Claude. The free Copilot Chat lists only OpenAI models.
- Claude Code (unofficially) and Codex can be pointed at open models through Ollama.
Red team with many agents
- Claude Code (
.claude/agents/), Codex (.codex/agents/), and Gemini CLI (.gemini/agents/) define subagents as plain-text files. - Claude Code’s agent teams are experimental and run only in the terminal.
- In code: the Claude Agent SDK, the OpenAI Agents SDK, and Google’s Agent Development Kit.
Pin a dated model version
- Anthropic: “Every Claude model ID is a pinned snapshot.”
- OpenAI’s GPT-6 models have only undated names. Gemini’s frontier model, 3.1 Pro, is a preview and can be retired on two weeks’ notice.
Licensed by MSU
- Source: MSU IT’s list of AI tools, tech.msu.edu/technology/ai/tools.
- Copilot, Gemini, and Gemini Notebook are free with an MSU login. ChatGPT Edu is bought through the MSU Tech Store with department billing.
- Codex is listed as “Codex Local,” included with ChatGPT Edu. Its cloud and GitHub features are turned off at MSU.
- ChatGPT Edu does not include API access. OpenAI’s API takes a separate request and department billing. Google’s Vertex AI is also listed as a developer service.
- Gemini CLI no longer works with free personal Google accounts, and at MSU it needs a billed Google Cloud project.
- Claude is not on MSU’s list.
Data MSU allows in it
- MSU allows public, internal, confidential research, and FERPA data in its licensed tools, including Codex Local and OpenAI’s API through MSU. Confidential research comes “with permissions.”
- MSU’s standards page adds that confidential and human subjects data must be anonymized and approved by MSU IT, with written consent, before going into an AI tool.
- A personal or free account is public data only, even for a tool MSU licenses. So is any tool without an MSU agreement. MSU has not named Claude, but by that rule it is public data only.
- Google’s Vertex AI is listed with its data rules “TBD.”
- No tool, licensed or not, may take HIPAA, payment card, CUI, or export-controlled data.
- Local models are not on MSU’s list. MSU’s guidance on DeepSeek says running an open model locally “enables the data to be processed internally,” but MSU has not ruled on local models in general.
Not yet verified
- Whether GPT-6 Sol and Luna return probabilities in a live call. OpenAI’s documentation says they do with reasoning set to “none.”
- How local setups handle a PDF attachment.
- Whether the desktop apps can pin a dated model version.
- Whether the Gemini app can run Python on pasted text.
- Whether Codex mode in the ChatGPT desktop app counts as MSU’s licensed “Codex Local.”
- Whether MSU treats Claude as public data only. That is MSU’s rule for tools without an agreement, but MSU has not named Claude.
Last verified October 2026. These products change every few months.