One AI prepares a spreadsheet, another edits a photograph, and a third writes a useful proposal without producing a file. A weaker model is not always the cause. We often compare different applications, plans and enabled tools as though they were equivalent.
This guide separates three decisions: which service to open, which working mode to use and which model to select. Features and prices were checked on 2 October 2026. Access also depends on region, device and account permissions. The recommendations are task-based selection guidance, not my measured quality leaderboard.
An application, a model and a tool are different things
ChatGPT is a service. GPT-6.1 Sol is a specific model. Image generation, a browser and spreadsheet handling are tools a service can make available to that model. The same model inside another application may have different capabilities and produce a different result.
Ask whether you receive an answer or an editable file. Can the system read your source material, retrieve evidence and verify the result? Does a connector require another paid account? Parameter counts and the most expensive subscription do not answer those questions.
ChatGPT: chat, Deep Research, Work and Codex
Chat is a useful starting point for questions, explanations, rewriting and developing an idea through conversation. Chat can also use tools; it is not limited to returning sentences. Deep Research suits questions requiring retrieval, comparison and citation across sources. You still need to read the research critically, especially when sources disagree.
Work targets substantial deliverables such as documents, spreadsheets and presentations. Codex provides a setting for code, repositories, commands and tests. Access to local files and applications is not automatic. OpenAI explains Work and Codex.
A practical Work request: "Compare these three supplier proposals in a spreadsheet. Separate stated prices from missing information. Do not place any order. Deliver an editable file and questions for each supplier." In Chat, you might first discuss which criteria matter.
Use the relevant image tool for generating or editing pictures. The name of the text model in the interface does not establish which image model produced an attachment. See the ChatGPT capabilities overview.
Gemini: research, Canvas and media creation
Include Gemini in your shortlist when you move between information work and visual creation. Deep Research handles research tasks; Canvas develops a document or prototype. Image generation/editing and video generation are separate capabilities. Check the scope and limits for your plan rather than relying on another person's demonstration.
A useful workflow is research first, a working document in Canvas second, and a visual brief afterward. One enormous prompt does not have to solve factual accuracy, final copy and art direction simultaneously. Official starting points: Deep Research, Canvas, images and video.
Claude: documents, artifacts and task execution
Consider Claude for writing, source analysis, coding and work deliverables. Artifacts support working with created content alongside the conversation. Claude Code addresses software-development tasks. The working interface is in transition: Anthropic describes a gradual merger of chat and Cowork, so accounts may not show identical controls. Current description.
For graphics, distinguish image analysis and code-generated charts, HTML or SVG from generating photographs with an image model. Anthropic explicitly explains this boundary. Connecting an external generator adds that provider's availability, account and cost. Orchestrating a tool does not turn Claude into the tool's image or video model.
Choosing between Claude and OpenAI models
The following table combines the current model families with practical selection guidance. Prices are standard API USD per million input / output tokens, not subscription prices or a fixed task cost. Tools, tax, caching, special processing modes and very-long-input premiums are excluded.
| Model | When to shortlist it | Input / output |
|---|---|---|
| GPT-6 Luna | Bounded high-volume classification and extraction with checks | $0.10 / $0.50 |
| GPT-6.1 Sol | Everyday analysis, code and deliverables with attention to cost | $2 / $10 |
| GPT-6 Astra | Difficult reasoning, complex development and interdependent tasks | $10 / $50 |
| Claude Haiku 4.5 | Fast routine processing with clear rules | $1 / $5 |
| Claude Sonnet 5.5 | Everyday balance of speed, writing, analysis and coding | $2 / $10 |
| Claude Opus 5.5 | Demanding development and longer work tasks | $4 / $20 |
| Claude Fable 5.1 | The most demanding reasoning and long agentic tasks | $10 / $50 |
Rate and positioning sources: Luna, Sol, Astra, Claude overview. Chat names and availability may differ from the API. Include required corrections when comparing task costs.
Renaming twenty files does not automatically need the most expensive model. Conversely, choosing the architecture of an application with authentication and payments requires attention to failure risk. For difficult work, establish which model handles it reliably before reducing unnecessary context and repetition.
What reasoning effort changes
Effort controls how much room a model has for reasoning. It is neither final-answer length nor a fixed token count. A short answer may follow substantial reasoning. Increasing effort also cannot repair missing requirements, an inaccessible file or an unreliable source.
| Situation | Sensible starting approach |
|---|---|
| Formatting already verified copy | Lower effort, narrow change and output check |
| Comparing several proposals | Medium effort and identical comparison criteria |
| Conflicting evidence or a difficult bug | Higher effort with verification and explicit uncertainty |
| Long agentic work | Higher effort where appropriate, a budget and a completion condition |
These are workflow suggestions, not measured equivalences between providers. Effort support is model-specific: Astra has no none setting, Luna supports it, and Haiku 4.5 does not expose the same effort control as current higher Claude tiers. OpenAI reasoning, Claude effort.
How many tokens do low, high and max consume?
There is no universal table. Consumption depends on the task, documents, history, internal reasoning, answer and tool calls. Repeated steps may resend context; caching affects billing. A token is not reliably one word, particularly across Czech, English and code.
This is an illustrative calculation, not an effort measurement. With 10,000 uncached input tokens and 2,000 total billed output tokens, a standard Sol API request costs $0.04; Astra costs $0.20 at identical token counts. If Sol uses 8,000 output tokens, it costs $0.10. These numbers do not predict the consumption of high effort. OpenAI reasoning tokens are billed as output; do not add them to total output a second time.
To measure your own work, record model/version, effort, identical input, caching, billed tokens, tools, time and quality. Repeat each variant several times. Without the same task and correctness criteria, a cheaper incomplete answer can look more efficient. Do not mechanically translate API arithmetic into a subscription allowance; see credits and resets.
Kimi, Qwen and DeepSeek: what open models mean
A web service, an API and a locally run model are three different options. A chat provides the vendor's prepared environment. An API connects to your own application. Running available weights also makes you responsible for execution, memory and tools. Check the exact version's licence: Kimi K2 uses a modified MIT licence, while current Kimi K3 lists a Kimi K3 License. Kimi K2, Kimi K3.
Apply the same process to Qwen and DeepSeek: choose a specific checkpoint and read its model card and licence. "Open" does not promise that a large model fits an ordinary laptop or that all supporting services are free. Size, quantisation and context length affect hardware requirements. Local execution can provide more control over data flows, but connected cloud tools may still send data away.
Choose around the deliverable you need
Start with three repeated tasks, such as revising a proposal, comparing a spreadsheet and preparing a visual. Define completion for each: correct figures, preserved formatting, an editable file or an accurate product appearance. Then choose the service and model.
Continue with image generators, prompts and styles and AI video. To turn tool selection into team practice, practical AI training should centre on your actual tasks and their verification.
Frequently asked questions
Does maximum effort always give the best result?
No. Better source material, clearer requirements or a suitable tool may help more. Compare task completion and correction costs.
Is a paid chat subscription the same as API credit?
No. A subscription, extra service credits and developer API billing may be separate products. Check the specific account and pricing rules.
Which model should I choose if I am unsure?
Start with a bounded task and the provider's current recommendation. Do not sacrifice correctness on difficult work for a lower token rate; validate representative cases first.