Your app - choosing a model
Which model to pick: Sol every day, Astra for big things
In short ↳Click the model name next to the prompt box and a list opens. Set GPT-6.1 Sol for everyday work. Switch to Astra when you plan something big, build a whole app or a 3D project.
Where to find it: Codex → prompt box → model name
Select the screenshot to enlarge it.
What you see on screen
- 1
Model name and effort
Clicking opens the list. Beside the name is the effort level (Ultra on this screenshot, Light and High elsewhere).
- 2
Default
'Recommended set of models': the app picks for you. Fine to start with, but choose manually when you want control over cost and quality.
- 3
GPT-6.1 Sol: your main model
Set it as your default. Near-Astra quality on coding and professional work (per OpenAI's claim), and the API price is 2 / 10 USD (input / output per million tokens) - about 1/5 of Astra.
- 4
GPT-6 Astra: only for big tasks
The most capable and most expensive (API: 10 / 50 USD). Use it for planning, building a whole app, hard reasoning and 3D projects. It is ticked on the screenshot.
- 5
GPT-6 Sol
A model for everyday work and agents (API: 2 / 10 USD). The older sibling of 6.1 Sol: if 6.1 Sol is on your list there is no reason to pick this one.
- 6
GPT-6 Luna
The fastest and cheapest (API: 0.1 / 0.5 USD). For simple, high-volume tasks: sorting, short summaries, extracting data. Too light for hard work.
- 7
GPT-5.6 Sol, Terra, Luna
The older family from July 2026. Since 7-8 Oct 2026 it is no longer the Chat default, but per the documentation it stays on the list during the rollout. Pick GPT-6 for new tasks.
How to use it - in plain words
A simple rule: Every day GPT-6.1 Sol. Astra only when the task is truly big.
- Planning an architecture, building a whole app or a 3D project: Astra.
- Everyday work: code, documents, analysis, an agent for tasks: 6.1 Sol.
- Lots of simple tasks at once: Luna.
A tip from the author's practice, not an official OpenAI recommendation: even Astra at low effort ('Light') is often enough, and it is faster and uses less of your limit. Raise effort only when the result is shallow.
Prices refer to the API and are recorded in the course (lesson 13); on a subscription your plan limit applies. The model list depends on your account and plan.
Try it on your account
Set GPT-6.1 Sol as your default. For your next big task switch to Astra, and return to Sol when it is done.
Lesson 13 · Practical guide
How to use GPT-6: the right model for the job
Start with Sol and review an answer on a small sample. Compare Luna for simple tasks and Astra when Sol misses your criteria. These profiles cover Astra, Sol and Luna; GPT-6.1 Sol has a separate section.
The prices below are API rates per 1 million input and output tokens. ChatGPT subscriptions are billed separately. Task cost also depends on actual usage and tools.
GPT-6 Luna
gpt-6-luna
0.1 / 0.5 USD
API · input / output · per 1M tokens
The fast specialist for simple tasks
Example tasks
- Summaries of emails, notes and documents
- Extraction: tax IDs, amounts, dates, names from invoices and contracts
- Classifying and tagging tickets (e.g. 500 customer reviews)
- Quick questions, translations, style fixes
When to consider another model
- You are making a decision with many dependencies
- You are running a long task across many tools
Reasoning level: none / low for a first trial. Check quality; if it misses your criteria, compare Sol on the same task.
50.9% on Agents' Last Exam at about 20× lower price than Sol. Also on Free and Go (desktop app).
GPT-6 Sol
gpt-6-sol
2 / 10 USD
API · input / output · per 1M tokens
The default for everyday work
Example tasks
- Offers, reports and analyses from several files
- ChatGPT Work tasks: mail + calendar + CRM + spreadsheet
- Code in Codex: fixes, tests, reviews, new features
- Automations and skills that run every day
When to consider another model
- You repeat a simple job in bulk: check whether Luna meets the same criteria at a lower API rate
Reasoning level: medium for a first trial. Compare high / xhigh on a hard task; higher effort does not guarantee a correct result. Assess quality, time and usage.
56.4% on Agents' Last Exam (max) and about half the factual mistakes of GPT-5.6 Sol. Do not convert a benchmark into a general quality percentage; compare results on your own task.
GPT-6 Astra
gpt-6-astra
10 / 50 USD
API · input / output · per 1M tokens
The expert for the hardest problems
Example tasks
- Complex research and analysis across many sources
- Computer use and long agent tasks
- System architecture, hard bugs
- Decisions with big money or risk behind them
When to consider another model
- You write a short email, summary or ask a simple question
- Sol already meets your criteria; Astra’s recorded API token rates are 5× higher
Reasoning level: medium or high. Keep max and Extra High for tasks where a mistake costs more than time and tokens.
Family leader: 59.3% on Agents' Last Exam, 73.5% on computer use, 41.4% on AutomationBench.
Which model? Answer 3 questions
More detailReasoning levels and prompt examplesSettings, available surfaces, cost comparison and templates for each model. Open after your first trial.
Reasoning levels - what to set
Reasoning effort sets how much work the model may spend preparing an answer. Higher effort may help with a hard task, while increasing time and usage. Compare results; correctness is not guaranteed. Check supported levels and cache rules for the chosen model.
| API | In the app | When |
|---|---|---|
none | none / Instant | Instant answer: translation, rewrite, simple question. |
low | Low | Short tasks with a clear goal: summary, extraction, email. |
medium | Medium (default) | Most work: report, offer, spreadsheet analysis, code fix. |
high | High | Many files and dependencies, output goes to a client or the board. |
xhigh | Extra High | Hard problems: strategy, complex code, multi-source research. |
max | Max | Compare it on a particularly hard task. It may increase time and usage; check whether the improvement justifies that. |
Where to find these models
Chat
Quick conversation. Since 7 Oct 2026 it has used GPT-6 Sol on paid plans, while Free and Go receive GPT-6 Luna from 8 Oct. Answers can use Intelligent UI: graphics, buttons, forms, charts and interactive tools.
ChatGPT Work
Pick Astra, Sol or Luna in the model picker, next to the reasoning level. Default choice: Sol + medium.
Codex
Same picker. Sol for most coding, Luna for simple edits, Astra for architecture and hard bugs.
API
Set the model parameter (gpt-6-sol, gpt-6-luna, gpt-6-astra) and reasoning.effort. Billed per token, separately from the subscription.
API cost example with equal usage
100 API calls with 10K input and 2K output tokens each (e.g. 100 reports from documents), standard rates, no cache.
With equal token usage, the recorded rates make Luna's cost about 20× lower than Sol's and 100× lower than Astra's. This example covers tokens, excluding tools, corrections and human work. If a model is wrong in one case out of five, corrections may cost more than the saving — so test on a sample.
Ready prompts for each model
Extracting data from invoices
From the attached invoices, output a table: invoice number, seller tax ID, issue date, net amount, VAT, gross, due date. If a field is missing, write “MISSING” - do not guess. Finish with the gross total and a list of invoices with missing fields.
Sales offer from files and CRM
Goal: a 2-page B2B offer for client [name]. Sources: meeting note (attached), 2026 price list (Drive), client history (CRM). Done means: 3 pricing options, benefits matched to the problems in the note, no prices outside the price list. Show the plan first, then the draft. Do not send anything - leave it for my approval.
Strategic decision with risk analysis
We are considering entering [country] with product [X]. Analyse the sales data (attached), the market report and 3 competitors. Deliver: a YES/NO/CONDITIONAL recommendation, 3 scenarios (pessimistic, base, optimistic) with numbers, the 5 biggest risks with mitigation, and a list of assumptions we must verify. Mark what is a sourced fact and what is your estimate.
5 rules for working with GPT-6
- Default to Sol + medium. Change only for a specific reason.
- Escalate step by step: first a higher reasoning level, then a stronger model.
- Describe the goal and acceptance criteria. Include the sequence when your procedure requires one.
- Test bulk jobs on 10 examples with Luna before running 1,000.
- A stable API prompt prefix may use caching, meaning reused context. Check the chosen model’s rate and tokens actually billed as cached input; the discount applies only to that part of the input.
Common mistakes
- Choosing Astra and Extra High for every task without comparison. You may increase usage even when a simpler setup already meets your criteria.
- Judging a model on one answer. Compare on the same 5–10 cases.
- Accepting an analysis with many dependencies because it sounds convincing. Check conclusions and exceptions regardless of the model.
- Counting only the token price. What matters is the cost of a correct result: revisions and your time cost too.
- Assuming GPT-5.6 in Chat is the same as GPT-6 Sol in Work. They are two different generations.
Benchmark results (Agents' Last Exam, AutomationBench, computer use) come from OpenAI's 22 Sep 2026 materials and show differences between models, not a guarantee for your process. Access and limits depend on your plan - check your account.
More detailAPI cost calculator and model tableCalculate input, output and call volume. The result uses the recorded API rates, not subscription billing.
Lesson 13 · Model selection
The GPT-6 family (Astra, Sol, Luna): quality, time and cost
Compare results on the same data. Measure the whole process, including revisions and your review.
This calculator illustrates rates recorded in the course. It is not a current offer or bill forecast. Before deciding, check model, cached rates, tool charges and billing terms in official pricing.
GPT-6 and GPT-6.1 models
GPT-6 Astra
gpt-6-astra · 10 / 50 USD
The most capable (launched 3–4 Sep 2026). Hard reasoning, research, large documents and code.
GPT-6.1 Sol
gpt-6.1-sol · 2 / 10 USD
New from DevDay (29 Sep 2026). Near-Astra quality on agentic coding, computer use and professional work, at 1/5 of the price. Cached input 0.10 USD.
GPT-6 Sol
gpt-6-sol · 2 / 10 USD
Everyday work and agents (launched 22 Sep 2026). About half the factual mistakes of GPT-5.6 Sol, ~5× cheaper than Astra.
GPT-6 Luna
gpt-6-luna · 0.1 / 0.5 USD
Fastest and cheapest (launched 22 Sep 2026). High volume: classification, summaries, simple extraction.
Since 7 Oct 2026, regular Chat on Plus, Pro, Business and Enterprise has used GPT-6 Sol; Free and Go receive GPT-6 Luna from 8 Oct, subject to rollout. GPT-6 in Chat introduces Intelligent UI: answers can combine text, graphics, buttons, forms, charts and interactive tools. This release did not change the models in Work or Codex, where GPT-6.1 Sol has remained available since 29 Sep 2026. All models in the table have a ~1.05M-token window and up to 128K output tokens. Rates shown: USD per million input / output tokens; cached input rates depend on the model.
When to test Astra
Astra supports demanding reasoning, coding, research and document work. Start with GPT-6.1 Sol (since 29 Sep 2026) - it delivers near-Astra quality at about 1/5 of the price. Switch to Astra when Sol misses your acceptance criteria. A higher token price does not determine the cost of a correctly completed task.
Example: give two models the same report and ask them to find inconsistent numbers. Count correct findings, false alarms and your review time.
Subscription and API
Check access and limits in your account. Do not assume pooled team limits or a fixed task count based on the subscription price.
API billing is separate. Text cost depends on input and output tokens, including reasoning tokens. Request and token rate limits also apply. Elapsed minutes are not the billing unit in this price table.
Estimate API cost
Rates recorded in the material for GPT-6 Astra: USD 10 / million input tokens, USD 50 / million output tokens. Text-only estimate, excluding cache and tool fees. Check pricing thresholds and conditions for the particular model.
One agent task can involve multiple calls. This is neither a subscription price nor a guaranteed bill.
A model selection procedure
- 01Choose a representative task and acceptance criteria.
- 02Run the same brief on two available models.
- 03Compare correctness, review time and actual usage.
- 04Choose the model meeting your criteria at an acceptable total cost.
A prompt saying “work for at most 20 minutes” is an instruction, not a hard spending cap. Configure cost controls in your application and verify what actually stops execution.
Sources and rate updates: GPT-6 Astra · GPT-6 Sol & Luna · API model list · Reasoning tokens · Plans and access
More detailGPT-6.1 Sol screen and DevDay announcementsThe newer version and announcements from 29 September 2026. Check current terms before use.
Model M6.1 - GPT-6.1 Sol
'Introducing GPT-6.1 Sol' - should you switch?
In short ↳After DevDay (29 Sep 2026) the app offers GPT-6.1 Sol: near-Astra performance on complex work at 2 / 10 USD per million tokens - a fifth of Astra's price. In ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu; API id gpt-6.1-sol.
Select the screenshot to enlarge it.
What you see on screen
- 1
The promise
'Maximize usage' - a cheaper model burns less of your plan limit, so you get more work done before you hit it.
- 2
'Near-Astra performance'
OpenAI's claim, mostly on agentic coding, computer use and professional work. Verify it on your own tasks - it is not a guarantee for every job.
- 3
Continue with current model
Keeps your model for now. A safe choice in the middle of a task you already compare against a known result.
- 4
Try GPT-6.1 Sol now
Switches the model. Run it on a task you already did with Astra or GPT-6 Sol and compare quality, time and limit use.
How to use it - in plain words
Do not click on impulse. Take one real task (a report, a bug, a spreadsheet check), run it on GPT-6.1 Sol and on your current model, and compare the results against your acceptance criteria.
Default for daily work in Work and Codex: GPT-6.1 Sol. Keep Astra for the hardest tasks where 6.1 Sol misses your criteria, and Luna for high-volume simple jobs.
For advanced users: example prompt+
Task: <paste a task you already did with another model>. Do it again and end with a short self-check: 1) which acceptance criteria you met, 2) which assumptions you made, 3) what I should verify by hand.
DevDay - San Francisco, 29 September 2026
OpenAI DevDay 2026 announcement summary
Agents API - public beta
Not chat over an API, but task execution: hosted runtime, memory, tools and multi-agent work in a single task. Plus Computer Use in the API, Tool Search and automatic Context Compaction for long conversations.
Decisions API - a model that decides
Built on GPT-6 Luna. It doesn't generate text, it picks one of the options you define and returns it with a confidence score: classification, routing, the agent's next step. It answers in about 150 ms. Limited preview, pricing has not been published yet - check the current price before you build on it.
Dots - agents that keep working
OpenAI answers Meta's Muse with Dots: personal agents on GPT-6 Astra running 24/7 on their own cloud computer, connected to more than 4,000 apps. They keep context between conversations and work with the laptop closed. Background app access is read-only - changing data or sending anything needs your approval. One primary Dot for now; Pro and Business Premium (Enterprise/Edu: beta, off by default).
Codex - security and many environments
Codex Security Cloud scans your repo for vulnerabilities, investigates the findings itself, removes duplicates and prepares the fixes - and scanning can run continuously. Plus multiple cloud dev environments, Codex Cloud working with the laptop closed, voice control in the CLI, the /agents view and code review in the desktop app.
Ultrafast - fast, at a price
Codex up to 8x faster (up to 300 tok/s), the API up to 6x faster. OpenAI quotes 6x the standard price - for GPT-6 Astra that works out to 60 USD/M input and 300 USD/M output, though some of those line items are calculated rather than published in the DevDay materials. Available in the API and in ChatGPT Work and Codex for Enterprise and the Pro 500 plan (500 USD/mo).
Pro 200 and Pro 500 - recorded changes
The DevDay material describes a Work and Codex limit change from 20x to 10x, GPT-6 Pro messages from 200 to 100 per week, and a Pro 500 plan (500 USD/mo) with Work and @ChatGPT for Slack and Teams. It also describes a transition period and a one-off credit for existing customers. Confirm current terms on your account.
ChatGPT Space and Pages
A shared space where a team and ChatGPT build pages: interactive documents with charts, checklists and images that update themselves from connected sources. Pro, Business and Enterprise, desktop and web (mobile: view only). Collaborative slides with export to PowerPoint and Google Slides are due within weeks. Team Tasks and the Slack/Teams connection are Business and Enterprise only.
Plugin extensions and @ChatGPT in Slack and Teams
Plugins get a home in the ChatGPT sidebar: panels beside the conversation, fullscreen apps and file viewers (web support for Free and Go is still to come). You can mention @ChatGPT in Slack and Teams channels, threads and direct messages; teammates do not need their own ChatGPT licence.
Meetings plugin - notes without a bot
Takes notes on calls and in-person conversations without a bot joining, creates a summary with action items and adds it to a Space. Beta in the Mac desktop app for Pro and Business. Agree recording consent and GDPR rules with participants before you use it.