GPT-6.1 Terra Dictation: Audio Limits Explained in 2026
What a text-and-image model can do with your spoken draft, and which step has to happen first.
October 27, 2026
GPT-6.1 Terra can't accept audio through the API. OpenAI lists text and images as inputs, and text as the output. To dictate something for Terra to edit, first transcribe your speech, check the transcript, then send the text. An app's microphone may make this extra step less obvious, but it doesn't change what the model can do.[1]
That distinction deserves a place alongside the launch headlines. OpenAI announced Terra on October 26, and GIGAZINE covered it on October 27. Its report covers lower-cost computer work and stronger visual reasoning. Interesting Engineering also points to its use for office tasks.[2][3] Those are good reasons to explore the model. They don't explain what happens to a spoken note before Terra sees it.
I checked the model documentation because phrases like "writes naturally" and "understands what's on screen" can make it sound like Terra offers a complete voice workflow. The published API contract covers less than that. This guide explains what it includes and how to use spoken drafts with a text model. It isn't a hands-on test of Terra's accuracy.
A microphone button doesn't identify the model hearing you
Suppose you're looking at an order screenshot and need to email the supplier. In this fictional example, you say, "The replacement hinges arrived. Two are scratched. Ask for replacements, but don't approve another shipment charge."
A voice app could transcribe what you say, attach the screenshot, and ask Terra to draft a reply. It may feel like you're speaking directly to Terra, but another component turns your speech into text. If it misses "don't," the writing model starts with the wrong instruction.
OpenAI's audio guide treats voice input for an existing text agent separately from file transcription and conversational audio. It outlines a chained workflow using speech-to-text, the text agent, and text-to-speech when a spoken reply is needed.[4] For an email draft, you may not need text-to-speech.
Terra's model page also says it doesn't support audio transcription or Realtime endpoints.[1] ChatGPT or Codex could still have a microphone feature. App features and model inputs are different things. Check the interface you're using instead of assuming a screenshot from one app describes another product's API.
Check the transcript before asking for a better email
For the hinge example, save the spoken note as editable text before requesting a polished response. Check the sentence about the charge, then compare the quantity with the screenshot yourself. Use the screenshot for context, but don't let it replace what you meant.
Then give the writing model a clear task with firm limits. For example: "Write a short email to the supplier using this note. Ask for replacements, but don't approve shipping charges. Leave the order reference blank if it's missing."
That's a suggested instruction. Terra could still produce a draft that needs corrections. Review the draft before sending it. Make sure the email asks for the action you want and stays within the limits you set.
For the first trial, keep the transcript with the revised message. If the transcript gets the words wrong, check speech capture or recognition. If the transcript is accurate but the reply promises reimbursement without a basis, check the writing step. Without the transcript, both issues look like bad dictation, and you might spend an afternoon fixing the wrong tool.
For longer notes, flag any details you're unsure about. Leave "Check the order number" unresolved until someone checks it. Guessing a number to polish the paragraph just creates more work for the person who has to catch it later.
The launch prices cover text-model use
Terra lists standard rates of $1 per million input tokens and $5 per million output tokens. Cached input has its own rate. The page also lists extra charges for long contexts, Fast mode, and some tools.[1] Keep these costs in mind when estimating an integration.
Independent coverage suggests the appeal: a capable model at a lower price may make office automation worth trying.[2][3] But token prices don't cover the full cost of a voice experience. Account for transcription and any file storage or tool services your app uses. API rates also tell you nothing about an app subscription's price or included usage.
I wouldn't set a price per note until the app shows how much each request uses. A short voice note can come with a much longer prompt, including instructions, chat history, and attachments. Each revision uses more resources. For a small trial, you don't need a huge spreadsheet, but you should track which costs come from each stage.
Choose the amount of automation the task needs
OpenAI points developers building new conversational voice apps to GPT-Live. Its guide has separate paths for transcription and other audio tasks.[4] A voice conversation works differently from dictating an editable paragraph into an email field.
For an email, start with speech-to-text and a visible draft. Add a writing model only for a specific task, like turning rough notes into a message to a supplier. Giving an agent browser control just to insert a paragraph adds permissions and steps the task may not need.
DictaFlow is a separate tool for entering text. Its documented workflow is hold, speak, release, and the app inserts text at the cursor. It supports native desktop and mobile apps.[5] You can dictate a brief into an AI app's text field, check it, then submit it. This article doesn't announce a Terra integration or say DictaFlow uses Terra internally.
If your current dictation tool gets your words right, keep using it during the trial. Changing both dictation and the writing model makes it harder to tell what affected the results. If you need to send spoken text to the active app without building an audio API pipeline, see DictaFlow's setup guide.
Try one ordinary message before rebuilding your workflow
Try this with the fictional hinge note or another harmless task. Start by checking the raw transcript. Then ask for a draft and set a clear restriction. Compare the draft with your source note. Finally, try the same task using your usual input method, and write down any corrections you make.
If a reply misses a charge restriction, beautiful writing doesn't make that okay. A correct reply can still be a poor fit if it takes longer to review than to type. Neither result tells you how Terra performs across everyone's work. It only tells you whether this workflow is worth testing again.
Before paying for a voice integration, ask the builder how it turns audio into text and whether you can review the text before the writing model responds. This matters for Terra because it has published input limits. Save the answer with your trial results and the app version you tested.
Sources
[1] OpenAI: GPT-6.1 Terra model inputs, endpoints and pricing
[2] GIGAZINE: October 27 Terra launch report
[3] Interesting Engineering: Terra office-work coverage