On July 23, 2026, Anthropic and OpenAI both shipped significant voice upgrades within hours of each other. Anthropic finally let Claude's voice mode use its good models — until now, talking to Claude meant talking to Haiku, its smallest and fastest one, no matter what you paid. OpenAI brought its full-duplex GPT-Live voice to the ChatGPT desktop app on both macOS and Windows, so you can now talk to ChatGPT while it runs agent tasks on your computer. Two very different bets landed on the same day, and the difference between them tells you which one you actually want.
The short version: Anthropic fixed a quality problem, OpenAI extended a capability. Claude's voice got smarter. ChatGPT's voice got somewhere new to live. Which matters more depends entirely on whether you want an assistant to think with or an assistant to work with.
Quick summary: Claude voice mode can now use Haiku, Sonnet or Opus, and you can switch between them mid-conversation — previously it was Haiku only. Connectors including Gmail, Google Calendar, Slack, Canva and Notion now work by voice, in ten languages, in beta across mobile, desktop and web. Free users get voice but are capped at Haiku and a single connector. Separately, OpenAI's GPT-Live voice arrived in the ChatGPT desktop app for macOS and Windows, letting you talk while it drives Codex and agent tasks, with a Mac-only Appshots feature that reads your frontmost window. That desktop voice is rolling out to Plus, Pro, Business, Edu and Enterprise — not the free tier.
Side by Side
| Claude voice | ChatGPT desktop voice | |
|---|---|---|
| What changed July 23 | Can now use Haiku, Sonnet or Opus | Full-duplex voice came to the desktop app |
| Model choice | Switch models mid-conversation | Not user-selectable |
| Platforms | Mobile, desktop and web (beta) | macOS and Windows |
| Free tier | Yes — Haiku plus one connector | No — Plus, Pro, Business, Edu, Enterprise |
| Connectors by voice | Gmail, Calendar, Slack, Canva, Notion | Drives Codex and agent tasks |
| Screen awareness | Not offered | Appshots reads frontmost window (Mac only) |
| Languages | Ten | Not specified in launch coverage |
What Anthropic Actually Fixed
Claude's voice mode has existed since 2025, and it had a specific, well-known problem: it ran on Haiku, Anthropic's smallest model, regardless of your subscription. So people who paid for Claude specifically because Opus writes and reasons well would open voice mode and get noticeably weaker answers, with no way to change it. That gap is now closed. You can pick Haiku, Sonnet or Opus, and switch between them without leaving the conversation — useful in practice, because you might start a quick question on Haiku and escalate to Opus when it turns out to be hard. Sources: TechCrunch; MacRumors.
The connector support is the more interesting half. Gmail, Google Calendar, Slack, Canva and Notion now work by voice, which turns Claude's voice mode from a conversation into something closer to an assistant that can look things up in your actual accounts while you talk. Free users get voice, capped at Haiku and one connector — a genuinely usable free tier rather than a demo.
One honest caveat that most launch coverage skipped: Anthropic did not change the underlying voice model itself. The models answering are better, but the speech layer — including interruption handling, the thing people complain about most in every voice assistant — was not improved. If talking over Claude felt awkward before, it will still feel awkward.
What OpenAI Actually Extended
OpenAI's move is about location rather than intelligence. GPT-Live, its full-duplex voice mode, now runs in the ChatGPT desktop app on macOS and Windows. Full-duplex means both sides can speak at once, so it handles interruption far more naturally than turn-based voice — which is precisely the area Anthropic did not touch. The genuinely new capability is that you can talk to it while it operates: it can be running Codex or agent work on your machine and you can narrate, redirect and ask questions while that happens. Sources: 9to5Mac; TechCrunch; Android Authority.
The Mac-only Appshots feature lets it reference whatever window you have in front, so you can ask about what you are looking at without describing it or pasting anything. That is a small feature with a large effect on how conversational the thing feels. The cost is that this desktop voice is a paid feature, rolling out to Plus, Pro, Business, Edu and Enterprise tiers, and not to free accounts.
Which One Should You Use?
- You want to talk to a genuinely strong model without paying: Claude. Its free tier includes voice on Haiku with one connector, and paid tiers now unlock Sonnet and Opus by voice.
- You want voice that handles interruption naturally: ChatGPT. Full-duplex is a real architectural advantage, and Anthropic explicitly did not improve its speech layer this round.
- You want an assistant that reaches into Gmail, Calendar, Slack or Notion while you talk: Claude, which now supports those connectors by voice.
- You want to narrate work while an agent executes it on your computer: ChatGPT desktop, which is the only one of the two built for that.
- You are on Windows: both work, but note Appshots screen awareness is macOS only.
- You mostly use voice on a phone: this particular ChatGPT update is about the desktop app, so it may change nothing for you.
The Pattern Underneath Both Launches
Look at what each company chose to fix and you can read their strategy. Anthropic's users were complaining that voice gave them a weaker model than they paid for, so Anthropic made voice reach the good models and the user's own accounts. OpenAI's users were increasingly running agent tasks, so OpenAI made voice the way you supervise those tasks. One company is treating voice as a better interface to thinking; the other is treating it as a control surface for doing.
Neither is wrong, and most people will end up wanting both at different moments — which is the recurring problem with committing to a single assistant. If your use is mostly the thinking half, being able to put the same question to several current models and compare how they handle it is more valuable than any one vendor's voice roadmap. LumiChats keeps many current models under one login at a pay-per-day price for exactly that, though it is a text tool and not a replacement for either voice mode.
01Can free users talk to Claude?
Yes. After the July 23, 2026 update, free Claude accounts get voice mode limited to the Haiku model and a single connector. Paid tiers can use Sonnet and Opus by voice and connect more services.
02Is ChatGPT voice on desktop free?
No. The GPT-Live desktop voice mode is rolling out to Plus, Pro, Business, Edu and Enterprise plans. Free accounts are not included in this rollout.
03What was wrong with Claude voice before?
It ran only on Haiku, Anthropic's smallest and fastest model, no matter which plan you paid for. Users who subscribed for Opus-level quality got noticeably weaker answers in voice mode with no way to change it. The July 2026 update fixed that.
04Which voice mode handles interruptions better?
ChatGPT. Its GPT-Live mode is full-duplex, meaning both sides can talk at once. Anthropic upgraded which models answer in Claude voice but did not change the underlying speech layer, so interruption handling is unchanged.
05Can Claude access my Gmail or Calendar by voice?
Yes, with connectors. Gmail, Google Calendar, Slack, Canva and Notion now work in voice mode across ten languages, in beta on mobile, desktop and web. Free accounts are limited to one connector.
06What is Appshots in ChatGPT?
A macOS-only feature that lets ChatGPT reference the window you currently have in front, so you can ask about what you are looking at without describing or pasting it. It is not available on Windows.
The bottom line: if you want the smartest model you can talk to, and especially if you are not paying, Claude just became the better voice assistant. If you want voice that feels natural to interrupt and that can supervise real work on your computer, ChatGPT desktop is the more capable tool and you will need a paid plan. The genuinely good news is that two rivals shipping on the same day means neither can afford to leave voice mediocre for long.
