explainx / blog
Jul 23, 2026: Claude Voice leaves Haiku-only — Opus/Sonnet on paid, connectors mid-call, 11 languages every plan. Turn-based vs GPT-Live; Free vs paid.
Jun 30, 2026
Claude Sonnet 5 launched June 30, 2026. Default on Free and Pro. API: claude-sonnet-5. Intro pricing $2/MTok input, $10/MTok output through Aug 31.
Jul 24, 2026
AMD and Anthropic locked a compute + equity partnership: up to 2 gigawatts of Instinct MI450-series Helios racks, a future AMD stake of up to $5 billion, and Claude-driven ROCm engineering — plus the MI355 “weekend bring-up” anecdote.
Jul 24, 2026
The Claude Cookbook isn’t a food blog — it’s Anthropic’s practical notebook index from RAG to async multi-agents. explainx.ai maps the 2025–2026 must-reads against the HN “prompt theatre” debate and CLAUDE.md minimalism.
On July 23, 2026, Anthropic shipped the voice upgrade Claude users had been waiting for: Voice mode no longer stuck on Haiku-for-speed. Paid chats can run Opus and Sonnet, reach connected tools mid-conversation, and speak many more languages on every plan — including Free.
The company framed it as “Think through hard problems in voice mode”. Same week OpenAI put ChatGPT Voice on desktop with GPT-Live. Two labs, two philosophies: Anthropic bet on stronger chat models + connectors; OpenAI bet on full-duplex Live + agent/computer control.
| Question | Answer |
|---|---|
| When? | Jul 23, 2026 — public beta |
| Where? | Mobile, desktop, web (best on phone) |
| Models? | Opus · Sonnet · Haiku (paid); Free = Haiku |
| Default model? | Last model used in text chat (fastest variant) |
| Tools? | Connected apps mid-call; permission before use |
| Free tools? | One connector |
| Languages? | 11 variants (see list) — manual switch |
| Architecture? | Turn-based (listen → think → speak) |
| Quota? | Counts toward regular usage limits |
| vs ChatGPT? | Depth + connectors vs GPT-Live duplex agents |
Anthropic added hands-free conversation earlier in 2026, then watched people use it for long sessions — pitches, offer decisions, process reviews — not just “what’s the weather.” Voice still routed through Haiku for latency. Fast replies, shallow reasoning. That mismatch is what this release fixes.
Official positioning:
Claude Opus and Sonnet, models designed for hard problem-solving, are now available in voice mode… Voice mode uses the fastest version of whichever model you’ve selected… It defaults to the last model you used in text chat, so you can move between voice and text without starting over.
You can switch models mid-conversation from the picker. That matters when a Haiku warm-up turns into an Opus strategy session without killing the thread.
| Capability | Free | Paid |
|---|---|---|
| Models in voice | Haiku only | Haiku + Sonnet + Opus |
| Connected tools | One | All connected tools |
| Languages | All listed languages | All listed languages |
| Surfaces | Mobile / desktop / web beta | Same |
| Usage | Regular plan limits | Regular plan limits |
If you are on Free and the demo video shows calendar moves and inbox drafts, that is a paid path. Languages are the egalitarian part of the launch.
For Max/Team users already juggling Fable credits and Opus 4.8 as the public Opus flagship (still no Opus 5), voice is another way to burn the same buckets — plan spoken Opus time like you plan long text Opus runs.
Anthropic’s “From talking to doing” section is the product punchline. Examples they publish:
Claude asks permission before using a connected tool. Connect under Settings → Connectors on mobile, desktop, or web. Press coverage also names Slack, Notion, and similar apps in the connector family — treat the official Gmail/Slack examples as the guaranteed story and your tenant’s connector list as source of truth.
This is closer to voice-driven Claude apps automation than to OpenAI’s desktop computer-control narrative. It also inherits connector risk: a spoken “send that email” with a loose approval habit is how agentic SaaS goes wrong — same category as AgentForger on the ChatGPT side and Cowork local escape on the laptop side. Keep approvals on for mail/calendar writes.
Official voice languages:
Critical UX notes from Anthropic:
For India and LATAM teams, Hindi + both Spanish locales are the practical headline. For product managers shipping multilingual support, this is the first time Claude Voice looks like a global sounding board rather than an English beta.
Anthropic is explicit: voice takes turns — listen, pause to think, respond. That is a deliberate trade for model quality and tool calls.
| Dimension | Claude Voice (Jul 2026) | ChatGPT Voice / GPT-Live |
|---|---|---|
| Duplex | Turn-based | Full-duplex (speak/listen together) |
| Model story | Opus / Sonnet / Haiku chat models | GPT-Live voice-native stack |
| Tools | App connectors (mail, calendar, Slack…) | Work/Codex agents, computer control, Appshots |
| Best surface | Phone (per Anthropic) | Desktop agent workflows (Jul 23 desktop ship) |
| Free | Haiku + 1 tool | Desktop Voice not named for Free/Go |
Neither “wins” universally. Walking a pitch → Claude Opus voice. Driving multiple coding agents away from the keyboard → ChatGPT desktop Voice. Many power users will keep both.
Deep background on OpenAI’s stack: GPT-Live launch.
From the official post — useful as a prompt library:
With connectors enabled, add:
Keep wait / don’t send / confirm in the spoken prompt. Voice makes over-eager tool use more likely than text.
| Moment | Suggested model |
|---|---|
| Warm-up / small talk / language practice | Haiku |
| Structured brainstorm, Sonnet-tier coding talk | Sonnet 5 |
| Offer negotiation, strategy, careful writing feedback | Opus 4.8 (or Fable if your plan/credits allow and the picker exposes it — verify in-app) |
| Tool-heavy mail triage | Paid plan + Sonnet/Opus + tight approvals |
Default-to-last-text-model is convenient and dangerous: finish a Fable-heavy coding chat, open voice, and you may inherit an expensive default. Check the picker before a 40-minute walk.
Advisor/executor habits from Fable + Sonnet Claude Code still apply mentally: use the cheap model to shape the problem, escalate when the argument gets hard.
Claude Voice’s July 2026 update is Anthropic admitting Haiku-only was a product dead end for “think out loud” work. Opus and Sonnet in voice, connectors with permission, and 11 languages on every plan make Claude a serious spoken sounding board — especially on mobile — while ChatGPT races ahead on duplex desktop agents.
Try it: open the Claude mobile app, tap the sound wave, set your language, pick Sonnet or Opus if paid, and run one pitch rehearsal with calendar/mail connectors off before you turn tools on.
Sources: Anthropic — Think through hard problems in voice mode (Jul 23, 2026) · Claude product announcement thread · TechCrunch / Engadget contemporaneous coverage
Model availability, connector lists, and language packs reflect Anthropic’s July 23, 2026 voice mode post and may change during beta. Verify Free vs paid entitlements and permission prompts in your Claude app before relying on voice for mail or calendar actions.
