ChatGPT Voice Rolls Out to Desktop Apps

On July 24, OpenAI announced that ChatGPT's new voice mode now works in its desktop apps. Voice conversations, previously a mostly mobile experience, can now be used in earnest on the desktop as well.

The headline feature is that ChatGPT Voice now **integrates with both ChatGPT Work and Codex**, so you can execute tasks and control agents with voice commands. Expect use cases like having it write code while you talk through the problem, or having it assemble whole workflows. As voice AI steadily shifts from "chatbot" to "work partner," bringing a voice interface to the desktop — the epicenter of productivity — looks like a meaningful step.

Claude Strengthens Its Voice Mode Too — Opus and Sonnet Now Selectable

Anthropic is pushing back. On July 23, **Claude's voice mode received an update**, making **Opus and Sonnet selectable** in addition to the existing Haiku. At the same time, external services such as Gmail, Google Calendar, Slack, Canva, and Notion can now be operated by voice.

On Zenn, a Japanese platform, an article has already appeared testing "the possibility of handing chores to Claude without stopping your work," and it will be interesting to see how this catches on in real-world use.

Sakana AI's Fugu Ultra v1.1 Adds a Claude Code-Compatible Endpoint

Japan's Sakana AI has updated **Fugu Ultra**, its multi-model routing system, to v1.1. The company claims gains of up to 7.9 points over v1.0, and a **Claude Code-compatible endpoint** has been added.

There is no independent verification yet, though, and the claim that it "outperforms Fable 5 even without including Fable 5 in the pool" is, at this point, based on the company's own announcements — best taken with a grain of salt. Its unavailability in the EU also remains a continuing constraint that users, including those in Japan, should keep in mind.

20 Companies Sign Joint Letter in Defense of Open Weights

More than **20 companies**, including NVIDIA, Meta, Microsoft, Palantir, and Hugging Face, have jointly signed a letter urging policymakers to avoid "premature regulation" of open weight models.

Alongside this, NVIDIA and Microsoft each released a document titled "Open Weights and American AI Leadership," making clear their position that open weights are core to American AI competitiveness. With regulatory debates heating up, this move amounts to the industry drawing a "defense line."

The Split Within Silicon Valley Over How to View "Chinese AI"

Alongside the open weights question in the previous section, another story worth noting is **the split inside Silicon Valley over how Chinese AI is perceived**, as reported by WIRED. AI startups raising enormous amounts of money emphasize Chinese AI as a threat, while smaller players reportedly see it in an entirely different light.

The industry is not a monolith, and where regulation ends up is still anyone's guess. The open weights letter, too, is best read with this split in mind.

Nous Research's Hermes Agent Gets About 80% Faster

On the agent front, Nous Research released **Hermes Agent v0.19.0** (Quicksilver Release) on July 20, announcing that cold start initialization time has been **cut by roughly 80%**, from about 4.3 seconds down to 0.9 seconds.

Task delivery, security approvals, and execution environment isolation have also been improved. Reworking the "first second" — often a bottleneck when running agents in production — is an unglamorous but promising update.

---