Notes from my first week at Speechify
Monday I logged into Speechify’s Slack for the first time. By Friday I’d shipped three things and watched the US government change how frontier AI models reach developers. Not a bad first week.
The Speechify Cookbook went up on Monday. A public GitHub repo of runnable recipes for the Speechify TTS API, in TypeScript and Python, with every recipe in two flavors: SDK and native REST. The idea was simple: docs tell you what an API does, a recipe shows you what it looks like to actually use it. Copy a folder, drop in your API key, run it. A Node.js guide and a Python streaming guide followed later in the week, both on the Speechify developer blog. Three posts in five days, and none of them went on my personal site.
Outside the Speechify bubble, the week belonged to Washington. OpenAI launched GPT-5.6 on Thursday as a three-model family (Sol, Terra, Luna) but limited the preview to about 20 government-approved partners. The White House had asked for the restriction after Sol scored 96.7% on OpenAI’s internal cyberattack benchmark. Anthropic got a narrow clearance the same day to redeploy Claude Mythos 5 for critical infrastructure, two weeks after export controls pulled it offline. OpenAI framed the arrangement as temporary. It did not feel temporary.
OpenAI and Broadcom unveiled Jalapeño, a custom LLM inference chip designed in nine months with AI assistance. Early tests running GPT-5.3-Codex-Spark showed better performance per watt than current alternatives, and the target is gigawatt-scale deployment by late 2026. Meanwhile Google Gemini 2.5 Pro with Deep Think landed on June 22, becoming the most capable model freely accessible while its competitors navigated government access controls.
Figma held Config 2026 and shipped code layers, native animation support, and custom AI skills. DeepMind lost four elite researchers in one week including Noam Shazeer and John Jumper, a turnover event that wiped roughly $270 billion from Alphabet’s market cap on aggregate. The pattern in all of this is that AI model releases are no longer just product decisions. They are regulatory events with national security implications, and that changes the calculus for every developer building on these platforms.
I joined Speechify because the product was good and the team was small. Watching the rest of the industry pivot toward government gatekeeping while I sat in onboarding felt weirdly validating. Not because regulation is straightforward (it is complicated) but because the timing made it clear that voice AI is infrastructure now, and infrastructure requires long-term thinking. Speechify’s TTS API ships on a platform that has been public since June 11 with Voice Agents, and the cookbook, the docs, and the guides are all designed for developers who want to build without waiting for government approval lists.
The cookbook is at github.com/SpeechifyInc/speechify-api-cookbook if you want to try it. More recipes landing next month.
FAQ
What is the Speechify Cookbook?
A public GitHub repository of runnable recipes for the Speechify TTS API. Each recipe is a self-contained folder with both SDK and native REST examples in TypeScript and Python. Clone the repo, copy a folder, set your API key in .env, and run it. The cookbook launched on June 22 with quickstart, streaming, voice cloning, SSML emotion, and speech marks recipes across three languages.
What did OpenAI release as part of GPT-5.6?
OpenAI launched three models on June 26: Sol (flagship, $5/$30 per million tokens), Terra (balanced, $2.50/$15), and Luna (fast and affordable, $1/$6). Sol also introduced two reasoning modes: max for deeper reasoning and ultra for coordinating sub-agents. The entire family was limited to about 20 government-approved partners at the White House’s request, with broader access expected within weeks.
What happened with Anthropic’s Mythos 5 that week?
Commerce Secretary Howard Lutnick authorized Anthropic to restore Mythos 5 for roughly 100 US organizations operating critical infrastructure on June 26. The model had been taken offline on June 12 under export controls triggered by its cybersecurity capabilities. The authorization required Anthropic to implement safeguards and submit to ongoing review. The decision did not apply to Fable 5, which remained suspended.
What is the OpenAI Jalapeño chip?
A custom LLM inference processor developed with Broadcom, designed in about nine months with help from OpenAI’s own models. Early lab tests running GPT-5.3-Codex-Spark showed substantially better performance per watt than current alternatives, though final benchmarks are pending. OpenAI aims for gigawatt-scale deployment by late 2026, representing its clearest move yet toward owning the full compute stack.