Claude Startups hits capacity; every application gets re-reviewed
Hundreds of thousands applied in days, freezing the $1K-credit and Team offers—plus a 43MB on-device voice-to-tool-call model and ttok retargeting GPT-6.
Copy markdown
Applied for Claude startup credits? Your application is being re-reviewed
Anthropic says it received "hundreds of thousands of applications over the past few days" and is over capacity on the $1,000 API-credit and Claude Team offers, so it is reshaping the program and re-reviewing every application—some statuses will change. Existing members keep Startup Stack access (up to $45K in third-party partner credits), biweekly office hours, and events; Max and Team subscribers keep their monthly Console API credits.
NeuDecide: a 43MB voice-to-tool-call model that runs on a Raspberry Pi
Neuphonic open-sourced (Apache 2.0) a 55.5M-parameter speech model that takes audio plus a list of JSON tool schemas and returns a tool call directly, with no transcript in between. It runs single-threaded on-device via ONNX—46ms time-to-call on an M3, 206ms on a Pi 5, ~150MB RAM—so you can add offline voice control to a robot, kiosk, or app this weekend by editing JSON instead of retraining (English only, up to 10 tools, 30s audio).
ttok 1.0 now counts tokens as GPT-6
Simon Willison’s token-counting CLI hit 1.0 and switched its default tokenizer from GPT-4 to the GPT-5/GPT-6 family (OpenAI hasn’t confirmed GPT-6 reuses it; a third-party test suggests it does). If you pipe prompts through `ttok` to stay under context limits, your counts just shifted—re-check any hard budgets in your scripts.