Launches on 2026-09-23

Google’s expressive text-to-speech model: 130 languages, voice design from a written description, $0.50 in and $9.00 out per million tokens until 31 December 2026 — and double that from 1 January 2027.

The cheaper Gemini 3.8 speech model for dubbing and voice agents: 101 languages, $0.50 in and $6.00 out per million tokens until 31 December 2026, doubling from 1 January 2027. Replaces the 3.1 Flash TTS preview.

The speed and cost tier of Recraft’s V4.1 text-to-image family: $0.007 an image through the API, against $0.035 for standard V4.1, with a median of 1.3 seconds per image on Recraft’s own tests.

Black Forest Labs’ 7B open-weight model that turns camera views, joint positions and a plain-language instruction into robot-arm actions. Commercial use is free only below US$5 million in annual revenue.

Hosted webhook infrastructure from SnapStack Technologies: capture events at the edge, inspect, transform and replay them, and deliver to private servers over outbound mutual TLS. Free to 10,000 requests a month.

A post-mortem debugger for Windows x86 and x64 crash dumps that interprets memory for the engineer — types, call stacks, heap metadata — and exposes its findings to AI coding assistants over MCP. Private beta.

Turns an OpenAPI, Swagger or Postman specification into a hosted Go MCP server behind a gateway with OAuth, audit logging and rate limits, so AI agents can call an existing API. Free for five servers.

Large-file delivery with links that can be changed after sending, signed proof-of-delivery certificates, per-recipient watermarks, country and identity restrictions, and a choice of US or EU storage.

NVIDIA’s open-weight 100M speaker-diarization model: it labels who spoke when for up to eight speakers, streaming or offline, and ranked first of twelve systems on Voice Arena’s independent early results.

Fireworks AI’s retrain of Kimi K3 to reason with fewer tokens: about 40% fewer on the developer’s own tests at similar quality, $3 in and $15 out per million tokens, as a two-week research preview.

Together AI’s experimental open 4B decision model, trained to pick one option from a list the way TypeSafe’s Jev does, with the full recipe published. The licence for the weights is still being finalised.

An open Apache-2.0 “contrastive language model” that scores candidate actions against a situation, so an agent can pick or check its next step without generating text. The authors report up to 9x lower latency than Jev.

An open 0.6B Swedish speech-to-text model from Klang AI, trained on about 50,000 hours of Swedish, with word error rates near 4.5% on the developer’s own tests and a CC BY 4.0 licence that allows commercial use.

A single 0–100 quality score for an AI agent in production, recalculated daily from outcome, reliability, cost, speed and safety, with the sessions behind it — a feature of Latitude’s agent monitoring, in early access.