Train your own open decision model for $17: Together ships Tev1
Open weights, $0.042/M in and free output on serverless. Plus a 340M router that beats Jev, Exa's deep-research API, and 22 Indian languages in one ASR model.

Copy markdown
Tev1: an open router you can fork for $17
Together AI released Tev1-4B-experimental, a Jev-style decision and classification model built on Qwen3.5-4B — open weights, live on Together serverless at $0.042/M input and $0/M output. The kicker: the full data recipe is public, and fine-tuning your own on 37,840 examples across six datasets took ~25 minutes and cost $17.
A 340M decision model that runs on CPU
Fastino's GLiNER2.5-Decide is a 340M open-weight, encoder-based model for fast, deterministic classification: hand it typed questions and rules, get structured answers back. It reportedly beats 1B models — and Jev — at agent routing, and runs on CPU, so your routing layer no longer needs a GPU.
Exa Agent Ultra: deep research across thousands of sources
Exa's new effort:"ultra" tier fans out a swarm of subagents for exhaustive list-building and entity enrichment — Exa claims 2,451 companies found on its Company Find-All benchmark versus Opus 5.5's 146. It's a hosted API only; runs cap at $20 by default ($1-$100 via maxCostDollars) and take 30 minutes to about 3 hours.
One ASR model for all 22 Indian languages
Sarvam's Saaras V4 transcribes all 22 scheduled Indian languages plus global-accented English from a single model, with five output modes (verbatim, code-mix, transliteration, translation), real-time streaming at sub-150ms to first token, and diarization on batch. It's API-only for now (model="saaras:v4") at ₹30/hour.