Launches on 2026-09-18

Takes text, images, audio and video in one request across a million tokens of context — up to an hour of continuous audio-video per call, at $0.15 per million input tokens.

An 8-29 MB automation model that runs tool calls and structured extraction on a phone, a wearable or a microcontroller, with no API bill at all.

Lets an agent run a program across roughly 164 billion rows of Reddit, Hacker News, arXiv and Wikipedia instead of fetching pages one at a time — priced by the second, and by how busy the machines are.

Grades how far a coding agent actually gets with your product, and records the real sessions where it fails.

Command-line tools for agents that load a command definition only when the agent asks for it, instead of spending tokens on a tool list it never uses — MIT licensed.

An outbound sales agent that keeps running after the campaign starts — reads the results, drops what is not working and proposes the next angle. From $250 a month.

Voice typing that adapts to whatever app you are writing in, and transcribes the meeting too — free on macOS, with Windows and iOS still to come.

A chief of staff that captures meetings, decisions and commitments as they happen, then chases the follow-through — free to start, $1 a day after.

Give it an outcome rather than a task and it decides what needs to exist — an app, an agent, an automation — then runs it for months on a budget you set. Free to start, $29 a month for more.

Fills in accelerator applications from one startup profile across 117 programmes — free for two a month, $29 for twenty.

The same GLM-5.3-Flash served faster: 200 tokens a second, 2.5x the price of the model it is built from, and no claim of extra capability anywhere in the vendor's own documentation.

A 2.8 MB model that decides which value belongs in which form field, in one forward pass, on the machine doing the work — 706,048 parameters, MIT licensed, with no API call and no bill.

An open-weight decision model that answers typed questions with probabilities in one forward pass, about 33 ms on a T4 — and whose own model card shows where it fails while staying confident.

Alibaba's streaming audio-and-video conversation model with spoken replies, tool calling and remote MCP, announced 18 September 2026; audio input costs $0.93 and audio output $1.87 per million tokens in Singapore.

xAI's second-generation speech-to-text model, released 18 September 2026, for batch and streaming transcription with diarisation and up to eight channels, priced at $0.10 per audio hour batch and $0.20 streaming.

Bad Theory Labs' open-weight fine-tune of Qwen3.8-Flash-Next for terminal and software-engineering agents, 177B parameters with 6.6B active and a 256K context, under the Qwen Community License; its benchmark claims are self-reported.

Plane now lets an AI agent be an assignee on a work item like any other teammate, picking up triage, ownership, blocker-tracking and status updates inside the project rather than beside it.