007
Changelog
— What we shipped, and when.
New models, a faster engine, a calmer console. Every release, newest first, with what it changes for you.
14 October 2026
New
Arc-7 Inference
Our largest model yet, on the new inference engine: speculative decoding, paged attention and a cache that keeps your system prompt warm.
256K context at the same price per token
Tool calls stream as they are written
Batch API at half the price
30 September 2026
Improved
Console 4.2
Every agent run is kept as a trace you can open, step through and run again against another model.
Replay a run against any model
Diff two runs side by side
Share a trace by link
12 September 2026
New
Guardrails 2.0
Write what a model must never say in a sentence; Guardrails turns it into a check that runs on every answer in under 10 ms.
Policies in plain English
PII masked before the model sees it
Every block logged with its reason
21 August 2026
New
Regions
A fourth region, with H200 capacity and data kept in Singapore for teams under its rules.
ap-southeast-1 for every plan
Latency under 40 ms across Southeast Asia
30 July 2026
Improved
Fine-tuning
A long run that loses a node now resumes from its last checkpoint on its own, and you are billed only for the minutes that counted.
Automatic resume
No charge for lost minutes
Evaluations after each epoch
2 July 2026
Fixed
SDK 3.0
The Python and Node SDKs speak the OpenAI format, so moving to Arclight is a change of one line.
OpenAI-compatible endpoints
Typed responses in TypeScript
Retries with backoff by default
One email when something ships, never more than two a month.
Or read the docs