News
What shipped,
and what
it changes.
Releases matter here only to the degree they change how you build. Every entry says what happened; the ones that affect your work say why.
- Moonshot AI
Kimi K2.8 Preview
Preview release of Moonshot's next K-series model.
Worth watching if you are cost-sensitive — the K series has consistently undercut Western frontier pricing.
Model - Shanghai AI Laboratory
Atria Dawn Preview released open source
Open-source preview from Shanghai AI Lab.
Open weights - DeepSeek
DeepSeek-V4.1-Flash
Open-source release optimised for speed.
Flash-class open weights are the practical option for the wide, cheap half of an agent pipeline.
Open weights - OpenAI
GPT-6 Astra
New proprietary frontier release from OpenAI.
A frontier release is the moment to re-run your evals, not to rewrite your prompts.
Model - Google
Gemini 3.8 Flash, plus a Cyber variant
Next-generation flash model with introductory pricing through 31 December 2026. The Cyber variant loosens some safeguards and requires Fairwind Program approval.
Introductory pricing with an expiry date is a trap if you build cost assumptions on it.
Model - Meta
Muse Spark 1.3, and a Contributor tier
Conversational model that asks clarification questions before acting. The discounted Contributor tier trains on your traffic.
Read the tier terms before the discount tempts you — training on your traffic is the price.
Model - Anthropic
Claude Fable 5.1 and Mythos 5.1
Both ship with three breaking API changes. Mythos is gated behind Anthropic's verification programmes.
Breaking API changes are exactly what the teardown session exists for.
Model - Perplexity
Hybrid compute with PPLX Qwen 3.8 27B
Cloud-first with on-device fallback on Apple silicon, aimed at sensitive workloads.
Platform - Inception
Mercury 2.5 Preview
A diffusion language model, launched with an 80% discount that expired 8 September.
Model - Z.ai
GLM-5.3 open weights, 753B
Weights released under a custom licence after a safety evaluation hold.
Open weights - Alibaba
Qwen 3.8-Flash, 1M token context
Flash-class release with a one-million-token context window.
Open weights - IBM
Granite 4.2 8B
Dense open model with selectable reasoning effort and a 131K context window.
Open weights
Snapshot taken September 2026 · wire this to a feed before launch or it will rot