After Intelligence — Edition 005
The flagships traded blows, the open stack answered back. Google shipped Gemini 3.8 Flash with a million-token window and DeepSWE-grade engineering chops, OpenAI crowned GPT-6 Astra its newest aligned flagship, and Moonshot pushed Kimi K2.7 Code into the open — a trillion-parameter agentic coder you can run yourself.
Gemini 3.8 Flash — Google's million-token workhorse gets agentic
Google's most intelligent Flash model to date, and the first to make a serious claim on end-to-end software engineering.
Gemini 3.8 Flash landed September 2 with a 1,048,576-token context window — the full million — and an introductory price of $0.75 per million input tokens, matching 3.7 Flash and holding through December 31, 2026 before doubling to $1.50. The pricing is the boring part. The interesting part is where Google says the gains landed: software engineering, agentic tasks, and multi-step reasoning, all areas where Flash-class models have historically been the B-team.
On DeepSWE v1.1, Google reports 3.8 Flash solving engineering problems end-to-end better than most larger frontier models — a claim that, if it holds up under independent evaluation, reframes Flash from "cheap routing tier" to "default agentic substrate." A million tokens of context at Flash latency and Flash pricing is the kind of bundle that makes long-horizon agent loops economically viable in a way they haven't been before. The open-weights crowd now has a concrete capability target to clear.
Read more → Google AI updates
GPT-6 Astra — OpenAI's new flagship, benchmarked and aligned
OpenAI's newest flagship, billed as "the world's most intelligent and aligned model," and by The New Stack's read, the benchmarks agree.
GPT-6 Astra's launch messaging leans hard on alignment — the framing is less "smarter" than "smarter and safer at the same time," a combination OpenAI has historically struggled to deliver simultaneously. The New Stack's early benchmark pass suggests the model lands where OpenAI claims it does, which is itself notable: the last several flagship cycles have arrived with marketing that outran the evals.
Details on architecture, pricing, and context length remain thin in the public writeup, but the positioning is clear — Astra is OpenAI's answer to a frontier tier that now includes Gemini 3.8 Flash on the efficient end and whatever Anthropic ships next on the agentic end. The question for builders is whether "aligned flagship" translates into the kind of tool-use and long-horizon reliability that actually matters in production, or whether it's a safety story layered over a capability bump. Early signs point to the former.
Read more → The New Stack
Kimi K2.7 Code — Moonshot's trillion-parameter agentic coder, open on Hugging Face
A ~1T-parameter mixture-of-experts coding model, 32B active across 384 experts, 256K context, Modified MIT — and you can pull it today.
Moonshot AI's Kimi K2.7 Code is the open-weights answer to a week dominated by proprietary flagship launches. The architecture is familiar — same MoE skeleton as prior Kimi releases — but the headline improvement is efficiency: roughly 30% fewer reasoning tokens for equivalent task performance. For agentic coding loops that pay per token in latency as much as in dollars, that's the metric that matters.
The 32B-active footprint puts K2.7 Code in the runnable-on-serious-hardware band, not the "wait for a distilled variant" band. Modified MIT licensing keeps it commercially usable. With 256K context it won't challenge Gemini 3.8 Flash's million-token window, but for repo-scale coding tasks — where the relevant context is usually a codebase, not a corpus — 256K is the right neighborhood. The open stack now has a credible agentic coder at frontier scale, released the same week the closed labs showed their newest hands. That timing is not accidental.
Read more → Kimi K2.7 Code on Hugging Face