
Alibaba launches Qwen4-Max with 1M-token context and 62% lower task costs
Alibaba says Qwen4-Max uses up to 35% fewer tokens than its predecessor, but it trails Anthropic and Moonshot on agentic coding benchmarks.

Alibaba says Qwen4-Max uses up to 35% fewer tokens than its predecessor, but it trails Anthropic and Moonshot on agentic coding benchmarks.

Will DePue's forensic report says DeepSeek, Moonshot, MiniMax and Ant Group models released since April 2026 were initialized on export-controlled Nvidia GPUs.

The new Gemini Decision 1 model answers up to 1,000 yes-or-no, choice or rubric questions per request, billed at $0.10 per million input tokens.

Gemini-powered agents produced about 6.5 million lines of machine-checked Lean proof in four weeks, and the Lean artifacts are being open-sourced.

Google's strongest model moves past its cyber-defender preview to paying API customers, enterprises and AI Ultra subscribers, with prices doubling after Nov. 30.

Priced at $0.10 per million input tokens, Ember ties Claude Haiku 5.5 on the Artificial Analysis index and ships with a 1M-token context window.

Fable 5.2 ships to all users with outputs up to 30% faster, while Mythos 5.2 stays limited to vetted cyber and life sciences organizations.

The two new eighth-generation systems make four TPU launches in 2026, a first for Google, as TPU customers near 100.

In Anthropic's CL-Bench test, a memory store lifted Claude Opus 5.5 from 0.655 to 0.838, and idle-time review helped fix bad notes.

The Physical Superintelligence system had a complete, Lean-checked proof on day three of its run, and Stanford researchers are now reviewing it.

An 11-week high-access evaluation puts the speedup at 1.9x since October 2025 and attributes perhaps 80% of it to Anthropic's own use of AI.

The model runs 17B active parameters, and the lab says Apache 2.0 weights follow next week once external red-teaming wraps.

The pre-training checkpoint activates 14B parameters per token, was trained on 33.5T tokens and ships with a 1M-token context configuration.

The video model adds multi-turn conversational editing and tighter reference following, priced at $0.07 per second of 720p video across the Gemini API and Cloud.

Two new developer models, Flare and Sunburst, arrive alongside reusable reference images, image markup, SVG export and up to 50% faster generation.

The new Devin coding model runs three times faster than SWE-2, is built on Thinking Machines' Inkling-Small, and is free through November 30.

The official Hy4 adds vision and computer use, cuts its internal hallucination rate to 7.2%, and keeps the preview's API price.

Astra handled the enclosure, circuit board, firmware and interface; a contract manufacturer delivered prototypes in under a month after several revision rounds.

The model activates about 48B parameters per token and scored 91.9 on BrowseComp, according to results Meituan reported.

The 11-domain agentic test runs models for up to 8 hours in simulated work environments; Claude Opus 5.5 is the best all-around performer.

Same weights as GPT-6 Astra, fewer dual-use refusals and a jump to 49.6% on MedChemBench, but only for vetted labs through trusted access.

Most rules become defaults that developers and users can change, detailed style guidance is cut, and hard limits stay only where OpenAI deems them necessary.

The checkpoint lands on Hugging Face three weeks after preview, scoring 41% on Terminal-Bench 4.0 and priced at $0.70 per million input tokens via API.

The report names Kimi K3.1 Pro and DeepSeek V4.1-Flash 1017, a month after OpenAI tied a reasoning-extraction campaign to Moonshot.

Carbon arrives two weeks after Gemini 4 Argon, goes first to cyber defenders, and keeps Argon's $2/$10 introductory per-million-token pricing.
The tightly edited AI news email engineers, founders, and investors actually open. One email. Every weekday. Five minutes to finish.
The briefing read inside teams at