News archive
October 12, 2026
Microsoft's Copilot home adds Code and Autopilot, from $9.99/month
Microsoft's new Copilot home combines Code, Cowork and Autopilot, with plans from $9.99/month; makers can test whether one work surface replaces several focused tools.
$9.99 individual monthly starting price$23.50 business monthly price per user, paid yearly$30 enterprise monthly price per user, paid yearly
Google studies AMIE with 98 completed primary-care patients
Google's AMIE study had 98 completed patients, one reported hallucination and no safety stops; builders should copy the supervised pre-visit summary workflow, not sell autonomous diagnosis.
98 patients who completed the study1 reported hallucination87–100% clinical evaluator favourable ratings
JetBrains releases Mellum2.1, a 12B open model for coding agents
JetBrains' Mellum2.1 is a 12B Apache 2.0 coding model with 2.5B active parameters; download it and compare real repo edits with Qwen3.5-9B before paying for an agent.
12B total parameters2.5B active parametersApache 2.0 model licence
Qwen-Image-2.1-Turbo cuts image generation to 8 denoising steps
Qwen-Image-2.1-Turbo is a 7B open image model built for 8 denoising steps, so creators can test faster local generation and editing without a closed image API.
8 denoising steps7B model size
Nace opens Drex 1.5 for probability-based option scoring
Nace's Drex 1.5 returns a probability for each option in one pass and scored 58.28 on Nace's Decision Index; it is a direct alternative to asking a chat model to choose.
1 forward pass for an option score58.28 Drex score on Decision Index 0.2.171 models on that index
Anthropic lets Managed Agents spawn up to 1,000 agents per run
Claude Managed Agents can run up to 1,000 agents per execution, with 16-way concurrency in the cookbook; use the beta on small jobs first because token cost can scale sharply.
1,000 maximum agents per run16 concurrent agents in the cookbook example66 bugs found in one reported dynamic-workflow test
AutoSprite turns one character sprite into a game animation sheet
AutoSprite turns one sprite or a prompt into idle, walk and run animations, then exports PNG sheets and atlases for Unity, Godot and other engines.
Cliptalk's video agent turns scripts, links and photos into edited clips
Cliptalk advertises 140+ voices, 30+ languages and a free 500-credit starter tier, so creators can test a complete short-video draft before paying for a timeline editor.
140+ voices listed by the service30+ languages listed by the service500 free starter credits
GLSlop publishes 200+ CC0 GLSL shaders for remixable visuals
GLSlop puts 200+ CC0 GLSL shaders in a remixable browser collection with GIF, MP4 and WebM export, giving visual makers a faster starting point than a blank shader file.
200+ CC0 shaders in the collection3 listed export formats
Minerva's Roger is a teleoperated humanoid for bomb disposal
Minerva says Roger was built in 5 months, uses VR teleoperation with onboard balance and recovery, and is aimed at bomb disposal, hazmat and other dangerous jobs.
5 months reported build time for Roger3 example dangerous operating settings
Cloudflare acquires Deno and will end Deno Deploy in six months
Cloudflare will run Deno Deploy for six months, support the Deno runtime for one year, and keep JSR alive; Deno Deploy users should plan a Workers migration now.
6 months reported Deno Deploy operating period1 year reported Deno runtime support period$26M Deno funding reported before the deal
Carbon-A annotates genomes with a 1.2B-parameter open model
Carbon-A is a 1.2B-parameter genome-annotation model with a 98,304-base context window and 4.6GB weights; research teams can run it locally, but it is not a general biology chatbot.
1.2B parameters98,304 bp context window0.944 reported macro nucleotide F1 on 42 genomes
Odyssey-3 opens a 14B interactive world-model preview
Odyssey-3's public preview generates interactive environments from prompts; reports describe a 14B model at 832×480, so game and robotics makers can test stateful worlds before building a simulator.
14B reported base model size832×480 reported preview output size
Sakana's MASS trains models on their own multi-agent workflows
Sakana's MASS trained from 12 synthetic tasks and reached a reported 69.9% held-out win rate over its base model; the code is available, but SWE-bench transfer fell below baseline.
12 synthetic tasks in the reported suite69.9% second-round win rate over base on held-out tasks0.94× reported SWE-bench score per token versus base
Sakana's MLR catches 73% of planted core-claim errors at $0.47 per review
Sakana's MLR caught 73.43% of planted core-claim errors at four reviews but only 16.11% of exact matches on real withdrawn papers; the benchmark gap is the story.
73.43% planted core-claim errors caught with four reviews16.11% exact matches on real withdrawn-paper errors$0.47 reported cost per review
