Aleph Alpha's Kolibri: 78B open MoE, Apache 2.0, 1M context
Germany's Aleph Alpha dropped Kolibri, an Apache 2.0 MoE with only 3.46B active parameters and a 1M-token window.
LLMs
Model releases, benchmarks and capabilities. 106 stories, page 1 of 5.
Xiaomi's MiMo-V2.6-Pro is now the smartest open-weight model
Xiaomi's MIT-licensed model scores 46 on Artificial Analysis, one point behind GPT-5.6 Sol, at about 1/15th of the cost per task.
Nano Banana 2.5 Flash spotted in Google Flow before any launch
References to an unannounced Nano Banana 2.5 Flash image model have appeared in Google Flow. Google hasn't said anything about it yet.
GPT-6 Sol vs Opus 5.5: OpenAI and Anthropic cut prices the same day
OpenAI halved GPT-6 Sol and Luna prices, and Anthropic's Opus 5.5 matches Fable 5.1 on most tasks for 40% less than Opus 5.
OpenAI drops GPT-6 Sol at half Opus 5.5's price, 90 min later
GPT-6 Sol and Luna cost half as much as GPT-5.6 and landed 90 minutes after Anthropic's Opus 5.5, turning launch day into a price war.
Claude Opus 5.5 lands: Fable-level coding at 40% lower cost
Anthropic's Opus 5.5 matches Fable 5.1 on most tasks, tops agentic coding and computer-use benchmarks, and costs 40% less than Opus 5 to run.
GitHub's Copilot app runs several coding agents side by side
Each Copilot app session gets its own git worktree, so several agents can code at once while you review and merge their work.
Grok Voice Transcribe 2.0 halves errors and keeps the $0.10/hr price
SpaceXAI's new speech-to-text model tops 32 streaming rivals on Artificial Analysis and costs the same as v1.0: $0.10/hr batch, $0.20/hr streaming.
A creeper broke GPT-6 Astra. It went to farm potatoes
OpenAI's GPT-6 Astra got further in Minecraft than any AI yet. Then a creeper blew up its stash, and it spent hours farming potatoes.
Perplexity's local AI agent lands on Windows — bring a 24GB RTX
Portable Computer now runs fully on-device on Windows PCs with a 24GB+ RTX GPU, adding local MCP servers and scheduled tasks.
Grok 4.8 is a 2.5T-parameter model, and it's still in the oven
Musk says Grok 4.8, with 2.5T parameters, finishes pretraining this week and then moves to RL. Grok 4.7 still hasn't shipped.
GPT-6 Astra just beat all 48 levels of 'I'm Not a Robot'
OpenAI's new model cleared every level of Neal Agarwal's CAPTCHA game. Impressive computer use, but the viral clip doesn't mean CAPTCHAs are dead.
DeepSeek's 'smallest' new model is 552B — and it beats V4 Pro
DeepSeek open-sourced V4.1 Flash under MIT: a 552B multimodal MoE running just 8-16B active params that it says tops its own V4 Pro.
GPT-6 Astra turned open anatomy data into a 2,234-piece 3D body
Astra didn't model the human body — it built the explorer around meshes that have been free and unused for years.
GPT-6 Astra Can Actually Use Your Computer, And People Noticed
OpenAI's GPT-6 Astra hits 72.6% on OSWorld 2.0 and drives real apps itself — and the viral stunts have already started.
GPT-6 Astra takes WebDev Arena #1, 35 points over Fable 5.1
OpenAI's new model scores 1,797 on Code Arena's WebDev board vs Claude Fable 5.1's 1,762 — at identical $10/$50 pricing.
OpenAI's most aligned model is also its hardest to audit
GPT-6 Astra cut unwanted behavior from 22% to 2.4% — while OpenAI's own tests found its reasoning harder to monitor.
OpenAI ships a zero-day hunter, then pledges $1B to defense
GPT-6 Astra is the first OpenAI model rated Critical for cyber risk. One day later, OpenAI put $1B behind the defenders.
Fable 5.1 Is Two Days Old and Devs Are One-Shotting GTA Clones
Anthropic's new frontier model doubled its agentic research score and cut cache reads 75% — and X is flooded with one-prompt games.
OpenAI ships GPT-6 Astra and calls it the start of AGI
Astra scores 98.6% on ARC-AGI-3, is OpenAI's first 'Critical' cyber-risk model, and Brockman says the AGI era is here.
Meta's Muse Spark 1.3 finally tops a coding leaderboard
Muse Spark 1.3 hits 75.4 on DeepSWE v1.1 — first place, above Gemini 3.8 Flash and GPT-5.6 Sol. Overall intelligence: still third.
Fable 5.1's party trick: tiny 3D dioramas Anthropic never pitched
Anthropic sold Fable 5.1 on agentic coding benchmarks. Devs are grading it on little one-shot Three.js dioramas instead.
Pliny leaked Fable 5.1's system prompt in under an hour
Anthropic's cheaper, less-restrictive Fable 5.1 shipped Sept 1. Its system prompt was reportedly public the same hour.
Claude Fable 5.1 more than doubles Anthropic's science score
Fable 5.1 hits 52.6% on Terminal-Bench-Science vs Fable 5's 24.7%, and cache reads just got 75% cheaper. Live today.