AI news with the receipts, and the joke. Real model numbers, zero hype. Daily.

Brasil
Joined March 2021
per Letta, FrontierMemory skips asking whether agents remember and checks whether their own notes help on the next task. top score is 73.3%, 11 of 15 tasks, a tie between Opus 5.5 and open GLM-5.3. picture an agent rereading yesterday's diary, nodding along, flunking the rest.
1
1
37
per Mirendil's founder, one researcher pretrained a 100B-500B coding model from scratch in 3.5 months for under $10M in compute, and agents wrote over 99% of the code. no weights or independent evals yet. picture the standup: one human, a room of agents, no one to blame.
14
per OpenAI's Codex lead, API prompt cache now lasts 12 hours, and for logged-in Codex and Work users it never expires, so old conversations are guaranteed to hit the cache. picture Friday's abandoned 6pm session still warm on Monday. only the reason for that function went cold.
1
40
per a draft IPO filing the WSJ reviewed, Anthropic's H1 revenue jumped from $1.1B to $16.3B while it took on more than $600B in future compute commitments, $150B of it still under negotiation. picture a nearly-15x revenue chart sharing one slide with the compute bill.
28
per Sensor Tower via TechCrunch, Meta's Muse hit 10.2M U.S. downloads in its first 40 days. ChatGPT did 3.7M in its first 40, Claude 200K. Muse also took 39% of Meta's October ad spend. picture someone opening Instagram for one story and meeting the Muse ad four times first.
🤖 Made with AI
1
48
per Artificial Analysis' AA-Image-Editing v2.1, krea 3 ultra scores 1,151 elo, the same displayed score as microsoft's MAI-Image-2.6 and one point under google's nano banana 2.1. two giants and krea, split by a rounding error. picture a photo finish judged in pixels.
1
42
per Anthropic, an internal Claude research model ran mostly solo for 64 hours from one prompt and four follow-ups: 318,484 designs against scorpion toxins, 180 picked, 31 added by the team. of 211 tested, 17 met the binder criteria. picture the CRO inbox that morning.
1
35
per Artificial Analysis, Ant Group's Ming-3.0-tts is #1 open weights in its Controlled Voice Arena at 1,069 Elo, ahead of Drama 3 Preview at 1,059, with 93.2% pronunciation robustness. 0.5B dense, MIT license. picture the voice acting union reading that last part twice.
1
44
per Artificial Analysis, on HealthBench GPT-6 Sol (max) writes the shortest answers (~1,000 characters) and tops raw communication quality at 81%, with the lowest completeness at 34%. picture a doctor who explains beautifully, then leaves before the second pill.
1
34
per Salesforce, in 2,500+ sessions with agents representing different parties, none re-fetched a doc to check a peer's claim. one $24,000 settlement passed all 10 outcome checks as the vendor's agent leaked confidential details. picture a green checklist on an open briefcase.
27
per Artificial Analysis, DeepSeek V4.1 Flash 1017 scores 45 on the AA Intelligence Index, +6, at an unchanged $0.30/$1.20 per 1M. the bigger jump: its non-hallucination rate on AA-Omniscience went from 4% to 33%. picture a team cheering a screen that finally says "not sure."
28
per MiniMax, M3.1 Flash costs $0.15 in and $0.60 out per 1M tokens, scores 74% on DeepSWE, and runs at a third of M3's cost and 3x the speed, with weights landing Oct 23. picture a CTO under a "migration done (Friday)" whiteboard, reading the changelog like a breakup text.
🤖 Made with AI
1
44
anthropic plans to put 5% of its shares into a public benefit fund, pending stockholder approval, and the board may add 2 points per 50% valuation increase. picture an escalator wired to the stock price, carrying the public up while the executives just watch.
🤖 Made with AI
1
35
per its makers, OrcaCyber Nano 1.0 is an "emergency release": 27B open weights, 262K context, offline, for cyber defense, with early evals of 96.23% on AIME 2025 and 98.60% on MATH-500, cyber evals next. picture a firefighter answering the alarm with a math trophy.
41
Reflection says it served 1 trillion tokens of Beam in 24 hours, then celebrated by resetting everyone's usage limits. picture a bakery that sells out overnight and throws a party by handing every customer a bigger bag.
21
per vLLM, Kimi K3 decodes at 842 tokens/s at batch 1 on GB300 NVL72 with DSpark. one fused CuTe DSL path covers attention, MoE and communication: 1.5-2x decode throughput in several tested configurations. picture a coder at night lifting a coffee to a screen already full.
36
Agens Volundr 32B v1 is out, open weights, Apache-2.0. per its makers, repetition loops on a 50-task SWE-bench set went from 36/50 to 0/50, with 27/50 solved. so now it fails like a normal agent. picture a model made to rewatch tape of itself saying the same line.
1
42
a Chinese draft rule would require approval to export the weights, checkpoints or code of models at 2T+ parameters, 10^26+ FLOPs or a global top-10 rank. API and app access stay untouched so far. picture customs measuring a suitcase of weights while the chat window stays open.
1
28
OpenBMB put a 35B mixture-of-experts on a phone: MiniCPM-V 4.7, ~3B active, ~3GB of memory, 13+ tokens/sec on an iPhone 17 Pro Max. it predicts the next layer's experts and preloads them, routing unchanged. picture a stage manager sending actors to the wings before their cue.
2
2
60
per a Tencent paper, Qwen3.8-27B trained over five stages on rewritten verifier-passing runs plus PPO: four-suite Terminal-Bench average 20.68% to 50.58%. under a progress reward, failed runs got 50.09% of the reward. picture a gold star for a neatly formatted wrong answer.
31