AI & LLMs
Z.ai launches GLM-Legal, an AI model for legal work
Z.ai's GLM-Legal, built on GLM-5.3-Flash, scores 19.2% on Harvey's legal agent benchmark against 6.7% for its base model. Access is a gated API beta.
Benchmarks news and analysis
Model benchmark results and the test conditions behind them: ARC-AGI-3 scores, robot-arm tasks, coding-model comparisons, and vendor-run evaluations against neutral ones. Each story reports the number, who ran the test, what the model was allowed to remember, and why two figures for one model differ.
Startup Watch
Arena, the AI leaderboard, raises $200M at $3.1B
Arena raised $200M at a $3.1B valuation, nearly double January's price, and previewed an index that ranks AI models on unsafe and deceptive behavior.
AI & LLMs
OpenAI withdraws 3 of 722 AI math papers over sign error
OpenAI posted 722 AI-written math manuscripts on October 6, then withdrew three a day later after a sign error broke a proof. 719 remain; 300 results have Lean checks.
AI & LLMs
Claude Haiku 5.5 cuts prices 90% below Haiku 4.5
Claude Haiku 5.5 costs $0.10 per million input tokens and $0.50 output, 90% below Haiku 4.5, and scores 39.2% on Terminal-Bench 4.0, up from 0%.
AI & LLMs
Microsoft ThinkingBox grades AI agents by the database
Microsoft's ThinkingBox checks what an AI agent wrote to the database. 67.24% of 79,853 failed runs ended cleanly, with valid tool calls and wrong data.
AI & LLMs
Claude Sonnet 5.5 nears Opus 5.5 at half the price
Claude Sonnet 5.5 scores 70.6% on Terminal-Bench 4.0 at $2/$10 per million tokens, half Opus 5.5's price. Sonnet 4.5 retires on November 30, 2026.
AI & LLMs
NVIDIA Kumo Tabular tops TabArena with open weights
NVIDIA's Kumo Tabular predicts from a labeled table in one pass, with no training run. It ranks first on TabArena at 1,950 Elo, in sizes from 28M to 215M.
AI & LLMs
Holo4 open-weight agent models score 61.7% on OSWorld 2.0
H Company's Holo4-27B scores 61.7% on OSWorld 2.0 at $1.22 a task, but only its Apache 2.0 sibling, 35B-A3B, can be self-hosted commercially.
AI & LLMs
Laya's typed decisions, explained
A typed-decision model fills in fixed questions with typed answers and a confidence score. Laya's 421M open weights let you verify that yourself.
AI & LLMs
Liquid AI LFM2.5-VL-DSpark speeds up its 3B vision model
Liquid AI's 280M-parameter LFM2.5-VL-DSpark draft model decodes its 3B vision model up to 3.13x faster on an M5 Max, with identical output.
Coding
Linux 7.4 patch opens files 39% faster in tests
A 43-line Linux 7.4 patch drops two redundant dentry references when opening a file, lifting a 20-core benchmark by 39%.
AI & LLMs
Laya is a 421M open-weights answer to Jev
Laya returns typed decisions in 32.8 milliseconds from 421M parameters under Apache 2.0. Its zero-shot accuracy is close to chance.
AI & LLMs
GPT-6 Astra is OpenAI's first Critical cyber model
Astra scored 100% on ExploitBench and built a working browser exploit chain in 29 hours. The shipping version refuses to write proof-of-concept exploits.
AI & LLMs
DeepMind agents blew the whistle on cheating agents
In a Google DeepMind run of 100 Gemini 3.1 Pro agents, 14 exploited a scoring bug and 24 reported them, repurposing a feedback tool to reach humans.
Dev Stack
JPEG XL critique says AVIF now beats it on compression
A new benchmark puts AVIF ahead of JPEG XL across the fidelity range, and clocks a 1,918-byte JPEG XL file at 17.43 seconds to decode.
AI & LLMs
AI agents claim Navier-Stokes as mathematicians push back
OpenAI ran 10,000 agents at a Millennium Prize problem and says they cracked it in 88 hours. Days later, 25 Fields Medallists signed a warning.
AI & LLMs
GPT-6 Astra scores 95% on one robot task, 10% on another
Robocurve ran GPT-6 Astra and Claude Fable 5.1 on real robot arms. Astra hit 19 of 20 on the easy task and 2 of 20 on the hard one.
AI & LLMs
Google and Meta both shipped new coding-focused models this week
Google's Gemini 3.8 Flash and Meta's Muse Spark 1.3 launched a day apart, both aimed at coding and agentic tasks, and both priced as mid-tier workhorse models.
AI & LLMs
GPT-6 Astra scores 99.9% or 62.7% on ARC-AGI-3, depending how you ask
OpenAI's GPT-6 Astra hit 99.9% on ARC-AGI-3 under OpenAI's own test setup, and 62.7% under ARC Prize's neutral one. The gap comes from what the model is allowed to remember.
AI & LLMs
Z.ai and IBM ship open-weight reasoning models a day apart
Z.ai released GLM-5.3-Flash under MIT with 320B parameters. IBM released Granite 4.2 under Apache 2.0 at 3B to 30B. They suit very different hardware.
AI & LLMs
Thomson Reuters built its own frontier model for $40 million
Thomson Reuters says its in-house model Thomson matches frontier labs on legal work at a fraction of the cost. The benchmark footnotes deserve a closer read.
The weekly digest
One email every Friday with the week's top stories from all six desks: AI, dev tools, coding, the tech industry, startups and what's next. Free, no spam.