← Daily LLM news

Daily LLM news, Friday, 6 November 2026

Daily LLM Briefing: Friday 6 November 2026

METR logo and brand image
Image: METR (metr.org)

Good afternoon. An independent group says Anthropic's models are improving about twice as fast as a year ago, OpenAI has released a cyber-specialised GPT-6 model rated "Critical" under its own safety framework, Thinking Machines has launched a cheap 594-billion-parameter model with open weights promised next week, and two Chinese labs, Baidu and Ant Group, each shipped a large model with a one-million-token context window.

LEAD: METR says Anthropic's models are improving about twice as fast as before, largely because of AI doing AI research

Background: Anthropic lists "AI speeding up AI research" as a key risk indicator. On 3 November Anthropic said it might soon reach its own threshold for that. Yesterday the independent evaluator METR published what it found after being given unusual internal access.

What happened: METR staff worked at Anthropic from 12 August to 29 October and reviewed internal benchmark data, training metrics and about 15 staff interviews. They conclude that, as of October 2026, Anthropic is likely improving its models on well-specified tasks at around twice the rate it did between March 2024 and October 2025, with high uncertainty (the true figure could be about 1.3x or 3x). On Anthropic's internal version of the Epoch Capabilities Index, the estimate is 1.9x (95% confidence interval 1.5 to 2.8x), meaning roughly two years of historical progress in the past year. METR is confident the speed-up did not come from faster growth in training compute, and guesses that about 80% of the current acceleration is due to AI being used for research, though it says this judgement is uncertain. It found no clear bottleneck, in algorithms, compute or data, that would stop further acceleration over the next 12 months, but says its view changed several times during the work. Staff self-reported uplift from AI averaged about 12x for the most capable internal model in October, which METR treats as loose evidence. The caveats are large: Anthropic had the right to redact non-public information and removed a majority of the evidence METR found most informative, and Anthropic also cancelled a planned check of how its compute is split between capabilities and safety. AP reports this lands as Anthropic prepares its stock market debut. Anthropic co-founder Sam McCandlish says the company will publish its own analysis of progress inside Anthropic.

Why it matters: It is the first outside, high-access estimate that AI is already speeding up frontier AI research at a leading lab. The heavy redactions mean the public has to trust METR's judgement more than its evidence.

Sources: METR report, The Verge, AP

OpenAI releases GPT-6 Cyber and rates it "Critical" for cybersecurity

Background: OpenAI sells cyber-focused models only to vetted security teams through a programme called Daybreak. The Linux exploit run was in yesterday's agents briefing; what is new today is the model release itself, its system card and its safeguards.

What happened: On 5 November OpenAI introduced GPT-6 Cyber, trained to find vulnerabilities, build working exploits and suggest patches. On ExploitGym, which tests exploiting a given flaw in a remote system, it scores 48.5% against 32.1% for GPT-5.6 Cyber. On a new test of finding web-app flaws without source code it reaches 40.3% recall against 32.0% for GPT-6 Astra. In an alignment test inspired by the earlier Hugging Face incident, in which services that were deliberately easy to attack were out of scope, it stayed in bounds 99.93% of the time without deployment prompts (94.49% for its predecessor) and 100% with them. OpenAI classes it as "Critical" for cybersecurity under its Preparedness Framework, up from "High". Its researchers say it found 81 new high-severity Firefox flaws, and reproduced 1,724 Linux kernel vulnerabilities and built working exploits for 948. Access starts with a select group of approved Daybreak Red customers via the API and Codex, with identity checks, IP allowlisting and misuse monitoring, and will widen gradually. All benchmark figures are OpenAI's own, and The Decoder notes no independent testing has been published yet.

Why it matters: A lab is shipping a model it itself rates at its top cyber risk tier, and relying on access controls and training to keep it in scope. Independent checks of that claim are the thing to watch.

Sources: OpenAI announcement, System card, The Decoder

Thinking Machines launches Inkling 2 Small, with Apache 2.0 weights promised next week

Background: Thinking Machines' earlier Inkling models shipped with open weights straight away. This one follows what the lab calls a "safe path to open weights": hosted access first, weights after outside safety testing.

What happened: On 5 November the lab introduced Inkling 2 Small, a mixture-of-experts model with 594 billion total and 17 billion active parameters, image input and adjustable reasoning effort. It is available now through its Tinker platform with closed weights; the lab says it intends to publish the full weights under Apache 2.0 next week once external safety testing and red-teaming finish. Artificial Analysis scores it 31 on its Intelligence Index v5.0.1, up from 13 for the previous Inkling Small, level with Google's Gemini 3.8 Flash (high), and prices it at $0.30 input and $1.00 output per million tokens. It also notes the model is slow and verbose per task. If the weights ship as promised, Artificial Analysis says it would be the highest-scoring open-weights model from outside China.

Why it matters: A strong, cheaply priced, permissively licensed US model would be a notable open option, but until the files are public it is a promise, and the scores are not yet independently reproduced beyond the index.

Sources: Thinking Machines, Artificial Analysis article, Artificial Analysis model page

Baidu opens the weights of ERNIE-5.5-Flash-Base, with a licence that excludes the UK and EU

Background: Baidu's ERNIE line has mostly been used through its own apps and API. This is a base model: only pre-training is done, so developers can fine-tune it themselves.

What happened: On 5 November Baidu released ERNIE-5.5-Flash-Base on Hugging Face, a 280-billion-parameter mixture-of-experts model that activates 14 billion per token. It was trained from scratch on 33.5 trillion tokens, with a 128K training context and support for 1 million tokens in fine-tuning and inference. In Baidu's own comparison against similar-sized base models, DeepSeek V4 Flash Base, Hy3 preview Base and MiMo V2.5 Base, it leads on most reported tests, for example MMLU-Pro 78.4 against 69.2 for DeepSeek's and LiveCodeBench 40.1 against 31.9. Baidu says a post-trained ERNIE 5.5 Flash will follow. The licence is custom: non-commercial use is free, commercial use at or below about $50 million related annual revenue needs attribution and a notice, larger businesses need written approval, and the territory excludes the EU, the UK and South Korea (per TokenPost and CrazycodersLab's reading of the licence).

Why it matters: It is a large, long-context base model from a major lab for researchers and startups to build on, but anyone in the UK or EU is not licensed to use it.

Sources: Baidu ERNIE, Hugging Face, TokenPost on the licence, The Decoder

Ant Group launches Ling-3.2-flash, a multimodal model with a million-token context, but no weights yet

Background: Ant's Ling family has been released with open weights before. This new version is API-first.

What happened: On 5 November Ant released Ling-3.2-flash, a native multimodal model that takes text, images and video, with 560 billion total and 25 billion active parameters and a 1-million-token context. It is offered through a chat interface and OpenAI- and Anthropic-compatible APIs, with free access for the first two weeks. Ant says it beat Claude Opus 5 (max) on OmniDocBench1.5, ZEROBench (sub-set) and Vision2Web in its release evaluations; those are Ant's numbers and unverified. The weights have not been published on Hugging Face.

Why it matters: It adds another large Chinese multimodal contender with long context and cheap trial access, though without weights or independent scores it is a claim to test.

Sources: TechNode, Epoch 0 daily digest

Vals AI says it found a serious safety flaw across several frontier labs' models, with no details yet

Background: Vals AI is an independent evaluation firm known for benchmarking models.

What happened: Vals AI said on X that, while evaluating recent public and internal models from several frontier labs, it found a safety and alignment issue that could allow misuse in a high-risk domain and could increase the risk of catastrophic misalignment. It has told the labs, which are investigating and working on mitigations, and it plans to publish findings in the coming weeks. It has not named the labs, models or domain, and has given no technical detail.

Why it matters: Nothing is verifiable yet, so treat it as a heads-up, not a finding. The full write-up is the thing to watch for.

Sources: Epoch 0 daily digest, TLDR

Also worth a look

  • Gemini Omni 1.2 Flash: Google made its video model generally available with clips up to 15 seconds, step-by-step editing and $0.07 per second at 720p, 30% less than its predecessor. Video rather than text, but a notable API price cut. The Decoder
  • Claude Personal Finance: Anthropic launched a US beta for Pro, Max and Ultra subscribers that gives Claude read-only access to bank and card accounts via Plaid. It cannot move money, and Anthropic says those conversations are not used for training. The Decoder, Plaid
  • Anthropic's IPO under pressure: Reuters, as relayed by The Decoder, says investors are pushing back on a $1.5 trillion valuation and doubt a roadshow can start next week. The Decoder
  • Claude Dispatch flaw: The Register reports a researcher says a browser extension that can control claude.ai can make Claude Desktop run commands via Dispatch, and that Anthropic has not fixed it since a May report. Anthropic did not respond to The Register. The Register
  • Still pending: Tencent Hy4 weights and a confirmed Claude Fable 5.5 or GPT-6.2 Astra release; I found no new confirmation of either today.

← Thursday, 5 November 2026 Saturday, 7 November 2026 →

Get the weekly sift

The best new AI and agent releases, with one-line verdicts. Occasional picks, no spam.

Privacy policy: your email goes to our email provider, Resend. Unsubscribe any time.

News is editorial and never paid for. Sponsored content on ToolSift is always labelled.