TabLOID

OpenAI’s public repository lists 722 manuscripts in 372 families, produced by an unreleased internal model. The company says it tested roughly 4,000 problems and used about three hours of ChatGPT Pro thinking compute per result on average.

The catch: verification varies. Some results have machine-checkable Lean proofs; others do not. OpenAI explicitly warns that unformalized work may contain issues. This is a research release, not a blanket certificate that every claim is correct.

Why it matters: generating pages is getting cheap. Checking whether those pages deserve to exist is still a job. Your math teacher’s “show your work” era is thriving.

OpenAI’s repository · Release announcement

OpenAI introduced GPT-6 with Intelligent UI on October 7. Responses can include interactive diagrams, forms, charts and small tools such as calculators. The company’s examples include a bicycle breakdown, trip planning and a bill splitter.

Rollout starts with Plus, Pro, Business and Enterprise on October 7, then expands to Free and Go on October 8. Workplace settings can affect Enterprise access. This update concerns ChatGPT’s Chat experience; it does not change the models powering Work or Codex.

Why it matters: a tool you can adjust can explain more than another paragraph. These are the company’s launch claims and demos, not our hands-on verdict.

OpenAI’s announcement

Anthropic launched Claude Haiku 5.5 for high-volume work such as summaries, classification and routing. It calls this its fastest model at standard speed and says average running costs are about 75% lower than Haiku 4.5.

For prompts up to 100,000 tokens, listed API prices are $0.10 per million input tokens and $0.50 per million output tokens. Longer prompts cost more. Haiku also gains an adjustable effort setting.

Anthropic’s own launch table shows stronger results than Haiku 4.5, while larger models remain better suited to complex coding work. Treat benchmarks as vendor-reported results. “Cheaper per token” and “cheaper for your exact job” are not automatically the same thing.

Anthropic’s launch and pricing

Anthropic’s October 6 update combines its cyber access programs into three tiers. Defense Access covers work such as incident response and vulnerability analysis. Red Team Access adds authorized penetration testing. Specialized Access is reserved for a limited set of verified organizations testing high-risk systems.

Applications and safeguards depend on the tier. Red Team Access currently requires an organization; individual researchers can qualify for some defensive work. The program also has data-retention requirements and specific exceptions, so teams need to read the terms before connecting sensitive systems.

Why it matters: better tools for people fixing software can be useful. A tier name is still not permission to test someone else’s systems.

Anthropic’s program details

Google’s EmbeddingGemma 2 maps text, code, images, audio and video into a shared searchable representation. The full model has 740 million parameters; developers can use smaller components for text-only work. Google released it under Apache 2.0 and describes it as designed for on-device use.

Think “find that clip from a description,” rather than “write me a novel.” Embeddings help retrieve material; they are not a chatbot answer by themselves.

Why it matters: local search can reduce cloud dependence. Actual privacy still depends on the app using the model, and real speed depends on your hardware. These capabilities are Google’s announcement, not a TabBOO feature we have shipped.

Google’s launch

Meta and Sierra, with partners including Shopify, Stripe, Walmart and Instinct, are developing the Personal Agent Protocol. It aims to give agents a consistent way to work with businesses while customers choose access and businesses define what actions they allow.

The design uses OAuth sessions and distinguishes read-only from write access. Sierra says a v0.1 specification is planned for later this month. This is an announced standard in development, not proof that every retailer already supports it.

Why it matters: “check my order” and “change my order” should be different permissions. More capable assistants need smaller, clearer keys.

Sierra’s protocol announcement

Sesame’s getting-started guide says its assistants have their own computer for browsing and can work with connected apps such as Gmail, Calendar and Drive after you authorize access. You can talk, text or send voice notes, then receive links and files when tasks finish.

Each assistant has its own app access. Sharing with one does not automatically share with another, and connections can be removed.

Why it matters: a voice interface becomes more useful when it can finish an errand. Giving it an inbox is still a real permission choice, not something to click through while making coffee.

Sesame’s guide

A University of Nottingham team studied male fruit flies under an almost eliminated geomagnetic field. The effects differed between healthy flies and flies with a Pink1 mutation used as a model of neurodegeneration.

The university reports that shielding Pink1 flies increased lifespan by around 20%, while their physical performance declined. The study also measured changes in mitochondrial function. It was published in Aging on September 23.

Why it matters: Earth’s magnetic field may be a variable in basic biology that researchers should examine more closely. The experiment does not establish a treatment for people. Please do not turn your bedroom into a magnetic bunker because a fly had an interesting week.

Published study · University research release