Tuesday, 27 October 2026

OpenAI released GPT-6.1 Terra on 26 October, a fast, efficient model it pitches at visual reasoning, computer use and clearer writing [1]. In the API it is `gpt-6.1-terra`, at $1 per million input tokens, $0.05 per million cached input tokens and $5 per million output tokens [1][2]. GPT-5.6 Terra, the model it follows, lists at $2 per million input tokens and $12 per million output tokens [5]. The model page lists a 1,050,000-token context window and 128,000 maximum output tokens, and prompts over 272K input tokens are billed at twice the input rate [2].

In ChatGPT, the release notes say Terra is rolling out to Plus, Pro, Business, Enterprise and Edu plans in the Work and Codex experiences, picked from the model list, and that each plan's overall usage limits are unchanged [3]. The company says the same allowance covers more tasks with Terra than with GPT-5.6 Terra, and substantially more than with GPT-6.1 Sol or GPT-6 Astra [1]. Some Plus and Pro users may see Terra at Medium in place of GPT-6.1 Sol at Light in the default Power presets [3].

The performance figures are OpenAI's own. The company says that at maximum reasoning effort Terra solves 50.3% of Terminal-Bench 4.0 tasks, against 21.5% for GPT-5.6 Terra, and that its error rate on an internal set of error-inducing prompts fell from 53.6% to 16.7% at low effort [1]. It also says GPT-6.1 Sol and GPT-6 Astra remain the better choice for the most complex coding [1]. Simon Willison notes that Terra sits where Sonnet sits in Anthropic's line-up, neither cheapest nor most expensive, which makes it harder to decide when to use it [4].

Cloudflare says that between 27 August and 9 September a small number of customers using the GLM-5.3 model on Workers AI received responses that contained data from other customers' requests to the same model [1]. In total 155 responses were affected, all to prompts longer than 681,280 tokens [1]. Security researcher Gadi Evron of Knostic reported the issue on 8 September through Cloudflare's HackerOne programme, and Cloudflare took the model offline and restored it on 9 September [1]. The company says it found no evidence of deliberate exploitation and that no other model on Workers AI was affected [1]. It published the report on 26 October, together with a post from the SGLang maintainers [1][2].

The cause was two problems stacked together. Cloudflare deployed GLM-5.3 with a 1,048,576-token context window, but the KV cache pool on its prefill GPUs held only 681,280 tokens, a mismatch it says it knew about before launch [1]. In SGLang, the open-source inference engine it uses, an over-length request in disaggregated prefill/decode mode was marked as rejected and swapped for a one-token placeholder, but not aborted, so the decode node answered while reading cache pages left over from earlier users' requests [1][2]. The client got HTTP 200 instead of an error [1].

The SGLang maintainers say default disaggregated configurations were exposed to a related cache race from the v0.4.x line until v0.5.21, or v0.5.22 for the MORI backend [2]. Cloudflare recommends that every operator move to SGLang v0.5.22 or later, and says it now checks pool size against context length before any disaggregated deployment change ships [1]. If you send long documents to a hosted open model, the provider's isolation between customers is part of your own data exposure.

Google DeepMind announced Genie 4 on 27 October, a world model that generates an explorable scene in real time [1]. Users move with standard keyboard controls and can type prompts while the simulation runs, for example to change the weather, move the camera or add a person or object; Gemini adapts the prompt to the scene and, between inputs, monitors and steers it so the world stays coherent [1]. DeepMind says it is its first real-time world model to generate audio natively, including ambient sound and dialogue in which characters react to the user [1]. TechCrunch reports that Genie 4 keeps a consistent scene for up to two minutes, double its predecessor's limit [2].

Access is narrow. DeepMind is rolling it out as a research preview for trusted testers, and says it will reach Google AI Ultra subscribers in Project Genie and early access partners of the Gemini API and Google AI Studio next [1]. TechCrunch reports that this is planned for the coming weeks [2]. A price for Gemini API access is not public. There is a waitlist in Project Genie [1].

The team told TechCrunch that details and objects can drift the longer a simulation runs, and that it has not yet worked out how to make audio accurate or in-world text legible [2]. The pitched uses are gaming, interactive learning and robotics, and trusted testers point to prototyping films as a possible new use [2]. In one research experiment tied to Gemini Robotics, a Gemini model broke tasks into steps that Genie 4 simulated in real time, a possible way to test robot scenarios before running them for real [2].

OpenAI published three results it says its internal models produced: a proof of the local smoothing conjecture in every dimension, a refutation of the Exponential Time Hypothesis (ETH), and an existence proof for certain four-dimensional quantum field theories, including QCD [1]. ETH is a long-standing conjecture about how hard it is to solve Boolean satisfiability, and it has been used in hundreds of papers to derive conditional hardness results [1]. OpenAI says its AI found the disproof on 16 October and that it was formalised in Lean on 17 October; both the write-up and the Lean proof are in its public GitHub repository [1].

OpenAI says it does not seek academic credit and has handed each result to outside experts [1]. The ETH work will be checked by a group led by Russell Impagliazzo and Ramamohan Paturi of UC San Diego and Ryan Williams of MIT; the group asked for all raw data to be public and for OpenAI to fund a conference on the implications, and OpenAI agreed [1]. The physics group, led by David Gross, will hold back the AI-produced material and write it up itself [1].

Williams told TechCrunch that some parts of the proof are routine but "two or three alien-looking parts" made him think the whole community needs to work on it together [2]. He said he asked whether the result breaks any cryptosystems, and that OpenAI had checked and was reasonably confident it did not [2]. According to TechCrunch, Anthropic released its own refutation of ETH hours after OpenAI's [2]. None of this is a product. For now it is a claim with a machine-checked proof and no expert write-up yet.

Oracle and OpenAI announced an expanded partnership at Oracle AI World on 27 October [2]. The centre of it is Oracle AI Workspace, which Oracle says uses plugins in ChatGPT to connect Oracle Fusion Applications, Oracle NetSuite, Oracle AI Database, Oracle Cloud Infrastructure and other tools, so employees can look up information and complete approved workflows within their organisation's permissions [1][2]. OpenAI describes the workspace as something Oracle is developing, and gives the example of a new finance team member using ChatGPT to investigate costs before month-end close [1].

For small businesses, Oracle announced the Oracle AI SMB Bundle, a free-to-start package that includes $12,000 in Oracle cloud credits each year, ChatGPT Business, Oracle AI Workspace, Codex, Oracle App Composer, OCI and NetSuite QuickStart for invoicing and cash flow [2]. Oracle adds that terms, conditions and restrictions apply; the full terms and what happens after the start are not public. Oracle customers in the U.S. can also buy the OpenAI API, Codex and ChatGPT Work through Oracle Marketplace using eligible Oracle Universal Credits [1].

The two companies also announced Oracle AI Security Mesh, a managed service that uses OpenAI cyber models to find and fix vulnerabilities, with customers setting where testing is allowed [1][2]. OpenAI says Oracle itself reported 130,000 active ChatGPT users and more than 95,000 active Codex users [1]. No release dates are given, and Oracle's release says the announcement outlines product direction and is not a commitment to deliver [2].

Apple says Keynote, Pages and Numbers will get an assistant, built on Apple models and OpenAI, that understands the context of the document you are working in [1]. In Keynote it helps design, animate and refine slides; in Pages it drafts, restructures and fine-tunes text; in Numbers it calculates, visualises and organises data, suggests chart types and builds custom tables [1]. The assistant is in beta, needs an Apple Intelligence-capable device on iOS 27, iPadOS 27 or macOS 27 or later, and works in U.S. English only [1].

The bigger change for anyone already using agents is Model Context Protocol support. Apple says MCP lets external AI agents and tools read content, make edits and run automated actions directly in the three apps, and it requires macOS 27 or later [1]. The Verge notes that this lets third-party agents plug into the apps [2]. Keynote also adds the Keynote Interchange Format (.keyx), an open JSON-based file format Apple says is readable by external tools and AI agents [1].

The features are part of an Apple Creator Studio update arriving on the App Store in November, free for existing subscribers [1][2]. New subscribers pay $12.99 a month or $129 a year in the U.S., with a one-month free trial [1]. Apple says the free versions of Keynote, Pages and Numbers remain available, and its release ties the new intelligence features to subscribers [1]. The same update adds an assistant in Logic Pro and a Generate Fill tool in Pixelmator Pro [1][2].

Mozilla has added two features to Smart Window, Firefox's opt-in AI browsing mode [1]. The first fills in web forms from tabs you choose. When a form has at least four text fields it can help with, a "Fill from sources" option appears with a recommended source page, and you can change or add tabs before reviewing the suggestions [1]. Mozilla's example is pulling work history from your portfolio or professional profile into a job application [1]. Smart Window does not submit the form for you, and Firefox's regular autofill still handles saved details such as your name and address without an AI model [1].

The data handling is the part to read. When you use form filling, information from the selected source pages is shared with Google's Gemini 3.1 Flash Lite, which detects the fields and extracts what to enter [1][2]. The feature supports text fields only for now; Mozilla says it is exploring dropdowns and PDF forms [1][2].

The second feature watches pages. You type "/watch" in the Smart Window chat, describe what you are looking for, such as a sale or a job listing, and add up to five pages per task; Firefox checks daily or weekly, but only while the browser is open [1]. You can have up to 15 active tasks, which pause after 60 days without a match and stop after 90 days [1]. Page monitoring uses a Qwen model [1]. Both features are in the Smart Window beta on desktop, in English in the U.S. and Canada, with more markets in the coming months [1][2].