Morning edition — Thursday 5 November 2026

Cognition, the maker of the Devin coding agent, released SWE-2 Flash on Wednesday in Devin Desktop and the Devin CLI. It says the model runs at one-fifth the price and three times the speed of its larger SWE-2, and its launch page lists it as free through 30 November.

SWE-2 Flash is built on Thinking Machines' open-weight Inkling-Small, which has 276 billion total and 12 billion active parameters. Cognition added a mid-training stage before reinforcement learning and says this raised Terminal-Bench 4.0 from 0.9% to 20.0% and FrontierCode 2.0 from 3.2% to 24.0% over the base model, and that SWE-2 Flash beats all American open-weight models on leading evaluations.

Cognition also trained the model for security work, lifting its ExploitBench score from 11.3% to 34.8% by its account. Citing cyber risk, it is not publishing the weights and is building a programme to share them with trusted industry partners and research labs, with details promised in a few weeks.

OpenAI released Images 2.6 in ChatGPT, Codex and the API on Wednesday. It says the update generates images 50% faster and adds animated GIFs, and that people now create 4 billion images a week across its products.

In the API the release ships as two models, according to OpenAI's changelog: GPT-Image-2.6 Flare for fast iteration and GPT-Image-2.6 Sunburst for detail and consistency across edits. Developers also get markup-guided edits, SVG output and a complexity setting that controls how much planning the model does.

OpenAI's system card says GIFs bring risks that single images do not, such as harmful meaning conveyed across frames, and that motion-related violations remain challenging to detect.

Meta has acquired Conductor, a start-up whose Mac app lets developers run several AI coding agents at once in separate copies of a project, Axios reported on Wednesday. Meta confirmed the deal to TechCrunch but gave no further details, and terms were not disclosed.

Conductor said it has stopped new sign-ups and downloads. Conductor Cloud goes read-only on 13 November and shuts down on 20 November, when the company begins deleting cloud user data; active subscribers get refunds. The local app will keep receiving new models and critical fixes until early 2027, and co-founder Charlie Holtz promised something new to use early next year.

Axios set the deal beside OpenAI's planned purchase of Pierre Computer, announced the same day and covered in the previous edition. TechCrunch notes that Conductor runs agents from labs such as Anthropic and OpenAI rather than its own.

OpenAI published research on Wednesday on an automated pipeline that, in a mid-October run with an upcoming cyber model, analysed 2,308 candidate Linux kernel findings, reproduced 1,724 vulnerabilities and produced 948 local privilege escalation exploits. Such flaws let an unprivileged local user gain root. The run cost $1,213,430 in API credits, or $1,279 per exploit, OpenAI said.

The pipeline has five stages: discovery, validation in isolated virtual machines, exploitability testing, patch generation and human review before anything reaches maintainers. OpenAI said the targets ran a stock desktop operating system with standard protections, that it began with critical bugs on the Linux community's guidance, and that it is sending further reports and patches with its Patch the Planet partners.

OpenAI said the October model found more than five times as many novel exploits as a GPT-5.6 Sol run in July, and that its agents used at least 10 exploitation techniques new to outside kernel experts; it withheld their details. In an early run, 20 of 200 reported successes reached root through a different zero-day than the one assigned, which OpenAI cites as a reason to verify how agents reach a result.

OpenAI President Greg Brockman said on Wednesday that the company now generates the majority of its revenue from enterprise sales, Reuters reported from a Wall Street Journal technology conference in Napa, California. Asked on stage whether OpenAI would be ready for an IPO next year, he said: "I think so."

According to Reuters, Brockman said OpenAI has been ramping up sales to businesses, including its Codex agent, in competition with Anthropic, which Reuters said is likely to go public before the end of the year.

The US government will lay out all the laws and regulations that apply to domestic AI labs later this month, Pentagon Chief Technology Officer Emil Michael said on Wednesday, Reuters reported. Michael, vice chair of the government's Super Intelligence Task Force, said the group would publicise the regulatory landscape "in the next week or two".

Speaking at the Wall Street Journal conference in Napa, he gave no further details. "What we're trying to do is create a framework that makes the American people have confidence that these companies are going to do the right thing," he said. "Or else I think what you get is 535 legislators writing a piece of regulation that could kill the industry."

Anthropic signed a content and data licensing deal with the Associated Press just before the US midterm elections, Adweek reported, citing two people familiar with the matter. According to Adweek, Claude users who asked for updates on election night saw up-to-date information attributed to the AP.

Adweek says this is the first time Anthropic has licensed content from any digital publisher, while OpenAI has negotiated more than a dozen such deals. Financial terms and duration were not disclosed; Anthropic did not respond to Adweek's request for comment and the AP gave no comment on the record.

Apple lowered the daily AI spending cap for some engineers from $300 to $200 last month and plans further cuts that could take some engineering teams to $100 a day, Bloomberg reported on Wednesday, according to The Next Web and 9to5Mac. Many staff outside engineering already have a $100 cap, which Apple has considered cutting to as little as $25 for some teams; employees can ask senior leaders for exemptions.

The budgets cover tokens used with tools such as Anthropic's Claude, OpenAI's ChatGPT and Google's Gemini, and some teams are being told to switch to cheaper or on-device models. Apple declined to comment, The Next Web says, and notes that Apple's research and development costs rose 32% in its 2026 fiscal year, partly on infrastructure and AI costs.