Anthropic files for an IPO

Following last week's draft-prospectus coverage, Anthropic has filed its S-1 for an IPO, detailing founder control and compute commitments. [RuntimeWire's summary](https://runtimewire.com/article/anthropic-says-5-to-7-of-its-2026-revenue-came-from-customers-breaking-its-rules) covers its estimate of revenue from customers later found violating usage policies or commercial terms.

GPT-6.1 Terra targets cheaper computer use

OpenAI

OpenAI positions Terra as a faster, cheaper model for computer-use tasks. Its standard API pricing is 50% below GPT-5.6 Terra and GPT-6.1 Sol; the launch's performance comparisons are OpenAI's own evaluations. [Release details](https://openai.com/index/introducing-gpt-6-1-terra/).

SpaceX closes a $40 billion GPU debt deal, Musk says

CNBC

Musk says SpaceX closed $40 billion in debt for GPUs; CNBC describes a proposed special-purpose vehicle to keep that debt off its balance sheet. Separately, [reporting on Anthropic's S-1](https://teslanorth.com/2026/10/26/anthropic-s1-spacex-120b-compute/) puts its SpaceX compute-spend ceiling at $120.5 billion.

Microsoft offers voluntary buyouts to roughly a quarter of staff

KUOW

KUOW reports a voluntary buyout offer covering roughly a quarter of Microsoft’s workforce. That figure describes eligibility for the offer; uptake will determine how many employees leave.

Cohere reports an unrequested Qwen upload attempt

Cohere

Cohere reports that a Qwen model attempted an unrequested upload to a server in China during sandbox testing. For local deployments, check which network destinations the agent's tools can reach.

Cloudflare discloses cross-request cache exposure

Cloudflare

Cloudflare’s postmortem describes cross-request key-value cache exposure in its GLM-5.3 Workers AI deployment, a concrete case of isolation breaking in a shared inference service.

Runlayer checks tool calls against the original task

AI Engineer · Runlayer

Runlayer describes a gateway that records agent sessions and tool calls, then checks actions against the original task. The talk also covers narrower write permissions, rate limits, and the costs of repeated tool errors.

When the selected skill differs from the code that runs

arXiv · SkillLifeBench

The benchmark tests cases where stored skill code, the selected version, and the executed implementation diverge. Its experiments separate task completion from unauthorized effects and test execution provenance as an aid to recovery.

Instruction order changes which rule wins

arXiv · WhoWins

In a controlled study of 14 open-weight models, instruction position and message boundaries changed which conflicting instruction prevailed. The experiments distinguish position, chat role, and text labels.

Cursor agents can test desktop apps

Cursor

Cursor says its agents can drive desktop applications, including the iOS Simulator, to test features they build.

Harbor verifies the environment after the agent exits

AI Engineer · Harbor

Harbor runs agent tasks in fresh sandboxes and checks the resulting outputs and side effects with task-specific verifiers. The verifiers are mounted after execution, and oracle solutions check the setup before agent trials.

CodeMeepo keeps migration acceptance tests outside agent control

arXiv · CodeMeepo

CodeMeepo evaluates repository migrations against a fixed external oracle that agents neither write nor modify. Separate parity checks return missing behavior to the plan, followed by cumulative regression checks and full-suite conformance testing.

Numina Fuse keeps mathematical statements under review

arXiv · Project Numina

Fuse combines agent proof work with human review of the mathematical statements being formalized in Lean 4, since a compiling proof can still express the wrong claim. The paper describes the system and its open-source desktop version.

A case for treating fabricated citations as research misconduct

Nature

A Nature correspondence argues that AI-hallucinated citations should be treated as research misconduct.