AI & Tech Brief — August 29, 2026

TL;DR

OpenAI is pulling its models out of Cursor now that Cursor belongs to SpaceX — it gave the maximum contractual notice and set a November 12 shutoff, saying it can’t trust Musk’s companies to honor its terms after admitted violations.
Z.ai shipped GLM-5.3 as open weights, and it’s the strongest open coding model yet — the gains come entirely from post-training on the same base as GLM-5.2, which is a telling signal about where progress is coming from.
htmx 4.0 is out after eight months — same API, but rebuilt on fetch() with explicit attribute inheritance, cleaner events, and first-class morphing and streaming.

Key Stories

OpenAI winds down its Cursor contract after the SpaceX acquisition OpenAI notified SpaceX that it intends to stop supplying models to Cursor, with a proposed shutoff of November 12, 2026 — the maximum notice its contract allows after a change of control. The stated reason is trust: after Musk acquired Twitter (now part of SpaceX) it broke OpenAI’s contract, and Musk admitted under oath that xAI (also now part of SpaceX) had violated OpenAI’s terms. OpenAI also framed it around accountability for its upcoming “Astra” model, which it won’t supply to Cursor. The HN thread (547 points) splits between “standard circling of the wagons ahead of a frontier-AI fight” and genuine sympathy for developers relying on OpenAI models in Cursor — Anthropic already banned xAI over similar ToS violations, and the open question is whether it now extends that to Cursor. Source: https://openai.com/index/our-decision-on-cursor-following-its-acquisition-by-spacex/
GLM-5.3 is released as open weights Z.ai’s latest is the most capable open-weights coding model, claiming a 50% jump over GLM-5.2 on its in-house code bench and open-source SOTA on Terminal Bench 3.0 and Agents’ Last Exam. The interesting part: it uses the same base model as GLM-5.2 — every gain comes from post-training, which the HN crowd read as evidence that better environments, verifiers, and training trajectories may now matter as much as another huge pretraining run. It also posted state-of-the-art cyber results, drawing both interest and unease. At 753B params it needs serious hardware to run locally, but the consensus is it’s a legitimate workhorse for on-prem and security work that US models tend to refuse. Source: https://huggingface.co/zai-org/GLM-5.3
htmx 4.0 ships Eight months in the making, htmx 4 keeps the 2.x API surface almost identical but rebuilds the internals on fetch() (dropping XMLHttpRequest). The three user-facing changes: attribute inheritance is now explicit (hx-confirm:inherited), event names are standardized to htmx:phase:action, and history no longer snapshots pages in localStorage (it re-fetches on back navigation, fixing a long-standing source of bugs with third-party JS). New features include built-in morph swaps, a <hx-partial> tag, and streaming extensions (hx-sse, hx-ws, hx-multipart). It ships with an npx htmx.org@4.0.0 upgrade-check tool and — tellingly — LLM skill files for migrating codebases. 2.x stays latest on npm until early 2027, so there’s no forced upgrade. Source: https://four.htmx.org/announcements/2026-08-28-htmx-4.0.0-is-released
OpenAI’s Python SDK migrates to HTTPX2 The openai package now uses HTTPX2 for its sync and async HTTP clients and no longer installs the legacy httpx package. The headline breaking change: HTTPX2 verifies TLS against the operating-system trust store instead of the bundled certifi CA bundle, which can break certificate verification in minimal containers and behind corporate TLS-inspecting proxies. Default usage is unchanged, but custom clients, auth handlers, transports, and RESPX-based mocks all need to target HTTPX2 objects. There’s a runtime-only legacy escape hatch (inject an httpx client with a cast), but it fails static type checking and “may be discontinued.” Source: https://github.com/openai/openai-python/blob/main/httpx2.md
“I accidentally turned LLM memory into program analysis” A security researcher, frustrated that LLM agents lose track of what’s been established during long vulnerability investigations, built Lemmalog — a Datalog engine that sits beside the LLM. The model turns messy debugger output and notes into structured facts; the engine derives conclusions, tracks provenance, and automatically invalidates anything that depended on a fact that later proves wrong. It tops the published field on LongMemEval’s “knowledge update” category while using ~38x less context per query than full-context prompting. Why it matters: it’s a concrete, benchmarked argument that much of “LLM memory” is really a state-maintenance problem databases solved decades ago — not a reason to buy a bigger context window. Source: https://pwning.systems/posts/llm-memory-program-analysis/
Autonomous mathematical discovery in a multi-agent “Station” An arXiv paper describes an open-world multi-agent environment where AI agents from different model families pursue a shared research goal with no central coordinator or scripted pipeline — choosing their own directions, running experiments, and building a shared literature. Across 12 AlphaEvolve construction problems plus two case studies, the agents produced results novel relative to prior literature on five: a new infinite family of finite-field Kakeya sets, new exact 604-point kissing configurations in dimension 11, new records for the discretized Kakeya needle and sign-uncertainty problems, and an improved lower bound for Erdős’s minimum-overlap problem. Notably, they produced theorems and analyses explaining the constructions, not just numbers. Why it matters: it’s a credible data point for multi-agent systems doing real, verifiable mathematical research rather than toy demos. Source: https://arxiv.org/abs/2608.23691

Quiet but interesting

Claude Code’s August 28 release adds PreModelSwitch/PostModelSwitch hooks (block, confirm, or annotate a model switch), a per-session prompt-cache line in /cost, and live streaming of a foreground subagent’s tool calls to Remote Control clients — plus a long list of security fixes around symlink handling, plugin path traversal, and sandboxed Bash output. Source: https://code.claude.com/docs/en/changelog
Codex CLI 0.151.0 lets extensions inspect or replace MCP tool results before they reach the model, adds a configurable grace period for discovering optional MCP servers, and counts nested subagent token usage toward root goal budgets. Source: https://developers.openai.com/codex/changelog
U.S. sanctions against the A/I Collective — the long-running Italian autonomous hosting collective behind inventati.org — drew a large HN thread (612 points). It’s more internet-governance than AI, but it matters to anyone running infrastructure on activist or privacy-focused hosting. Source: https://www.inventati.org/

Skip

Superhuman’s “Claude automates legal and small biz work” — the newsletter’s lead item today, but it’s an aggregator blurb pointing at a product/tutorial angle rather than a verifiable primary announcement; nothing concrete to act on. Source: https://www.superhuman.ai/

Sources checked: Claude Code changelog, Claude release notes, Gemini CLI changelogs, OpenAI API & Codex changelogs, Superhuman, ByteByteGo, darioamodei.com, blog.samaltman.com, Google DeepMind blog, Hacker News front page. Quiet today: Claude release notes (latest Aug 6), Gemini CLI (latest v0.54.0, Aug 6), OpenAI API changelog (latest Aug 21), darioamodei.com (no new essays), blog.samaltman.com (latest is the Aug 28 post covered yesterday), DeepMind blog (latest Aug 27 double-blind evals, covered yesterday), ByteByteGo (latest Aug 27, covered yesterday).