Terminal-Bench 4.0 shipped August 28, quietly superseding version 3.0 — a benchmark this blog spent four weeks declining to cite because its live leaderboard kept returning contradictory numbers. The primary source finally explains why: ‘continuous benchmarks’ don’t have a fixed scoreboard, they have a timestamp.
OpenAI announced August 28 that it will end Cursor’s direct access to its models on November 12, 2026, citing SpaceX and xAI’s history of violating its terms of service. Cursor CEO Michael Truell says OpenAI models are only ~5% of Cursor’s traffic — but the episode is a clean case study in what happens when your product is a reseller of other labs’ models.
Salesforce and Anthropic announced ‘Claudeforce’ on August 26 — led by a Claude CoWork plugin called Salesforce in Claude with 37 prebuilt sales skills that let Claude read and act on live CRM data through an MCP server, without opening Salesforce’s own UI. Pilot customers have it now; open beta is set for September.
On August 26, Z.ai confirmed that Ox Alpha — the anonymous free coding model that spent six days on OpenRouter and OpenCode with Claude Code among its top consumers — was GLM-5.3-Flash, a 320B-parameter MoE model, all along. Weights went to Hugging Face under an MIT license the same day, served entirely on domestic Chinese chips.
On August 25, OpenAI reinstated the rolling 5-hour usage cap on Codex and ChatGPT Work for Plus subscribers, six weeks after removing it on July 12 as an open-ended ’temporary’ change. In between: a silently discovered paid-reset test and a free reset handed out for hitting 20 million users. Pro subscribers stay uncapped for now — Plus users are back where they started.
A frontier-class coding model called Ox Alpha appeared for free on OpenRouter and OpenCode on August 20, 2026, with no disclosed maker. Fingerprinting points toward Z.ai’s GLM family — a US Commerce Department Entity-listed company — and Bloomberg reports Claude Code is already among the top consumers, routing billions of tokens to a backend nobody has confirmed.
GitHub shipped a public preview of GitHub Copilot in Slack on August 21, 2026 — @mention it to triage issues, investigate failures, and run sandboxed agent sessions. Anthropic’s Claude Tag has done on-demand and proactive, ambient Slack monitoring since June 23. The gap between the two reveals exactly where Copilot’s architecture still tops out.
On September 1, GitHub Copilot drops Gemini 3.1 Pro, Claude Opus 4.5, Opus 4.6, Sonnet 4.5, Sonnet 4.6, and Raptor Mini from every Copilot surface. Individual annual-plan subscribers get a one-model carve-out; everyone else has one week left to update their configs.
Around August 16, OpenAI was caught testing paid usage resets for Codex — $5-8 for Plus, up to $80 for Pro — discovered in app checkout code rather than announced. Five days later, on hitting 20M Codex users, OpenAI credited every paid user a free ‘banked’ reset instead — one that rolled out roughly a day late and, per user reports, drained within hours for some.
Anthropic’s second company-wide Risk Report, published August 14, raised its catastrophic-misalignment rating from ‘very low’ to ’low’ and disclosed an unreleased internal model, ‘Model 2,’ that beats Mythos 5 on many internal tasks but hasn’t cleared Anthropic’s full predeployment safety suite — and has no release date.