A Sept 22 Traictory write-up flags a 12-point gap on Terminal-Bench 4.0 between xAI’s 38% (xHigh, harness unnamed) and Artificial Analysis’s 26% on mini-swe-agent. Price stays $2/$6 in Cursor; AA also calls the default xhigh config slow and verbose versus the twice-as-fast paid tier.
AWS released Strands Harness (Apache 2.0) — a ready made general purpose agent for Python/TypeScript that runs locally or in any Linux cloud, with shell/file/web tools, context compaction, memory, and subagents. On matching Claude/GPT models it claims 28% lower token cost vs other harnesses across six benchmarks, and with Fable 5 it scored higher than Claude Code on Terminal Bench 2.1 at 77% lower cost…
DeepSeek will be among firms briefing the UN Security Council on Wednesday on AI and international security, sources told Reuters, alongside expected participation from OpenAI and Anthropic. Moonshot was also invited; DeepSeek founder Liang Wenfeng does not plan to attend…
In September parliamentary submissions, OpenAI and Anthropic urged Australia to allow a narrow, conditional path for model training on local creative works after Canberra ruled out a broad copyright exemption. Both tied the ask to data centre plans in Sydney and Queensland; a joint committee report is due in November…
JetBrains today folded Central and related agent tooling into JetBrains Air—an open, multi vendor system spanning IDE agent workflows, Air Teams coordination, and Air Governance. Junie remains the coding agent across surfaces; Agent Client Protocol lets third party agents plug into JetBrains IDEs…
White hat firm Hacktron chained a Discourse/libheif image bug with an OpenAI SSO flaw, then used an employee’s Codex session to open a harmless PR in an internal repo—under 72 hours end to end. Claude Opus 5 built the working exploit overnight after Opus 4.8 failed; OpenAI patched in 14 hours and paid a $6,500 bounty…
Spotted something we missed? Start a thread.
