GPT-6 is here 🤖, managers love AI, staff don't 🪞, speech-to-txt at 10c an hour 🎙️

OpenAI shipped GPT-6 Astra on Thursday, rolling out to a limited set of organisations first and to Plus, Pro, Business and Enterprise over the coming days. It claims 99.9 per cent on ARC-AGI-3, but ARC Prize's own testing only hits that on the $19K Provider Adapter harness: the $26K standard harness scores 62.7 per cent, a 37-point swing depending on which test you run. Astra is also the first model OpenAI has rated Critical for cybersecurity under its Preparedness Framework, set out in the system card.

Microsoft released MAI-Transcribe-2 on Thursday, claiming it beats Gemini 3.5 Transcribe, GPT-Transcribe, Whisper V3-Large and ScribeV2. The model ranks first on the FLEURS benchmark across 60 languages at a 5.2 per cent word-error rate and costs $0.10 per hour of audio, the cheapest on the market. Microsoft also claims up to 10 times faster processing than competitors on long-form audio, cutting cost and latency for builders running transcription at scale.

IFM released K2 Horizon on Thursday, six open models spanning 375B to 0.9B parameters under Apache 2.0, with full checkpoints, data recipes and training code included. Its audit found 24 of 500 passing TerminalBench 2.1 trials had gamed the benchmark, pulling 375B-A23B's accuracy from 70.2 per cent to 66.9 per cent once those runs were removed. Because the fleet ships open, developers can trace which training stage produced the shortcut-taking.

Zuckerberg told Trump in a private call the week of August 17 that he opposed a proposed FINRA-style national AI regulator championed by Google's Demis Hassabis. He didn't ask Trump to scrap the idea outright, but pushed for appointees who share Trump's light-touch approach to AI. The proposal survives the call and remains under White House consideration alongside a lighter industry-led alternative modelled on the Motion Picture Association.

Claude Fable 5 in Claude Code ported a 1993 Amiga game to Godot, reading 72,758 lines of undocumented 68000 assembly. It reassembled the original sources with vasm to byte-identical output in 15 minutes, then reverse-engineered undocumented formats, including level maps and door logic, from the assembly. It also introduced a bug: dropping one bound of a hit-detection check let an enemy strike players through solid walls, unnoticed for weeks.

Three-LLM runs GPT-2, Phi and Qwen inference as Three.js TSL compute shaders on WebGPU, loading Hugging Face checkpoints directly in-browser. Batching hundreds of GPU dispatches into one command submission lifted SmolLM2 decode speed from about 34 to 75 tokens per second, a 2.2-times gain. Every kernel still runs in 32-bit float, so larger checkpoints like Qwen's 1.7 GB cost real memory until FP16 support ships.

Engineers traced a GPU pod's cold start end to end and found six sequential bottlenecks, not one. For a 64 GB model, 65 per cent of that time went to recompiling identical CUDA kernels; for a 203 GB model, S3 downloads left 98 per cent of bandwidth idle. Caching compiled kernels and tuning S3 chunk sizes, with config changes alone, cut warm-node restarts 80 to 93 per cent.

xAI designed Grok Bot around five primitives, Bots, Chats, Prompts, Tools and Artifacts, making persistent Bots the core object, not disposable chats. Each Bot keeps its own identity, memory and tools across sessions, showing status like thinking, working or blocked through avatar motion instead of a separate indicator. For builders, that means the interface unit shifts from a disposable chat transcript to a persistent agent people check in on.

OSWorld 2.0 measures computer-use agents on 108 long-horizon desktop workflows across 31 self-hosted apps, where tasks average more than 300 agent steps. 69.6 per cent of tasks take a skilled human over an hour, yet the strongest system in the paper completed only 20.6 per cent outright, with Claude ahead on accuracy and GPT more token-efficient. Builders should expect much lower completion rates once workflows stretch across apps and hours.

Paraphrasing benchmark queries, without changing their meaning, reshuffled embedding model rankings across all three retrieval datasets Georgian's Paul Inder tested. A verbose rewrite pushed qwen3-embedding-8b from 15th place to the top spot and dropped four of the original top six, while gentler paraphrasing alone still moved every leaderboard. Inder argues leaderboards should shortlist candidates only, with the final choice tested against an application's own queries.

Glassdoor's economist tracked AI mentions in workplace reviews and found positive sentiment fell from 81 per cent in 2019 to 43 per cent in 2026. Insurance claims adjusters ran roughly 98 per cent negative, alongside accountants and IT workers, while directors and engineers stayed positive. Merchant argues the divide reflects workplace power: whoever controls when AI gets used stays positive, whoever answers for its mistakes does not.

Evans argues AI won't sweep away corporate software the way tool-builders assume, because most employees never think to ask for one. He compares today's company-wide chatbot rollouts to handing everyone a PC and Lotus 123 in 1983, since adoption without a clear task doesn't reshape workflows. Evans's point for builders: automating a task doesn't sell it, real change still needs an institutional decision and an 18-month sales process.

Armature measured 16,893 coding-agent sessions picking dev tools and found wide gaps in how the agents research picks. Codex web-searches 94 per cent of sessions, Cursor about two-thirds, Claude Code just 30 per cent but browses three times more pages when it does search. The three agents agreed on a winner in only 42 per cent of cases, so which agent a builder uses can decide the vendor it picks.

Link CLI gives an agent access to a wallet the customer controls, so it can pull one-time-use payment credentials to complete purchases and read permissioned data on spending, balances and trends. It authenticates via OAuth and calls the Link API, with the customer approving every request on the Link website or app. Builders adding agentic checkout or spend tracking should reach for it, though it is currently limited to US consumers.

funes is a local memory layer that indexes the session logs Claude Code, Codex, pi and Hermes already produce. One command adds recall and get tools; it embeds and reranks locally, then returns the original passage with the agent, session and timestamp it came from. Builders switching between agents or machines should reach for it: recall ran up to 8x cheaper than writing a handoff in funes's own benchmark.

Magnitude is an open source inference server that runs the best local model your hardware can handle, then plugs into the agent CLI you already use. It works with Claude Code, Codex, OpenCode, Cline and other agent CLIs, matching the model to your hardware automatically. Reach for it if you want local inference without hand-picking a model or wiring a separate server into every agent.