AI — Page 3

The September 24th post promises access across Nous Portal tiers, while the Portal's pricing page describes hosted tools as a paid-tier benefit.

Aniket's September 24th post quoted $0.07 per million input tokens, while Isoquant's website now shows different rates and performance claims based on its own benchmark.

TestingCatalog reports that the proposed ChatGPT tier would add 'Fastest Work and Codex' to Pro, with higher usage limits also possible. OpenAI has not listed it on its pricing page.

Founder Ali Ansari says flow-transform 1.0 scored 96% F1 on PrivacyBench; micro1's new end-to-end benchmark is synthetic and limited to tabular data.

Anthropic is locking in seven years of CPU capacity; Akamai says the deal could expand by another $9B and estimates about $5.5B in related capital spending.

Sundar Pichai says a Planet-built prototype will fly next week, beginning an orbital test of Google's plan to put AI computing in space.

One user described delayed email alerts. The Information reported that Instinct had sometimes warned users it was at full capacity and responses might be slower; that warning does not establish how many users experienced delays.

Ramin Hasani's team says the experimental add-on cuts decoding time by up to 3.13x on an M5 Max, while leaving image processing and prompt prefill unchanged.

JetBrains CEO Kirill Skrygan is pushing the IDE maker into team workflows, cost controls and governance without demanding loyalty to Junie.

Alexandr Wang demonstrated the feature on September 24th, a day after Meta outlined the real-time avatar technology behind it.

Alexandr Wang's team pairs Muse Realtime Voice with a video model Meta says runs at about 870 milliseconds and watermarks generated output.

The company pitches a shared workspace for people and AI agents; its product site lists Google Drive and Notion as the sources currently available.

Alexandr Wang said the device is shipping in December, in time for the holidays. Demo screens show gesture timings and an avatar interface that Meta has not confirmed as shipping specifications.

The deal was reported in August 2025, after WaveForms raised $40 million from a16z; Alexis Conneau says some of the team's work is set for Meta Connect.

At Connect, Mark Zuckerberg also described new Mac control for the agent, extending a product that already works across files and native apps.

Alexandr Wang said users can wake Muse by saying its name. Meta has not given a release date or explained how glasses will approve consequential actions.

Inferact reports 709 tokens per second on 16 TPU v7 chips, versus 452 on 16 GB200 GPUs, in a low-concurrency speculative-decoding test. Co-founder Woosuk Kwon, a vLLM co-creator, helped build the open-source kernel.

Pavel Rabtsevich's linked test classified 8,054 historical signals, but its stronger result came after the first run informed a redesigned prompt.

The company’s prompt arrived after OS3 opened beyond invite-only beta, pitching one assistant across models, computers and its r1 device.

Across 360 real-robot trials, Opus 5.5 averaged 36% progress at an estimated 90 cents per run; Astra completed more tasks outright.

Co-founder Inky Ganbold is pitching a browser workspace where engineers can fork open boards, edit them with AI and export files for fabrication.

The September 23rd update pairs voice with GPT-6 models and workplace tools, extending ChatGPT Work to web and mobile.

Karina Nguyen and Jakub Zegzulka built the research demo around synthetic versions of Elon Musk and Joe Rogan, with generated dialogue labeled on screen.

Anthropic says a two-week sprint improved 13 performance measurements across claude.ai and the Claude desktop app, with engineers setting targets and steering the work.