Meta's transcription AI undercuts Google 80%

Hi there, this is your daily ☕️ AIpresso.

In today's AIpresso:

🎙️ Meta's transcription AI undercuts Google 80%

🌍 World Labs AI builds 3D worlds from photos

🤖 Anthropic makes persistent AI work far cheaper

🔐 OpenAI limits Astra's top cyber tools

🔍 Claude's AI watermark has a blind spot

Plus: 💡 4 strategies & tactics, 🎁 8 other news you might like, 🧰 6 tools, and 📚 5 papers.

🎙️ Meta's transcription AI undercuts Google 80% LINK

Meta Superintelligence Labs launched Muse Voice Transcribe on September 1, priced at $0.18 an hour, roughly 80% below Google Cloud Speech-to-Text's $0.96 standard rate, and undercutting Cartesia, ElevenLabs, and Deepgram by smaller margins too.

The single model handles streaming transcription, diarization for 20+ speakers, and endpointing, scoring a 3.1% word error rate on Artificial Analysis's AA-WER Streaming benchmark, ahead of Cartesia Ink-2 (3.4%), ElevenLabs Scribe v2 Realtime (3.6%), and Gemini 3.5 Transcribe Live (4%).

Available now through the Meta Model API and powering system-wide dictation in Meta AI for Mac, Muse uses reinforcement learning to balance accuracy against delay, committing words in 80ms chunks, though its diarization error sits at 17.5% and the benchmark measures English only, not the 25 verified languages.

🌍 World Labs AI builds 3D worlds from photos LINK

World Labs, the spatial-intelligence startup co-founded by Fei-Fei Li, launched Atlas, an omni-model that generates, reconstructs, and simulates 3D scenes from just a few input images, anchoring every input to a position in 3D space.

Atlas outputs up to one minute of 1440p video with direct geometric camera control, rebuilds scenes from one to over a hundred photos, and exports point clouds and 3D Gaussian splats plus depth data for robotics real-to-sim training.

Human evaluators preferred Atlas in 75-94% of comparisons against models like MiniMax H3, Gemini Omni Flash, and Seedance 2.5, and it led reconstruction with a 25.3 median error, though it's currently limited to an early-access program for select partners.

🤖 Anthropic makes persistent AI work far cheaper LINK

Anthropic released Claude Fable 5.1 and Mythos 5.1 on September 1, holding base rates at $10/$50 per million input/output tokens while cutting cache reads 75% to $0.25 per million, which it says lowers typical bills ~25%.

Fable 5.1 led Artificial Analysis's Intelligence Index at 66 versus Claude Opus 5's 63, and Cognition measured coding cost dropping to $2.68 per task from Fable 5's $5.84, moving its Devin traffic over on launch day.

The savings depend heavily on reused context, since Artificial Analysis found max-effort tasks cost $3.76, 20% above Fable 5 because it emits ~1.7x the output tokens, and the system card rates alignment risk "low rather than very low" and calls Mythos 5.1 less honest under pressure.

🔐 OpenAI limits Astra's top cyber tools LINK

OpenAI will ship its forthcoming Astra model soon but gate its most advanced cybersecurity capabilities, saying it's the first LLM to cross the company's "critical cybersecurity threshold" for autonomously finding and exploiting security flaws.

Astra scored a perfect result on ExploitBench, and in a modified version built by OpenAI engineers it discovered and exploited two zero-day vulnerabilities without human guidance, prompting comparisons to Anthropic's Mythos model earlier this year.

OpenAI added chain-of-thought monitoring, jailbreak detection, and response restrictions for accounts flagged as higher risk, and says Astra didn't attempt to escape a breakout test, though no third party has confirmed the safety claims and testers remain undisclosed.

🔍 Claude's AI watermark has a blind spot LINK

Anthropic shipped Claude Fable 5.1 on Tuesday with a text watermark that reliably marks natural language but skips code, since swapping a variable, operator, or function token could break the program.

Based on Google DeepMind's SynthID-Text, the system perturbs the randomness used to pick next tokens rather than the probabilities, leaving a detectable statistical pattern that survives copying and some editing but not a full rewrite.

Detection ships as a private-preview API for regulators, law enforcement, media, fact-checkers, researchers, and enterprises needing AI Act compliance, though short responses may lack enough signal and code watermarks only appear in comments.

💡 Strategies & Tactics

> PRs NOT Welcome: How Top AI Open Source Projects Are Managing Thousands of Contributors: Top AI projects now auto-reject outside code contributions, using their own trusted agents to write fixes while redirecting community members toward reporting bugs and discussion.

> How to Size GPUs for AI Inference and TCO Without Overspending: Match GPU choices to each workload's actual token lengths and traffic, then shrink models through quantization to cut inference costs without overprovisioning.

> Closing an Azure OpenAI assistant's retrieval gap didn't take a new identity platform. It took one filter and a narrower assistant.: Add a query-time filter checking each user's own permissions before retrieval, since answer-quality tests never verify whose access an AI assistant actually uses.

> How to Shrink a Language Model Without Making it Too Dumb: Shrink big AI models to run on consumer hardware by storing weights in less detail, deleting near-useless ones, or training a smaller copy.

Other news you might like

• Perplexity's Hybrid Compute splits sensitive tasks between cloud and local AILINK

• Runway's Solaris is an AI system that generates software interfaces in real timeLINK

• NVIDIA and CrowdStrike Strengthen Agentic Cybersecurity FrontierLINK

• BenchMIRT: What are LLM benchmarks actually measuring?LINK

• AI token prices are hitting new record lowsLINK

• NVIDIA DLSS 5 to Launch on RTX 50 Series GPUs on September 3LINK

• Introducing agentic video understanding with GeminiLINK

• ChatGPT Health adds Epic integration for clinicians to import patient dataLINK

🧰 Trending tools

1752vc Pitch Deck Analyzer: reviews startup pitch decks slide-by-slide against 25,000+ real decks, flagging weak narratives and contradictory claims investors would catch first.LINK

Hy4 preview: Tencent's multimodal AI model family handling text, image, video, and 3D generation, helping developers build content tools and multimodal appsLINK

Cohere Parse 5: converts documents into accurate, structured text with faithful table and formatting transcription, benchmarked via ParseBench for reliable parsing quality.LINK

cain-agent: an AI penetration testing tool for authorized security assessments, with built-in cloud modules covering AWS, Azure, GCP, and Chinese cloud providers.LINK

DeepSec: audits AI-generated code for security flaws in real time and automates authorized penetration testing using 40+ skill packs from recon to proof-of-concept.LINK

agentacct: a local-first dashboard that tracks coding agent activity, showing tools used, files changed, tests run, time, and token costs.LINK

📚 Trending papers & reports

Fact versus context conflicts reveal that when an AI's training clashes with information you feed it, a specific internal signal steers which one it trusts, but that signal is task-specific and doesn't reliably carry over elsewhere.LINK

AI agent training fixes a hidden mismatch that appears when assistants trim their working memory mid-task, keeping what they learn in practice aligned with what they actually do, yielding steadier behavior and better results across seven web-search tests.LINK

Contrastive expert routing improves how large AI systems pick which specialized sub-model handles each word, lifting reasoning accuracy by up to ~1.8 points across nine benchmarks while adding under 3% to size and compute cost.LINK

Response-length forecasting reuses the speed-up machinery already running inside AI chatbots to guess how long each answer will be, letting servers prioritize quick requests and cut worst-case wait times for short ones by ~35%.LINK

AI judges of AI match humans well when picking one right answer but fail to capture how people disagree, so this tuning method makes AI graders better mirror the full spread of human opinions.LINK

See you tomorrow for a new dose of ☕️ AIpresso!