AI NewsOctober 6, 2026

Bank of America strategists say the market has priced AI spending almost perfectly but may be underestimating how much US consumers will keep spending, suggesting a shift toward consumer-linked stocks.
In a Monday note led by strategist Savita Subramanian, Bank of America said AI-related capital expenditure and shrinking discretionary spending are "already baked into investment positions," with AI-exposed sectors like software and consumer finance trading near record lows relative to industrial stocks. The strategists called it "dangerous to underestimate" US consumer demand.
This follows a Goldman Sachs forecast that the five largest US tech companies will spend $1.2 trillion on AI infrastructure in 2027, up from $800 billion this year, while PYMNTS and Visa data show consumers adjusting spending rather than cutting back.
Key Capabilities:

• The call: Bank of America recommends selectively shifting from AI capex plays toward consumer stocks.
• The gap: AI-exposed sectors trade near record lows versus industrials, a valuation gap BofA calls overdone.
• Context: Goldman projects hyperscaler AI spending hits $1.2 trillion in 2027.

North 2 adds over 15 features, letting a company's AI agents remember past sessions, share reusable skills, and stay within strict per-agent spending limits, aimed at regulated industries like banking and healthcare.
Cohere launched North 2 on October 5, rebuilding its agent orchestration system and adding cross-session memory so an agent doesn't start cold in every conversation. North Admin, its management console, adds token spending caps, user quotas, and organization-wide limits, tracked down to the individual agent.
The platform runs in the cloud, on-premises, or fully air-gapped, and works with Cohere's own models or ones a customer already uses elsewhere. Cohere holds SOC 2, ISO 27001, and ISO 42001 certifications, aimed at buyers who can't send data to a public cloud.
Key Capabilities:

• Persistent memory: Agents retain context across sessions instead of starting fresh each time.
• Cost controls: North Admin tracks and caps token spending per user and per agent.
• Flexible deployment: Runs in the cloud, on-premises, or fully air-gapped for sensitive data.

Most AI systems need separate models stitched together, one for language, one for video, one for controlling a robot arm. Reka's Rho-1 puts all of it in a single 19-billion-parameter network.
Reka released a research preview of Rho-1 on October 5, trained on 320 Nvidia H100 chips over three months. It can watch real-time video, explain what it's seeing, generate new video, and output robot commands, all using the same shared understanding rather than passing information between separate tools. It also accepts a new instruction mid-video, adjusting the output as it generates.
Reka pairs it with a separate "inverse-dynamics" model that works backward from a video to guess what robot actions would have produced it. It's early: video output is capped at a low 672x384 resolution, and it's a research preview, not a production system.
Key Capabilities:

• Unified system: Text, image, video, and robot actions share one internal model, not separate tools.
• Real-time steering: Accepts new instructions while a video is still generating.
• Early stage: Video resolution is capped at 672x384, and it's a research preview only.

A new system called textGrain tags AI-written text from ChatGPT and Codex with an invisible signal, rolling out to EU users to comply with the EU AI Act, while API users worldwide can turn it on voluntarily.
OpenAI's textGrain embeds a statistical pattern in word choices rather than changing how text looks or reads. It's required under Article 50 of the EU AI Act, which took effect in August and mandates machine-readable labeling of AI-generated content. For now, only approved researchers can access the detector that checks for the watermark.
OpenAI's own numbers show the limits: editing just 10% of words in a passage cuts detection from about 92% to 66%, and editing a quarter of the words drops it to 17%. Anthropic added a similar watermark to Claude two months ago.
Key Capabilities:

• EU-only for now: ChatGPT and Codex watermarking applies only to text generated in the EU.
• Global API opt-in: Developers anywhere can turn textGrain on for select models starting today.
• Easily weakened: Light editing or paraphrasing sharply cuts detection accuracy.

What was reported yesterday as an unconfirmed scoop is now official. Reflection AI's Beam is a 501-billion-parameter open-weight model built to rival top Chinese systems like GLM and DeepSeek, with full weights coming later this month.
Reflection, founded by ex-Google DeepMind researchers, trained Beam on 23.8 trillion tokens using 10,500 Nvidia GB300 GPUs, then ran a four-week reinforcement learning process generating over 100 million practice rollouts. It scores 80.9 on SWE-bench Verified, a test of fixing real software bugs, and 80.1 on Terminal-Bench 2.1, roughly matching Z.ai's GLM-5.2 while reportedly using three to four times less compute.
Weights aren't public yet; Beam is available now only through a waitlist-gated API, with Apache 2.0 open weights promised later in October.
Key Capabilities:

• Scores published: 80.9 on SWE-bench Verified, 80.1 on Terminal-Bench 2.1, both Reflection's own numbers.
• Efficiency claim: Said to use 3-4x less inference compute than GLM-5.2 for similar reasoning results.
• Not fully open yet: Only an API waitlist is live; full weights arrive later this month.

GLM-5.3, Z.ai's latest coding and agent model, is now available through Amazon Bedrock, giving enterprise AWS customers a managed way to run it without building their own hosting.
Z.ai made GLM-5.3 and the faster GLM-5.3-Flash available on Amazon Bedrock on October 5, following earlier listings on AWS Marketplace. Z.ai says the model improved 50% over its predecessor, GLM-5.2, on the company's internal coding benchmark, a claim that hasn't been independently verified.
Bedrock access means enterprise buyers get prompt caching, cross-region inference, and AWS's standard data-retention terms, without routing requests through a separate Chinese-hosted API.
Key Capabilities:

• New Bedrock access: Live through AWS's managed console as of October 5.
• Enterprise-ready: Comes with prompt caching and cross-region inference support.
• Unverified gain: Z.ai's 50% coding-benchmark improvement over GLM-5.2 is a vendor claim.

Liquid AI's d1 answers yes/no and multiple-choice questions by returning a probability instead of generating text, and it just added vision, cutting classification costs by up to 200x versus GPT-6.1 Sol and Claude Opus 5.5.
Liquid tested the updated d1 against both rival models on six real tasks, from filtering support tickets to inspecting circuit boards, and it matched or beat GPT-6.1 Sol on four of six while answering in 200 to 300 milliseconds, far faster than either rival. Because it returns a probability in one pass rather than writing out an answer token by token, there's no output cost at all.
d1 is meant for decisions with a fixed set of possible answers, like moderation or ticket routing, not open-ended writing tasks. It's available now through Liquid's API console, free during this experimental phase.
Key Capabilities:

• Vision added: Now accepts images, text, or both as input in one call.
• Big cost cut: 19x to 200x cheaper per 1,000 runs than GPT-6.1 Sol or Opus 5.5.
• No output tokens: Returns a probability directly, instead of generating text.

Kivi turns speech into clean, ready-to-send text in any Mac app, powered by Sarvam's own speech-to-text model built for Indian languages and English.
Sarvam AI launched Kivi for Mac, letting users dictate in any application instead of typing. It runs on Saaras V4, Sarvam's latest speech model, which auto-detects the spoken language across English and 22 Indian languages and cleans up filler words before sending text.
Saaras V4 posts the lowest average word-error rate among models Sarvam benchmarked, including against OpenAI's GPT-4o Transcribe, on the company's own tests. It's free to try now at heykivi.ai.
Key Capabilities:

• Any-app dictation: Works system-wide on Mac, not limited to one application.
• 22+ languages: Auto-detects and transcribes across English and all 22 Indian languages.
• Cleanup built in: Removes filler words and formats speech into ready-to-send text automatically.