Claude Haiku 5.5 launches at about 75% lower running cost

Overview

9 stories in this issue. The first 3 are today's priorities.

Hot Model Updates

1. Top · Claude Haiku 5.5 launches at about 75% lower average running cost than Haiku 4.5
2. Top · Fine-tuned NVIDIA Nemotron models reach gold-medal level at IOI and IMO 2026
3. Top · Liquid AI releases open weights for d1-3B and d1-omni-600M multimodal decision models

Global AI News

4. Microsoft reveals specs and prices for its Nvidia-chip Surface Laptop Ultra
5. Google launches Playground for building browser games from text prompts
6. ChatGPT for Teens will add College Planner for college applications
7. US government, tech companies and Biohub commit $1.8B to AI biology datasets

Regional & Early Signals

8. Nous Research raises a $90M Series B and launches agents for business users
9. Amazon Quick and Bedrock Knowledge Bases add a query-time permission check to RAG

Jiufeng graphic based on the sources cited in this issue.

Hot Model Updates

Anthropic's new small model targets subagents and high-volume workloads, and it is available on Amazon Bedrock.

Anthropic introduced Claude Haiku 5.5 on October 7th, calling it the cheapest, fastest and most capable small model it has released. The company says it costs around 75% less to run on average than Haiku 4.5. AWS has announced that Haiku 5.5 is available on Amazon Bedrock and Claude Platform on AWS. Citing Anthropic, AWS says it is the fastest and most efficient model in the Claude 5.5 family, built for subagents and cost-sensitive, high-volume work.

• Bedrock: Regional data residency, with IAM access control, CloudTrail audit, CloudWatch monitoring and Bedrock Guardrails; usage is billed through AWS
• Claude Platform on AWS: Anthropic's native APIs, features and console, accessed through the AWS Management Console with AWS billing and authentication
• Getting started: aws-samples has published a Getting Started notebook

Limitations: RuntimeWire reports that Anthropic's announcement gives no per-token prices, benchmark results, context limit or API identifier. Haiku 4.5 is documented at $1 per million input tokens and $5 per million output tokens. Until new prices are published, the 75% average running-cost claim cannot be converted into a new token price.

Image source: Amazon Web Services; mirrored on Jiufeng R2.

Source: RuntimeWire · AWS Machine Learning Blog · Getting Started notebook

Two specialist versions built on the same Nemotron 3 base reached gold-medal level in informatics and mathematics olympiads.

NVIDIA researchers write on the Hugging Face blog that they started from Nemotron 3 and used supervised fine-tuning (SFT), reinforcement learning (RL) and feedback-driven inference to build two specialist systems. For IOI 2026, Nemotron-3-Ultra-CC with SFT and GenCorrect scored 535.4/600. For IMO 2026, a generate-verify-refine system also reached gold-medal level.

The IMO training approach is described in an arXiv paper, and the NeMo-Skills repository includes the IMO inference pipeline recipe.

Limitations: These results are self-reported by NVIDIA's team in its own blog post and paper. Each competition used a separately specialized model, so this is not one general-purpose model winning both.

Source: Hugging Face Blog · arXiv paper · NeMo-Skills

Liquid AI released Open d1, two open-weight models in its d1 decision model family. They accept image and audio inputs and do not generate text.

Liquid AI released Open d1, two open-weight multimodal models:

• d1-3B: Accepts text and images
• d1-omni-600M: Accepts text with an image, or text with audio
• Output: No text output; each model returns calibrated, typed answers in one forward pass with zero output tokens
• Deployment: Weights are on Hugging Face and load through Transformers, with day-one llama.cpp support; target hardware is DGX, RTX workstations and Jetson edge boards
• License: LFM Open License v1.0, which allows free commercial use below $10 million in annual revenue

Limitations: d1-omni-600M is an early research release with no published latency figures.

Source: MarkTechPost · d1-3B model card

Global AI News

Microsoft priced AI PCs that run on Nvidia chips and are designed to run AI models and agents.

At an event in San Francisco on Wednesday during the city's Tech Week, Microsoft revealed the specs and prices for the Surface Laptop Ultra. These AI PCs run on Nvidia chips and are designed to run AI models and agents, and they come with a revamped Windows 11.

Limitations: Microsoft says the top configuration is already out of stock. TechCrunch is the only reporting source so far.

Source: TechCrunch

Playground is an experimental game-creation platform powered by Gemini, Nano Banana and Lyria, and it is free for adults in the US.

In Playground, users describe changes in text to adjust game rules, physics, characters and environments, then test the results right away. Finished games can be shared by link or published to a public gallery, and some genres support leaderboards and multiplayer. The platform is free, and Google One subscribers get higher weekly usage limits depending on their plan. On the same day, Google and Unity announced Unity Spark, a tool for professional developers with access to the Unity Asset Store.

Limitations: Playground is an experiment open only to US adults. Published games go through safety reviews based on Google's community guidelines, and Unity Spark is still headed for a planned closed beta.

Source: Google Blog · The Decoder

OpenAI wants ChatGPT for Teens to cover the full application process, from deadlines to financial aid, while its teen safeguards face outside criticism.

OpenAI says College Planner will gather requirements, deadlines, tasks and financial-aid steps for the schools a student is considering. It will also support progress tracking, scholarships and fee waivers. The first US version targets grades 10 through 12 for students applying to four-year colleges. OpenAI plans to expand later to other countries and to two-year colleges, technical colleges and trade schools. The same update adds flashcards, quizzes and a teen AI council.

Limitations: OpenAI has not given a release date. On the same day, a Common Sense Media assessment called the teen experience an unacceptable risk for users under 18, disputing whether the safeguards that are on by default work reliably in high-risk conversations.

Source: OpenAI · RuntimeWire

A consortium of more than six public and private organizations wants to build the training data needed for virtual-cell models.

SiliconANGLE reports that the consortium announced the effort on October 7th and committed $1.8 billion to it. The goal is to produce the large biology datasets needed to train biology-optimized AI models. According to the report, testing how a cell responds to a new therapy usually requires a physical lab setup that can take years and millions of dollars, and reliable simulated cells could shorten that process.

Limitations: The report says ultra-detailed simulations that can replace lab equipment do not exist yet. This commitment funds data production, not a finished model.

Source: SiliconANGLE

Regional & Early Signals

The developer of the open-source Hermes Agent confirmed a $1.5 billion valuation and released an agent product for business users.

Robot Ventures led the round, with participation from Nvidia, Union Square Ventures, Menlo Ventures, Samsung and 1789 Capital. The three-year-old startup's total funding is now $158 million. It also launched AI agents for business users.

Limitations: TechCrunch is the only source. The excerpt gives no details on the business agents' features, pricing or availability.

Source: TechCrunch

AWS uses two layers for enterprise RAG permissions: filtering on ACL attributes stored in the index, then a real-time check against the source at query time.

AWS says Amazon Quick and Amazon Bedrock Knowledge Bases add a real-time access control list (ACL) check at query time on top of the existing pre-retrieval ACL filtering, which relies on stored ACL attributes. The query-time layer verifies ACLs directly with authoritative sources such as SharePoint, Google Drive and Confluence, so that each employee only receives AI-generated answers drawn from documents they are authorized to access.

Limitations: This is an AWS vendor blog post describing its own products, and there is no independent evaluation yet.

Source: AWS Machine Learning Blog