- NATURAL 20
- Posts
- Anthropic Launches Haiku 5.5 as OpenAI and Microsoft Accelerate AI Agents
Anthropic Launches Haiku 5.5 as OpenAI and Microsoft Accelerate AI Agents
PLUS: GPT-6 reaches more ChatGPT users, and Google unveils its Gemini agent for work.
Wes Roth
October 09, 2026
In partnership with
Your SOC 2, handled end to end with AI agents.
Enterprise buyers won’t put your product near their customer data without a SOC 2 report. Twenty-one state privacy laws now hold them accountable for the vendors they share data with, so their diligence lands on you.
Sprinto gets you audit ready in 14 days, across three working sessions. AI agents connect your stack, collect the evidence auditors ask for, and close the gaps as they appear. You approve, they execute.
Sprinto also answers the security questionnaires your buyers send, using the same evidence base, so security reviews stop holding up deals.
No compliance hire. No consultant. Your auditor signs off.
Today:
Claude Haiku 5.5 targets faster, lower-cost tasks and coding subagents
OpenAI Ultrafast mode delivers up to eight times faster API processing
Windows combines local AI models, cloud systems and agent security
Claude adds live dashboards and editable animations
OpenAI rolls out GPT-6 and Intelligent UI to more ChatGPT users
AI Is Moving Toward Faster Models, Hybrid Computing and Agents
New AI releases focus on response time, operating costs and where models run. Anthropic is scaling down the price of routine intelligence, OpenAI is offering a faster service tier, and Microsoft wants Windows agents to switch between local hardware and cloud models as needed.
The common theme is deployment, not model size alone. As agents take on longer workflows, developers increasingly need to choose how much intelligence, latency, spending and device access each task requires.
Anthropic introduced Claude Haiku 5.5, its fastest and most capable small model, for frequent requests such as summaries, classification, database queries and routine information work. It is also designed to serve as a coding subagent alongside larger models and handle speed-sensitive applications including live customer support and browser use.
Anthropic estimates that the new model costs about 75% less to operate on average than Haiku 4.5. That matters for applications that make thousands of short model calls, or agent systems that delegate simpler steps to a cheaper model rather than using a flagship model for everything.
Haiku 5.5 adds adjustable reasoning effort, letting developers tune the trade-off between capability, latency and expense. It is available through the Claude Platform and supported cloud partners, including AWS, Google Cloud and Microsoft Azure.
Alongside the launch, Anthropic halved Claude Sonnet 5.5 cache-read prices, estimating about 20% lower costs for most agentic workloads. Claude Max and Team subscribers are also receiving monthly API credits for building applications on the Claude Platform.
OpenAI's Ultrafast mode is the fastest service tier in its API, offering up to eight times the speeds of Standard processing on supported workloads. It targets interactive applications and agent systems where a long chain of generation and tool calls makes latency especially noticeable.
Developers can opt into the tier using the service_tier setting set to ultrafast. OpenAI recommends persistent WebSocket connections for tool-heavy workflows so repeated network requests do not offset the benefits of faster model inference.
GPT-6 Astra is supported for API users subject to low default limits, and OpenAI's October 8 developer changelog added GPT-6.1 Sol to Ultrafast. Preview access for GPT-5.6 Sol is also described in the documentation. Availability, limits and regional processing support depend on the selected model.
Ultrafast is priced above standard processing, making it a workload-specific trade-off. High-volume background jobs may be better served by lower-cost tiers, while live agents and interactive tools can benefit more directly from lower latency.
Microsoft outlined plans for Windows as a hybrid intelligence platform where agents can perform work using either local models or cloud services. Its strategy combines device-level compute with security, identity and administrative controls for agents accessing files and tools.
Microsoft Execution Containers are now generally available on Windows, allowing policies to restrict an agent's access to resources. The company also wants enterprise administrators to distinguish agent activity from human activity and manage agents using tools such as Microsoft Agent 365 and Intune.
For local coding intelligence, Microsoft described deploying MAI Code 1.1 Flash, a model with 137 billion total parameters and 6.8 billion active parameters, using 3-bit compression and a 256,000-token context window. GitHub's HydraFusion will be able to route tasks to on-device models, with an experimental preview expected later in October.
Microsoft says Copilot will use permitted local context, take actions on Windows and select local models when appropriate. Those Copilot+ PC features are scheduled to begin rolling out over the coming months rather than being generally available today.
🧠RESEARCH
STEPQuant proposes a way to reduce memory used by the recurrent states of linear-attention AI models by varying numerical precision according to how quantization errors affect outputs over time. Tests on Qwen3.8-27B and Kimi-Linear-48B-A3B-Instruct reportedly matched FP32-state accuracy closely with a nominal six-bit budget, compressed recurrent states more than fivefold and lowered total serving memory by up to 68.7%. The findings are author-reported and depend on the evaluated architectures, kernels and serving conditions.
Long-WAM combines autoregressive video pretraining with longer visual context to guide robots through changing tasks without excessively slowing control. Its authors report RoboCasa GR-1 task success improving from 63.3% to 78.7% as visual history expands from zero to 19.2 seconds, plus 95% success in a dynamic cup-stacking test on a real robot. Their optimized system takes 107.4 milliseconds per action chunk on an RTX 5090, including future-video latent prediction.
OPD Before RL proposes a two-stage training framework for answers that cannot be graded by exact correctness alone. First, a student model learns from a rubric-aware teacher through on-policy distillation, and then reinforcement learning optimizes responses against rubric scores. Across HealthBench, ResearchQA and RubricHub Science, the authors report stronger scores than the comparison methods and fewer signs of reward hacking than a supervised-fine-tuning baseline.
📲SOCIAL MEDIA
🗞️MORE NEWS
Anthropic released Claude Dashboards in beta, allowing users to connect services such as BigQuery, Databricks, Snowflake and Salesforce and generate automatically refreshed dashboards from questions in ordinary language. Claude Motion creates editable code-based animations for presentations and reports that can be exported as MP4 videos without generating synthetic footage. Dashboards is in beta on paid plans and Motion is in beta for Team and Enterprise, while Claude Docs, Slides and Design are now out of beta for all plans, including Free.
OpenAI expanded GPT-6 in ChatGPT alongside Intelligent UI, which can mix explanations with charts, diagrams, buttons, forms and interactive tools. Interfaces can appear progressively as answers are generated, and GPT-6 can begin responding while it continues to reason. The rollout started October 7 for Plus, Pro, Business and Enterprise users and extended October 8 to Free and Go; the release does not change the models powering Work and Codex.
Google introduced a universal Gemini agent that can answer work questions, perform knowledge tasks, create media and write or run code using enterprise context. It is designed to work across Gmail, Drive, Docs, Slides, Sheets, Chat and Calendar, with reusable skills, connectors, security controls and orchestration involving more than one model family. Google's Gemini at Work announcements also include new analytics capabilities, specialized industry workflows and cost tools such as smart routing and spending caps.
OpenAI introduced the Decisions API in public beta to evaluate text and images and return typed results such as probabilities, category selections and scores against a rubric. It targets classification, request routing and prioritization, and OpenAI reports decisions around ten times faster than equivalent Responses API workflows. GPT-6 Luna is the only supported model at launch, accessed through the dedicated /v1/decisions endpoint.
What'd you think of today's edition? |
Reply