AI

AI models, tools, agents, generative AI, AI policy, and industry impact.

Flat illustration of a server cabinet beside an open laptop on overlapping blue and coral circles.

AI2 days agoBy megan

Recursive Flint’s local Mac setup calls for at least 48GB of memory

The coding model activates about 2.5B of its 39.8B parameters per token, but its laptop route uses separate 4-bit weights. Recursive’s BF16 checkpoint targets a different, larger hardware setup.

Read article

Abstract data streams pass through a glowing cube toward a grid of checks.

AI6 days agoBy megan

DeepSeek’s Flash API aliases now serve V4.1-Flash-1017

DeepSeek changed the model behind its Flash API names on October 17 while keeping the calling method and pricing unchanged. Teams should test their own prompts, tool calls and output handling against the new checkpoint.

Read article

AI7 days agoBy megan

Jev reached 35.3% of Vercel AI Gateway’s paid teams, but 0.27% of spend

Vercel’s October 15 index shows broad early use of a decision model on its own gateway. Team reach and estimated spend measure different things; neither establishes wider market adoption or model quality.

Read article

Two abstract compute slabs with divided blocks joined by green cables on a pale background.

AI8 days agoBy content

Apple runs Muse Spark Open across two M5 Ultra Mac Studios

Meta’s open weights are available alongside hosted APIs, but Apple’s local MLX test shows why the 816B-parameter model needs more than one top-memory Mac. Its 4-bit footprint is about 493GB before inference overhead.

Read article

A ribbed ceramic token sits in one recess of a terracotta disk, beside other empty recesses and a small green stop.

AI15 days agoBy megan

Google details Gemini agent identity, network and cost controls

Google says its new work agent will use a distinct identity, a policy gateway and a project spend cap. IT teams need to test how those controls behave and confirm which capabilities are available.

Read article

Abstract illustration of a vertical threshold separating cool grey and coral blocks, representing a price change above 100,000 prompt tokens.

AI16 days agoBy megan

Claude Haiku 5.5 costs more above 100K prompt tokens

Anthropic’s new small model offers low API rates for shorter prompts, but requests above 100,000 prompt tokens move to a higher price tier. A move from Haiku 4.5 also requires checks for token counts, thinking settings and request compatibility.

Read article

An open ivory tray and a closed steel-lidded tray sit beside a loose green enamel tab on a pale textured surface.

AI17 days agoBy megan

Meta and Sierra propose OAuth-based protocol for personal AI agents

Personal Agent Protocol would let an agent discover a business’s services, start a guest or account session, and work through its website, APIs, or business agent. Sierra plans to publish the v0.1 specification later in October 2026, so the announced flow is not yet an implementable standard.

Read article

AI18 days agoBy megan

Mistral Large 4 enters API preview while its model weights remain unreleased

Mistral’s multimodal model is available through a public preview API, but developers must wait for downloadable weights before assessing self-hosting.

Read article

Painted blue and orange paths pass through an adjustable metal collar.

AI19 days agoBy megan

Cohere’s North 2 reworks agent orchestration and adds token controls

Cohere’s October 5 upgrade combines a redesigned system for multi-step agents with cross-session memory and more granular usage limits. Operators still need to validate reliability, security, and cost in their own deployments.

Read article

Three overlapping paper panels have one, two and three square openings.

AI20 days agoBy megan

Gemini Apps model access changes start October 9 for users without a plan

Google’s October 2026 plan tiers will leave personal accounts without a subscription with Flash-Lite. AI Plus will offer Flash as well, while AI Pro and AI Ultra will offer Pro. The rollout timing differs by plan.

Read article

AI22 days agoBy megan

Strands Decider 2B Publishes Code and a Local Model Checkpoint

Strands has released a 1.9-billion-parameter model that scores supplied options rather than writing text. Its code and recipe are inspectable, but confidence thresholds and exact reproduction need careful testing.

Read article

AI23 days agoBy megan

Google Unveils Gemini 4 Argon With Limited Cyber Defender Access

Google is giving selected Fairwind partners the first access to Gemini 4 Argon. Its output limit and API prices are announced, but broader developer availability has no date.

Read article

AI25 days ago

Meta Muse Adds Business Connectors: How Approval Works

Muse for Small Business connects to tools including Shopify, Slack and QuickBooks. Operators should review permission scope: an approval can cover later actions without another prompt.

Read article

AIabout 1 month ago

GPT-6 Sol and Luna Cut API Prices, With Access Split Across Products

OpenAI’s new mid-tier and lightweight models cost less than their GPT-5.6 predecessors, but access differs across the API, ChatGPT Work, Codex, desktop apps and GitHub Copilot.

Read article

AIabout 1 month ago

OpenAI’s Math Advisory Group Is Independent, but Its Advice Is Not Binding

The group can scrutinise results, recommend release practices, publish its advice and criticise OpenAI. It cannot make company decisions or influence the pace of OpenAI’s internal mathematics research.

Read article

AIabout 1 month ago

Jev Trades Free-Form Text for Typed AI Decisions

TypeSafe AI’s early-access model returns bounded choices, scores, and probabilities instead of prose. That removes some integration failures, but correctness still depends on schemas, thresholds, evidence, and evaluation.

Read article

AIabout 1 month ago

DeepSeek-V4.1-Flash pairs cache changes with a new prompt format

The release changes how cached attention data is handled. Operators also need to account for its prompt format and the limits of its integration toolkit.

Read article

AIabout 2 months ago

GLM-5.3 Weights Arrive Under a Custom Licence, Not MIT

Z.ai released downloadable GLM-5.3 weights on August 28 under a custom licence. Broad rights remain, but very large Model-as-a-Service operators must clear a security review before commercial use.

Read article

AIabout 2 months ago

Anthropic previews Model Hardware Standard for AI-controlled equipment

MHS gives AI agents a common driver layer for programmable lab and factory hardware, but it remains an application-only research preview that requires expert oversight.

Read article

AI2 months ago

OpenAI Lowers GPT-5.6 Sol API Prices Through at Least November 21

GPT-5.6 Sol now costs $4 per million input tokens and $20 per million output tokens under promotional pricing available at least through November 21, 2026. Prompts above 272K input tokens use higher rates for the full request.

Read article

AI2 months ago

OpenAI previews cross-interaction safety checks for Zero Data Retention customers

Private Safety Processing is being tested with early customers. OpenAI says it can detect patterns across related interactions without giving personnel the underlying content, but the technical paper and measurable results are still pending.

Read article

AI2 months ago

Qwen3.8-27B Makes Reasoning Effort a Deployment Control

Alibaba’s Qwen team released a 27-billion-parameter open-weight model with tunable reasoning depth and a 262,144-token native context. The extra control brings latency, memory and task-quality tradeoffs that operators need to test.

Read article

AI2 months ago

The harder test for LLM mathematics is finding new methods

OpenAI reports ten results from its internal Astra model. Timothy Gowers argues that search-tree pruning and genuinely new methods are a harder capability test than result volume.

Read article

AI2 months ago

Why Hugging Face Download Counts Aren’t Model-Adoption Numbers

Hugging Face found only one repository in both its top 25 by downloads recorded during its 2026 measurement window and its top 25 by likes. The mismatch shows why request counts cannot stand in for users or production deployments.

Read article

AI2 months ago

Gemini 3.7 Flash Is GA, but Its 2026 Price Is Temporary

Google’s production model adds a new Flash target for coding and agents. Developers should budget for a January 1 price increase and retest migration settings, thinking levels and vendor benchmarks.

Read article