AI Tools Weekly: September 21--September 27, 2026
· 11 min read

AI Tools Weekly: September 21--September 27, 2026

Agents left the chat window this week, showing up in glasses, seller tools and team chat, while the cost of running AI models kept falling for builders.

AI agents spent this week climbing out of the chat box. Two major model makers cut prices, Meta said its Muse agent is moving onto smart glasses, Amazon opened its seller assistant to outside AI tools, and Google’s open-source agent runtime kept collecting stars — all inside seven days.

Sources and scope: this roundup is desk research from official announcements, release notes and pricing pages we read on September 25, 2026 — we did not run any of these models hands-on.

OpenAI ships GPT-6 Sol and Luna and halves API prices

OpenAI shipped GPT-6 Sol and GPT-6 Luna on September 22, 2026, halving API prices. Sol costs $2.00 per million input tokens and $10.00 per million output tokens; Luna costs $0.10 and $0.50 (OpenAI; QZ). An OpenAI spokesperson says the new rates carry no expiration date.

Go deeper: our OpenAI GPT-5.6 review covers the models these two replace.

Where this fits: Sol and Luna sit below GPT-6 Astra, OpenAI’s top model — a cheaper everyday tier for people building with the API, not a new flagship.

For scale, the previous GPT-5.6 Sol was $4 per million input tokens and $20 per million output tokens, and GPT-5.6 Luna was $0.20 and $1.20, so the cuts are large rather than cosmetic (9to5Mac). OpenAI says the change is permanent, not a promotion. The rollout reaches Codex and ChatGPT Work first, for Plus, Pro, Business, Enterprise and Edu users, while Free and Go users reach Luna through the desktop app (OpenAI). Both models appear in Work and Codex, not in Chat. OpenAI suggests Sol for complex coding and agent-style workflows and Luna for focused, high-volume jobs.

Anthropic releases Claude Opus 5.5

Anthropic released Claude Opus 5.5 on September 22, 2026, the first model in its new Claude 5.5 family. It costs $4 per million input tokens and $20 per million output tokens, twenty percent less than Opus 5 (Anthropic). It arrives with a one-million-token context window.

Go deeper: our Claude Code review shows how Anthropic’s models behave inside a real coding tool.

Where this fits: Opus 5.5 becomes Anthropic’s workhorse for heavy jobs, priced far below the company’s own Fable 5.1 tier.

Cache reads cost $0.20 per million tokens, sixty percent less than Opus 5, and cache writes cost $5 (Anthropic). A context window is the amount of text a model can hold in mind at once: Opus 5.5 holds 1M tokens, returns up to 128K tokens of output, and has a June 2026 knowledge cutoff (model docs). Anthropic says it costs about 40 percent less to run than Opus 5 on typical workloads and produces output more than 30 percent faster. The company says Opus 5.5 performs at the level of Claude Fable 5.1 on most work, and that Fable 5.1 is priced at $10 and $50. Anthropic also says Opus 5.5 scored best of any model to date on its automated behavioral audit. Two things for beginners to know: “thinking” can no longer be switched off on the API, and Anthropic raised five-hour usage limits on Pro, Max, Team and seat-based Enterprise plans and added a saveable rate-limit reset. Claude Sonnet 5.5 and Haiku 5.5 follow “in the coming weeks,” and Opus 5.5 is available on the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS (AWS).

xAI ships Grok 4.7 at unchanged pricing

xAI released Grok 4.7 on September 21, 2026, built on a larger base model than Grok 4.6 and trained longer on hard, multi-hour tasks. It runs at the same price and speed as Grok 4.6: $2 per million input tokens and $6 per million output tokens (xAI).

Go deeper: our Grok Build news post explains xAI’s coding tool.

Where this fits: Grok 4.7 is xAI’s coding workhorse, competing with Claude Opus 5.5 and GPT-6 Sol at a lower output-token price.

The model keeps a 500K-token context window. xAI reports Grok 4.7 scored 46.3 percent on CursorBench 4.0, up from 40.4 percent for Grok 4.6; 37.6 percent on Terminal-Bench 4.0, up from 20.3 percent; and 71.0 percent on DeepSWE v1.1 at high effort — company-reported numbers, so treat them as claims rather than independent results (xAI docs). A “Grok 4.7 Fast” variant runs at twice the speed for twice the price and is available only through Cursor and Grok Build, not the public xAI API. The standard model reaches users through Cursor, Grok Build, the Grok API, third-party coding harnesses, model routers and cloud platforms.

Meta Connect 2026: Muse brings the agent to your glasses

At Meta Connect 2026 on September 23 and 24, Meta put its Muse personal AI agent at the center of Mark Zuckerberg’s keynote and said Muse is coming to its AI glasses, activated by voice and able to act on what the wearer is looking at (Meta).

Full beginner guide coming: we are working on a walkthrough of wearing an AI agent — until then, our comparisons hub is a solid starting point.

Where this fits: Muse is Meta’s answer to ChatGPT and Gemini, except it ships inside hardware you wear instead of a browser tab.

Meta says Muse can act on what the wearer sees — a product on a shelf, a flier on a wall, a school-supply list. New this week: Muse Realtime Voice with voice design by description, Muse Realtime Avatar embodiment research, an email address for Muse, and computer use in Muse for Mac, where Muse can drive any app with permission (TechCrunch). Meta introduced a new AI glasses lineup across Ray-Ban, Oakley and Meta Glasses, including its first open-ear audio glasses, and says there will be over 100 styles of AI glasses by the end of 2026. It also debuted “Meta VR Glasses.” For context, Meta introduced Muse on September 8, 2026 as a personal AI agent running on a “Muse Secure VM” (Meta newsroom).

Amazon opens its seller tools to outside AI agents

At Amazon Accelerate 2026 on September 23, Amazon announced a Seller Assistant selling-partner plugin that brings the assistant into AI tools sellers already use, launching first with Amazon Quick and in beta with Anthropic’s Claude. Amazon is bundling a free twelve-month Quick Plus subscription (Amazon).

Go deeper: our comparisons hub tracks how assistant-style tools stack up against one another.

Where this fits: Amazon is letting rival assistants reach into its marketplace instead of forcing sellers into a single Amazon-only chatbot.

Seller Assistant gains persistent memory of each seller’s business plus new workflows that monitor and execute routine restocking and pricing tasks, with audit trails and seller-defined guardrails (GeekWire). Amazon says Seller Assistant has rolled out to over 90 percent of its selling partners in their native language, has hundreds of thousands of active users, and that sellers accept its recommendations over 90 percent of the time. It runs on Amazon Bedrock, using Amazon Nova and Anthropic Claude models.

Google’s Gemini 3.8 Flash TTS reaches general availability

Google’s Gemini API release notes for September 22, 2026 mark Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS as generally available, alongside a new Gemini API Voices endpoint (Google). The text-to-speech models turn written text into speech through the same API.

Go deeper: our Gemini 3 Flash review covers the model family these voices belong to.

Where this fits: voice is now a feature inside the Gemini API that developers already use for text, rather than a separate product to buy.

Gemini 3.8 Flash itself shipped on September 2, 2026 as Google’s most intelligent Flash model, built for long-running software engineering and autonomous agents at the speed and cost of 3.7 Flash (Google blog). “Generally available” means the models are out of preview and Google supports them for production use — the useful detail if you are choosing a voice for an app you plan to ship.

Ando raises $20M to treat AI agents as team members

Ando emerged from stealth on September 24, 2026 with $20 million in combined pre-seed and seed funding from Accel, Index Ventures and Emergence Capital (TechCrunch). Founder Sara Du built a team messaging platform where AI agents participate as workspace members rather than bolted-on bots.

Go deeper: our comparisons hub tracks the tools agents are moving into.

Where this fits: Ando bets that agents should have accounts and inboxes like colleagues, not just a sidebar inside someone else’s chat app.

TechCrunch framed the company as a Slack challenger, and that framing belongs to TechCrunch — Ando has published no adoption numbers, so treat the comparison as positioning rather than settled fact (RuntimeWire). The $20 million in combined pre-seed and seed funding is the concrete figure.

Google’s open-source AX runtime passes 11,000 GitHub stars

google/ax is Google’s open-source runtime for running AI agents as coordinated services — free plumbing rather than a chat app. It is Apache-2.0 licensed and had 11.1k GitHub stars and 633 commits as of September 25, 2026 (GitHub). A September 20 commit split it into three programs.

Go deeper: our comparisons hub is where we track agent runtime options; a full AX beginner guide is coming.

Where this fits: AX is infrastructure for teams that need to run many agent jobs reliably, not a tool an individual opens and chats with.

The restructuring turned what was a single command-line tool with an embedded Python harness into a general-purpose orchestration layer — orchestration meaning the job of keeping many automated steps running in the right order. It is built around three binaries: ax-server (a gRPC API, a way for programs to call each other over a network, for submitting Task manifests), ax-controller (a horizontally scaled reconciler that consumes work from Redis Streams, a message queue) and ax-task-runner (which executes tasks inside sandboxed workers). Task state moved to Redis rather than Kubernetes custom resources (GitHub). Google’s Go module documentation describes AX, short for “Agent eXecutor,” as a distributed agent runtime with native recovery and resumption — meaning a job can pick up after a crash instead of starting over (pkg.go.dev).

The week ahead

Watch Anthropic’s Claude Sonnet 5.5 and Haiku 5.5, which the company says are coming in the coming weeks. Watch the GPT-6 Sol and Luna rollout too: OpenAI says Plus, Pro, Business, Enterprise and Edu users see the new models first in Codex and Work, while Free and Go users reach Luna through the desktop app. And watch whether Meta’s Muse reaches existing AI glasses styles as the 100-plus style rollout continues toward year-end.

FAQ

Which of this week’s stories matter most for a beginner?

The three price stories. GPT-6 Sol and Luna, Claude Opus 5.5 and Grok 4.7 all target people who pay per use, and lower prices usually trickle into the tools you already open. Muse on AI glasses is the most interesting if you want to see where assistants are heading, Gemini 3.8 Flash TTS matters if you are building something with a voice, and AX matters if you write code for a living.

Do these lower API prices mean cheaper tools for me?

Not automatically. API prices are what developers pay per unit of text processed, and a subscription app can keep its price even when its costs fall. But cheaper models often show up later as lower usage fees or more generous free tiers. OpenAI says its new rates carry no expiration date, Anthropic says Opus 5.5 costs about 40 percent less to run on typical workloads, and xAI held Grok’s price flat while improving the model.

What is an AI agent, and why are agents in so many of this week’s stories?

An agent is a model that does more than answer — it takes steps toward a goal, calling tools, browsing, editing files or restocking inventory along the way. That is why this week’s news is all about places agents do work: on glasses with Meta’s Muse, inside Amazon seller tools, in team chat with Ando, and in the plumbing Google’s AX provides for running them.

Ready to pick something? See how these tools compare across everything we track.

  • NiteAgent — AI agent development, frameworks, and production patterns
  • ToolBrain — tool reviews, LLM comparisons, and AI workflow guides
  • NoCode Insider — AI workflow automation with no-code tools, agents, and APIs

Cross-links automatically generated from None.

Back to all posts