2026 · 09 · 04 · Fri

Updates for 9/4

Google and Meta both refreshed their flagship models this morning. Copilot also picked up a dirt-cheap transcription model, plus 4 fresh usage tips.

Today's moves

Model, pricing and availability changes we actually wrote into the map and the encyclopedia.

A · Theme of the day

Gemini's fast tier flips to 3.8

Google upgraded its fast model to 3.8.

Gemini's fast tier hits 3.8

Gemini (Google)

Compared to before

Two weeks ago the fast tier was still 3.7 Flash, and it stumbled on complex coding instructions.

What changed

Google's fast Gemini tier moved from 3.7 to 3.8 Flash, sharper on coding and reasoning.

Why it matters

Input stays $0.75 through year-end but doubles in January 2027, so heavy users should lean on it now.

Source: deepmind.google　·　→ View on chaos map

Read in depth →

B · Theme of the day

Meta quietly sharpens its reasoning model

Meta's Muse Spark reasoning model moved to 1.3.

Muse Spark bumped to 1.3

Llama (Meta)

Compared to before

Last month's 1.1 was priced right but a clear step behind the top models.

What changed

Meta's reasoning/agent model Muse Spark moved from 1.1 to 1.3, price unchanged at $1.25/$4.25.

Why it matters

Same price, better output is a real win for cost-conscious developers. Casual users won't notice the difference.

Source: ai.meta.com　·　→ View on chaos map

Read in depth →

C · Theme of the day

Copilot's transcription turns dirt cheap

Copilot picked up a dedicated cut-rate transcription model.

Copilot's transcription just got cheap

Microsoft Copilot

Compared to before

Until last month Copilot leaned on general-purpose third-party models for transcription, a mixed bag on cost and speed.

What changed

Microsoft added MAI-Transcribe-2 to Copilot, a transcription model cheaper and faster than the OpenAI/Google alternatives.

Why it matters

Anyone churning through meeting notes or captions at volume benefits most; occasional users may not notice.

Source: copilot.microsoft.com　·　→ View on chaos map

Read in depth →

Today's Tips

Practical ways to use these tools, picked up from social media and community threads.

T1 · Theme of the day

Tools start answering to other tools

More paths to drive tools via MCP or a CLI flag.

Runway ships an official MCP server

Runway

How to do it

Connect Runway's MCP server from Claude, ChatGPT, Cursor, or Replit to generate images/video with Gen-4.5 and more.

Why it helps

Generate assets without leaving your usual chat tool. If you edit video elsewhere, it's just one more connection to manage.

Source: x.com　·　→ View on chaos map

Copilot CLI gets a one-shot -p flag

GitHub Copilot

How to do it

Run copilot -p "fix this bug" to execute a one-shot prompt with no interactive session.

Why it helps

Great for calling Copilot from CI or scripts and getting just the result back. Interactive users may never touch it.

Source: x.com　·　→ View on chaos map

v0 can now run terminal commands

v0 by Vercel

How to do it

v0 runs terminal commands: testing in a browser, reading git history, and running tests, all inside the chat.

Why it helps

Check your generated UI without spinning up a local server yourself. Backend-heavy projects benefit less.

Source: x.com　·　→ View on chaos map

T2 · Theme of the day

A map appears for the Kiro migration

A quick note for anyone moving from Amazon Q to Kiro.

Amazon Q Developer to Kiro config map

Amazon Q Developer / Kiro

How to do it

Config files map as: .amazonq to .kiro, cli-agents to agents, rules to steering.

Why it helps

One glance tells you where each file goes when migrating. If you're still on Q Developer, skip this for now.

Source: x.com　·　→ View on chaos map

Past updates

A daily archive of changes actually applied to the site.