Daily AI Zine · No. 068

Today's Big Thing

The one thing that matters

Test it

Claude Opus 5.5 enters the Copilot workbench

The important part is placement: a new Claude model is now available where repository agents already run.

Claude Opus 5.5 is now available in GitHub Copilot, with GitHub positioning it for agentic coding, long-running agentic tasks, and knowledge work. That makes this a practical release rather than a remote model announcement: it can be tested inside the coding environment instead of through a separate workflow. Japanese coverage also reports that Anthropic introduced Opus 5.5 as a lower-cost successor in the Opus line, which matters if it can handle serious implementation and review work without pushing model spend up. Treat this as a controlled trial, not a default swap: give it one repo issue that needs planning, editing, and verification, then compare its output against your current Copilot model.

T1GitHub Changelog09/22 T2ASCII.jp09/22

My AI Ecosystem

Your actual stack

Test it

GPT-6 Sol and Luna widen Copilot’s model menu

OpenAI’s GPT-6 Sol and GPT-6 Luna are now available, and GitHub says both are being added to Copilot alongside GPT-6 Astra. The practical change is model routing: coding work can now be split by capability and cost instead of treated as one default lane. Do not change defaults yet. Run a small comparison across one bug fix, one refactor, and one documentation task.

T1GitHub Changelog09/22 T1OpenAI News09/22

Test it

Hot signal: agent-made changes are getting isolated preview rooms

Cloudflare introduced Worker Previews, giving every branch its own URL, configuration, state, and observability. The hot signal is the infrastructure pattern: agents can make parallel changes without sharing production state. For any agent-assisted site, API, or automation workflow, this is the direction to copy even if the stack is different. The value is fewer risky merges and cleaner reviews.

T1Cloudflare Blog09/22

Watch it

OpenAI formalizes third-party assessment expectations

OpenAI published priorities and principles for independent third-party assessments of frontier models and safeguards. This is not a feature release, but it matters for procurement and client assurance. When model vendors make safety or reliability claims, ask what was independently assessed, what was excluded, and whether the assessment covered real tool use rather than chat-only behavior.

T1OpenAI News09/22

CoWork Corner

Claude CoWork, day to day

CoWork remains steady, tighten the connector habit

Claude CoWork is the workspace Adrian uses for persistent projects, operational records, and AI-assisted workflows. No meaningful CoWork product change was found in today’s candidates. The useful workflow to revisit is connector discipline: Claude Desktop’s one-click MCP server installation makes adding tools easier, so the operating rule should be stricter, not looser. Add one connector only when a project needs it, record what data it can reach, then remove it when the job is done.

Tier 4 · quiet day, honest fallback

T1Anthropic Engineering Blog

GPT Desk

OpenAI, ChatGPT, Codex

Test it

Route Sol and Luna before making either a default

What changed: OpenAI introduced GPT-6 Sol and Luna, and GitHub says both are joining GPT-6 Astra in Copilot. Why Adrian should care: Grey Group OS, Goodsense, and SET work all mix writing, code, research, and deck logic, so one default model is likely wasteful. What to do: make a three-task bench today, one proposal paragraph, one repo issue, and one research summary, then score Sol, Luna, and Astra on usable output per attempt.

T1OpenAI News09/22 T1GitHub Changelog09/22

Test it

Use Astra’s research case as a cost-control test

What changed: OpenAI says Parallel used GPT-6 Astra to research and synthesize labor-market data in half the time and at half the cost versus prior models. Why Adrian should care: the same pattern fits Japan market scans, production references, and client briefing work for Grey Group and Street Attack Japan. What to do: run one paid-style briefing prompt through Astra, capture source trace quality, editing time, and whether the result is client-safe without a full rewrite.

T1OpenAI News09/22

Small Money Systems

Small, repeatable, real

Use it

Small system: Japan market snapshot desk

System: build a fast paid briefing desk for Japan labor, creator, venue, or category snapshots. Customer: overseas producers, brands, and founders who need a Japan read before committing budget. Offer: a short source-linked brief with risks, local context, and next-step recommendations. Price: ¥55,000 per snapshot. Existing assets: Grey Group OS, Goodsense positioning, Japan-English business workflow, and past production research habits. AI workflow: GPT-6 Astra drafts the synthesis, then Claude Code or Codex turns the brief template into a reusable intake and export flow. First action: write one intake form and one sample topic today. Repeatability: every brief reuses the same prompts, structure, and source checklist. Effort: one hour. Expected value: small direct revenue and better lead qualification.

T1OpenAI News09/22

Test it

Small system: model-routing audit for AI-heavy teams

System: sell a compact model-routing audit for teams using ChatGPT, Copilot, Claude, or Codex without a cost rule. Customer: small agencies, production companies, and startups already paying for multiple AI tools. Offer: a one-page routing map that says which model to use for writing, coding, research, and review, plus one sample task bench. Price: ¥77,000. Existing assets: Grey Group OS, Daily Ops, and Adrian’s working stack across GitHub, Codex, Claude Code, and OpenAI APIs. AI workflow: use the new Copilot model options as the test surface, with Claude Code recording the results. First action: create a three-task benchmark sheet. Repeatability: each client gets the same audit shell with different tasks. Effort: half day. Expected value: consulting revenue and lower wasted AI spend.

T1GitHub Changelog09/22 T1GitHub Changelog09/22

Build Next

Deployable now

Use it

Build a model bake-off logger for real repo work

What to build: a small GitHub issue template and results table that records model used, task type, first-pass quality, edit time, and final decision. Why now: Claude Opus 5.5, GPT-6 Sol, and GPT-6 Luna are all moving into Copilot, so model choice is becoming an operating decision. Effort: one hour. Expected impact: better output quality and less wasted model spend. Dependencies: an active GitHub repo, Copilot access, and three comparable tasks ready to test.

T1GitHub Changelog09/22 T1GitHub Changelog09/22

Use it

Build an agent preview checklist before the next merge

What to build: a lightweight checklist that requires every agent-made branch to show its preview target, config differences, state isolation, and rollback note before review. Why now: Cloudflare’s Worker Previews make isolated branch testing explicit, and the same discipline is useful even outside that platform. Effort: 15 minutes. Expected impact: fewer risky merges and cleaner handoff reviews. Dependencies: a branch-based workflow, preview deploys or staging links, and a reviewer willing to block incomplete agent work.

T1Cloudflare Blog09/22

Try This Today

One action, right now

Pick one real repo issue today. Ask Claude Opus 5.5, GPT-6 Sol, and GPT-6 Luna for the same plan or patch. Save only three notes: best first answer, least editing needed, and safest reviewer handoff. Keep the winner for that task type.