Daily AI Zine
Monday, August 24, 2026
Issue No. 039 · Tokyo · full edition

✦ Today's Big Thing

Cheaper flagship AI makes routing the real work

The useful move today is not chasing another model. It is putting expensive reasoning in the right lane and pushing routine work down to cheaper paths.

5 min read · 7 sections

In Brief
  1. OpenAI’s reported temporary GPT-5.6 Sol API discount makes model routing a budget decision today.
  2. Asana’s Codex migration case is another signal that bounded engineering debt is the right shape for coding agents.
  3. Runway’s Gen-4.5 claim keeps video generation on the production watchlist, but it still needs a controlled test.
  4. Cloudflare’s Kitesurf points to agent browsing becoming infrastructure, not a desktop trick.

Today's Big Thing

The one thing that matters

Test it

GPT-5.6 Sol’s price cut makes routing the practical lead

A cheaper flagship lane matters only if it changes what gets sent where.

OpenAI reportedly cut GPT-5.6 Sol API pricing for a limited period, with input down 20 percent, output down 33 percent, and the promotion running at least three months. The same report says ChatGPT Work and Codex credit consumption is also reduced, with spend tracking and control features announced. The practical shift is not “use the biggest model for everything.” It is to reroute work: high-judgment drafting, code review, and planning can get the premium lane, while summaries, extraction, and formatting stay on cheaper paths. Adrian should treat this as a budget window for measurement, not a shopping signal.

T2ITmedia AI+08/22T1OpenAI News08/13

My AI Ecosystem

Your actual stack

Use it

Codex keeps proving itself on bounded migrations

Asana says it used Codex to replace an outdated testing system in two weeks, completing work it expected to take five years for about $12K. The lesson is shape: Codex is strongest when the job has a clear target system, tests, and a finish line, not when it is asked to “clean up the repo.”

T1OpenAI News08/18
Watch it

Hot signal: Cloudflare puts agent browsing inside Workers

Cloudflare introduced Kitesurf, an agent-first browser that runs in V8 isolates on Cloudflare Workers. In plain terms, it gives agents a stateless browser designed for large-scale web tasks, rather than relying on a normal desktop browser session. This is early infrastructure, but it is relevant to research, monitoring, and repetitive web workflows.

T1Cloudflare Blog08/06
Test it

Runway Gen-4.5 stays interesting, but test before trusting

Runway describes Gen-4.5 as its best video model, citing motion quality, prompt adherence, and visual fidelity. That is relevant for commercial production and pitch assets, but the right move is a small test set: product motion, human movement, signage, and Tokyo street textures. Do not judge it from demo clips alone.

T1Runway Research
Use it

Claude Code work still comes down to harness and review

Anthropic’s Claude Code material continues to point at agentic coding as a managed workflow, not a prompt-and-pray habit. The useful pattern is stable: define the task, constrain the change, run tests, inspect the result, and only then merge. That is still the operating discipline to keep around Claude Code.

T1Anthropic Engineering BlogT1Anthropic Newsroom

CoWork Corner

Claude CoWork, day to day

CoWork: no product move, audit one project room

Claude CoWork is the workspace Adrian uses for persistent projects, operational records, and AI-assisted workflows. No meaningful CoWork product change showed up in today’s candidates. The useful workflow to revisit is tool hygiene: Claude Desktop has documented one-click MCP server installation, so pick one project room and list which connected tools are allowed, which are read-only, and which should never touch client-sensitive material. The goal is not more capability. It is a cleaner room before the next agent task starts.

Tier 4 · quiet day, honest fallback

T1Anthropic Engineering Blog

GPT Desk

OpenAI, ChatGPT, Codex

Test it

Use the GPT-5.6 Sol discount as a budget experiment

What changed: OpenAI reportedly lowered GPT-5.6 Sol API pricing for a limited window and reduced ChatGPT Work and Codex credit consumption. Why Adrian should care: Grey Group OS, Daily AI Zine, and Goodsense all have repeatable work where premium reasoning may help, but routine extraction does not deserve the top lane. Do this today: pick one recurring API workflow, run the same task through the current model and GPT-5.6 Sol, then log quality, latency, and cost before moving anything permanent.

T2ITmedia AI+08/22T1OpenAI News08/13
Use it

Give Codex one contained migration, not a vague backlog

What changed: Asana says Codex replaced an outdated testing system in two weeks, for about $12K, on work expected to take five years. Why Adrian should care: SET, Goodsense, and Grey Group OS likely have small pieces of old workflow glue that are annoying but bounded. Do this today: choose one outdated script, test path, or Netlify deployment helper, open an isolated branch, and ask Codex to replace it with an acceptance check before implementation.

T1OpenAI News08/18

Small Money Systems

Small, repeatable, real

Test it

Small system: QA-retirement sprint for old web workflows

System: build a one-week Codex-assisted cleanup sprint that replaces one outdated testing or release workflow. Customer: small production companies, studios, or Japan-facing operators with fragile web ops. Offer: a before-after branch, a simple test path, and a handoff note. Price: ¥120,000 to ¥180,000. Existing assets: Grey Group OS, GitHub, Netlify, Codex practice, and production-delivery discipline. AI workflow: Codex drafts the migration, Claude Code or a human reviews, and GitHub records the change. First action: find one internal old test or deploy helper and turn it into the demo case. Repeatability: each client gets the same audit, branch, test, and handoff structure. Effort: half day to package the offer. Expected value: repeatable project revenue and fewer fragile delivery steps.

T1OpenAI News08/18
Test it

Small system: activation audio recap add-on

System: build a fast sponsor recap pack that turns event notes, short interviews, and phone clips into a clean audio summary plus written highlights. Customer: Street Attack Japan sponsors, venues, and local brand partners. Offer: a short narrated recap, pull quotes, and a follow-up email draft. Price: ¥45,000 to ¥80,000 per activation. Existing assets: Street Attack Japan footage flow, production planning, and Adrian’s Japan-English business workflow. AI workflow: an MCP audio server handles speech generation or transcription, while Claude Desktop organizes the handoff. First action: make one internal sample from an old activation brief. Repeatability: the same recap template can run after each event. Effort: one hour for the sample. Expected value: small add-on revenue and better sponsor follow-up.

T2Perplexity Research08/19T1Anthropic Engineering Blog

Build Next

Deployable now

Test it

Build a ChatGPT visibility checker for one public offer

What to build: a small script that asks a fixed set of buyer prompts in ChatGPT search, records whether a public offer appears, and saves the answer with date, prompt, and source notes. Why now: a report says ChatGPT search is using the site operator at scale, and generative-engine visibility is becoming measurable. Effort: one hour. Expected impact: better sales copy and clearer positioning for public pages. Dependencies: a public page, a prompt list, and either Codex or Claude Code to build the first local version.

T2Simon Willison Weblog08/20T1OpenAI News08/12

Try This Today

One action, right now

Take one recurring workflow and split it into three lanes: cheap extraction, mid-tier drafting, premium judgment. Run the same source material once, record quality and cost, then decide which lane deserves GPT-5.6 Sol during the discount window.