Mirza Ishraq
← All issues

Issue #2 · By · · 10 min read

Sonnet 5.5, One Week In

This week I deleted a feature from Tradion. Chart analysis let you upload a chart screenshot and get an AI analysis back. You can paste a chart into ChatGPT, or any AI model out there, and get the same thing. If ChatGPT can do it, why would anyone use Tradion for it?

It felt like a GPT wrapper, so it’s gone. A GPT wrapper with a price tag is just ChatGPT with a billing page. Lens v2 replaced it. It reads the chart in real time and streams its analysis while the market moves, on whatever time frame you pick.

Inside, Sonnet 5.5 after a week of building on it, GPT-6.1 Sol, a compaction trick that made my live analyst 36% cheaper in a replay, an iced coffee video prompt, five tools, and a leaked key.

Talk of the Town

Sonnet 5.5 writes all my AI-assisted code now. It’s cheaper per task than Sonnet 5, much more accurate, and I need less time reviewing what it writes. I still review it, though.

Quick background. Anthropic released Claude Sonnet 5.5 on September 28. It’s the middle model in the lineup, under Opus.

Per token it costs the same as Sonnet 5, $2 per million tokens in and $10 per million out (a token is a small chunk of text, a word or part of one). Anthropic says it runs 30% or more faster and costs up to 30% less for most work. On Terminal-Bench 4.0, a multi-step coding test, it scored 70.6% to Sonnet 5’s 10.3%.

The launch page hit 884 points on Hacker News. On Reddit the love sounds like Opus 5.5 quality with half price, and the math holds, since Opus 5.5 costs $4 in and $20 out.

The loudest pushback was a thread where Sonnet cost 30x more than GPT-6.1 Sol on the same 3D animation, $57.34 to $1.78 at API prices. Most commenters called it a setup problem. Both ran on ultracode, the heaviest setting, and Sonnet split the job across 19 agents to Sol’s 4. Several thought Sonnet’s version looked more detailed.

Two days after launch, Anthropic also deprecated Sonnet 4.5, which stops working on the API on November 30.

If you build on the API, upgrading is more than a new model name. Anthropic’s what’s new page lists five breaking changes, things that worked on Sonnet 5 and now fail or behave differently. Three of them will hit most people.

If your code switches Claude’s thinking off, Sonnet 5.5 rejects it and points you to a new lowest setting called between_tools. If your code forces Claude to call one specific tool, that’s rejected too.

And the short notes Claude writes between tool calls now arrive as thinking instead of normal text, so an app that shows them to users just goes quiet. No error, the politest way software has ever broken.

I moved every Sonnet 5 call in Tradion over on September 29, and nothing broke. Anthropic put out a solid migration guide, and following it and their API docs was super helpful.

Then I built Lens v2’s live analyst on it, the part of Tradion that comments on every candle as the market moves. Before launch, I spent $49.30 replaying two recorded trading days through it to see what it would cost.

It comes to about $1.53 for every hour the analyst is switched on, on a 1-minute chart. The effort setting, how hard the model thinks before it answers, barely moved the price. Low, medium and high all landed between $1.53 and $1.58 an hour on the same day, so mine runs on high.

So here’s where I land. If you know how to code and understand what’s going on, use Opus for planning and Sonnet 5.5 for execution, even if you still have to type a command or review the code yourself.

This Week in AI

GPT-6.1 Sol

A day after Sonnet 5.5, OpenAI released GPT-6.1 Sol. Astra is OpenAI’s top model, and AWS, which added Sol to Bedrock the same day, says Sol approaches GPT-6 Astra across demanding evaluations at roughly one-fifth of the cost. On the API it costs $2 per million tokens in and $10 out, the same as Sonnet 5.5. So the two cheap-but-smart models now charge the same for tokens in and out.

I’ve barely used it. Codex runs my reviews on Astra, and the one time Claude handed Sol a task, Codex refused it as not supported on my ChatGPT account yet. Try it once it shows up in your model picker.

→ Introducing GPT-6.1 Sol

Claude Code v2.1.287

If you use Claude Code, update. Version 2.1.287 (October 1) added Claude Mods, which let plugins change deeper parts of how Claude Code works, plus a built-in one called You should know. That’s a side agent that watches your session and flags things you or Claude might miss. Turn it on with /plugin enable cc-plugin-you-should-know@builtin (it needs telemetry on).

The fix I care about is quieter. In my setup Opus plans and Sonnet builds, a mode called opusplan, so the model switches back and forth all day, and the switch could drop the earlier thinking. Update for that fix alone. Mods and the side agent I haven’t tried yet.

→ v2.1.287 release notes

Anthropic Python SDK v1.9.0

This one is for anyone calling Claude from Python. The SDK is Anthropic’s official library for talking to the API, and v1.9.0 (September 28) adds Sonnet 5.5 and that new between_tools setting. It also lets your code start running a tool while Claude’s reply is still streaming in, instead of waiting for the whole reply. Cache diagnostics are out of beta too.

Running tool calls while the reply still streams is the one I’d use. My live analyst waits on its tools, with a 30-second deadline on each. Upgrade if you build on the API.

→ v1.9.0 release notes

Workflows & Hacks (Steal This!)

Steal #1. Compact your agent early

What it is. Every time your app sends Claude a message, it sends the whole conversation along with it. So the longer a chat runs, the more every new message costs. Compaction means asking Claude to summarize the conversation so far, then carrying on from that summary instead of the full history. Anthropic has it built in now as on-demand compaction (in beta), and you decide when it happens.

When is the whole trick. My live analyst comments on every candle, so its conversation grows all day. I replayed the same trading morning twice.

Summarizing once the conversation passed 120,000 tokens cost $1.65 per live hour. Summarizing at 60,000 cost $1.06. That’s 36% cheaper, and the comments after each summary still named the same price levels and volumes.

How to use it. Nothing to install beyond a current Anthropic SDK. Anthropic’s compaction docs have a ready-made loop you can copy. After every reply, it adds up the tokens the last request used. Once that passes your limit, it asks for a summary and swaps it in for the old messages.

Pick your limit by testing, not guessing. Mine moved from 120K to 60K only because I replayed real sessions both ways.

Watch out for. Check that a summary actually came back. In my first full-day replays, the summary request failed without any error, the later turns weren’t usable, and $12.16 of test budget went nowhere. Sending the summary request as a stream fixed it.

Steal #2. A 30-second recipe explainer video

What it is. The fun one. One prompt gets Claude to build a 30-second recipe explainer animation, each ingredient and its measurement going into the cup, from empty to the finished drink. Mine is an iced coffee.

How to use it. Paste it into Claude, swap in your own drink, and see what you get.

The artifact.

We're going to try a little test. Do you think you could render a recipe
motion graphic animation using javascript or html (w/e you think will
produce the best) to show the full recipe from start to finish (empty cup to
finished iced coffee) - Explainer video style - Showing the recipe
ingredients + measurements as they're going into the cup. Recipe: iced
coffee in a tall cup, brown sugar, blonde espresso, light ice, no whipped
cream. Should be a 30s video.

Watch out for. The recipe has no amounts, so Claude picks the measurements. Add yours if they matter.

Builder Radar

  • LiveNerf reruns GPQA Diamond, MMLU-Pro and competition math questions on Opus 5.5 every day for 30 days to see if it gets worse. Its own repo says the first possible call is around October 24, so every “it got nerfed” post this week is early.
  • Offrun is a free Mac app that runs Claude Code, Codex, Antigravity and Grok Build side by side, with usage and reset time per account. Handy if your week is split across two plans’ limits, like mine.
  • Weave Router is open source routing for coding agents. It sends simple work to cheap models like DeepSeek V4 Flash and saves Astra for hard debugging.
  • ds4 is antirez’s C inference engine for high-memory Macs, CUDA and ROCm. It runs DeepSeek V4, GLM 5.x and Qwen locally, MIT licensed.
  • Ledge is a free, open source Markdown notebook that runs shell commands, code, SQL and AI prompts inside your notes. For everyone whose runbook is a wiki page they copy-paste from.

Fix it before you ship

Your host doesn’t know your key is a secret

This week a secret key leaked from one of my Netlify sites. Netlify didn’t catch it. The API provider did, with errors and logs. The fix was one checkbox.

Your keys live in environment variables, settings your host keeps outside your code. Netlify only scans your build for a variable’s value once you mark it Contains secret values. Smart detection of unmarked keys is on Personal, Pro and Enterprise plans.

And leaked keys get used fast. In one honeypot test, attackers used a key leaked on GitHub within a minute.

  1. Make a new key at the provider and retire the old one.
  2. In Netlify, open Project configuration, then Environment variables, edit the variable and select Contains secret values. Give it only the scopes that need it.
  3. Redeploy. If secret scanning fails the build, the deploy log shows where the value turned up.

Study how a platform handles secrets before you hand it one. Then ask your AI tool to check the rest.

Audit this app for leaked secrets before launch. Report each issue as: issue
| file | risk | exact fix. Change nothing until I approve.
Check that no secret, service-role, OpenAI, Stripe or AWS keys are in client
code, build output or git history, and that every secret in my hosting
provider's environment variables is marked as secret.

App Ideas

Pretty sure it exists, but here’s the app I wished for this week. It scans viral posts from Reddit, X and the rest, then plans my week for me. Monday morning it hands me post ideas with hooks, a posting calendar, and suggestions based on what’s trending where.

It’s for anyone who posts every week and does the trend research by scrolling.

How I’d build it with Claude. The scanning half is what my Saturday newsletter scan already does. Plain Python pulls Reddit and Hacker News and scores trends by counting, and Claude only writes the summary. The planner is the missing half.

Interview & Sketch prompt.

I want to design a content planning app that scans viral posts on Reddit, X
and other platforms, finds what's trending on each one, and plans my posting
week automatically: post ideas with hooks, a posting calendar, and
suggestions for which trend fits which platform. Before sketching anything,
interview me one question at a time about my own workflow: which platforms I
post on, how often, my niche, how I write hooks today, and what "viral"
should mean for me. Keep going until you can design around my real habits,
not a generic creator. Then sketch an architecture: which sources to pull and
how (official APIs, RSS, rate limits), how to score a trend with code before
any model sees it, where Claude writes hooks and plans the calendar, and what
I review before anything is scheduled. Give me the answer as a numbered list
of implementation phases.

Demo/Mockup prompt.

Build a dummy-data prototype called "Weekplan", a content planning app that
turns this week's viral posts into a posting plan.

## Design & Theme
- Light theme, warm off-white background, near-black text
- One accent color for "trending now" and nothing else colorful
- Clean sans-serif UI font, monospace for counts and dates
- Minimal, calm, editorial. No gradients, no glow, no emoji, no rounded
  bubbly cards
- Desktop-first, works on a laptop screen at minimum

## Pages

### 1. This Week (/)
- Top bar: week picker (Monday to Sunday), platform filter (Reddit, X,
  LinkedIn, YouTube), "Regenerate plan" button
- A 7-day calendar grid, one column per day, each slot a planned post card:
  platform badge, hook (one line), format (thread, short video, carousel),
  suggested time
- A right-hand panel: "Trending where" with the top 5 trends per platform,
  each with a score and a sparkline of the last 7 days

### 2. Ideas (/ideas)
- A list of 12 post ideas, each with 3 hook options, the trend it rides,
  the source posts it came from (title, platform, rank), and a
  "Add to calendar" button

### 3. Trend Detail (/trend/:id)
- The trend name, its score per platform, the top 10 source posts with
  title, platform and rank, and why it's trending in two sentences
- Suggested angles: 3 cards, each with a hook and a platform

### 4. Settings (/settings)
- Platforms I post on, posts per week per platform, niche keywords,
  topics to avoid, posting time windows

## Important Notes
- Prototype with dummy data only, no real backend, no real API calls
- All data hardcoded to look like a real week (realistic titles, scores,
  dates)
- Make navigation between the four pages work
- Make "what to post, where, and when" obvious at a glance

Build it? Hit reply.

Build & Experiment Log

  • Building. Lens v2 for Tradion, a real-time AI chart analyst that streams while the market moves, plus a new landing page, new positioning and Stripe endpoints for the new pricing.
  • Tried. Explainer videos. Pretty slick looking, and several are finished.
  • Broke. My token budget, on commands Claude didn’t need to run. I run some myself now, and I’m still figuring out how to enforce it.
  • Learned. Rust feels very dynamic for talking to embedded systems. I could be wrong.
  • Next week. Rerunning my plan-adversary evals to see if Opus 5.5 got worse.

From the Blog

P.S. Forward this to a friend who builds. Read it on the web at Sonnet 5.5, One Week In.

Thank you for reading! Follow me on X @myspaceio.

Hit reply and tell me, what are you building?

More issues