Opus 5.5, Sonnet 5.5, and What I Switch Between on Real Projects

By Muhammad Kashif · Draft · AI automation & cybersecurity builder, Pakistan

Opus 5.5 landing near Fable 5.1 power at roughly forty percent cheaper, Sonnet 5.5 on the horizon, OpenAI Sol and Luna at half price, NotebookLM voice tutors across ~100 languages, and law-focused GPT-6 Astra surfaces — my switchboard got more crowded. Amodei talking pace, Musk and Altman agreeing, political rejection of pacing — noise around the dial. On real projects I switch models when the eval says so, not when the timeline says I should feel impressed.

Price cuts are invitations to re-route

When Opus-class quality drops in price, I re-run the arena on my golden sets and move everyday coding up if the numbers hold. When Sol and Luna arrive cheaper, I test them on computer-use and decision-shaped tasks before touching client data. Price is a door. Eval is the key.

Teams that switch on announcement day without tests are performing modernity. Teams that switch on spreadsheet day are engineering.

Not best at everything

Kimi building a better Fall Guys clone, GPT winning on cost, Claude winning on polish — the notes keep me from monogamy. I maintain a per-task default and a documented exception path. Exceptions require a note in the PR.

Monogamy with a model feels like simplicity. It is often just unmeasured preference.

Voice tutors and law surfaces

NotebookLM voice tutors in many languages help training and onboarding. OpenAI for Law on Astra-class models looks like verticalisation. Both still need human professional responsibility. I will not let a voice tutor invent labour policy, and I will not let a law surface file anything without a licensed human.

Language coverage is a gift for Pakistan work when Urdu and English mix. Gift plus review beats gift alone.

One-person billion narratives

Stories about AI-run companies with outsourced doctors and pharmacies are warnings about accountability theatre. If the medical or financial act needs a licence, the agent is assistive. My switchboard never includes "pretend the licence is a prompt."

Founders who want that theatre get a no. Builders who want assistive depth get a routing plan.

Pacing debates vs my calendar

Calls to pace frontier models are policy. My calendar still has client deadlines. I comply with law and contract, keep eval harnesses ready for sudden SKU changes, and refuse to treat Twitter agreement between CEOs as an SLA.

Operational readiness beats opinion cosplay.

How I switch mid-project

Freeze the golden set. Run old default vs candidate blind. Compare defect rates and cost. If candidate wins, update AGENTS.md and pin versions. If mixed, split by task class. Announce the switch to humans so review expectations update.

Silent switches create silent regressions.

What I sell

Switchboard design, golden-set construction, and quarterly re-routing after price cuts. I do not sell eternal Opus maximalism or eternal cheapness. Real projects need a dial with labels.

Sticky note

Opus when wrong is expensive. Sonnet when volume matters. Vertical models with licensed humans. Price cuts trigger tests, not parties.

Blind switch protocol

Candidate models get anonymised labels in the eval sheet so my bias toward Opus or Sonnet cannot steer the pencil. Price cuts trigger the protocol; Twitter heat does not. Luna and Sol candidates get the same blind treatment.

Vertical model + licence clause

Law and health surfaces in the stack require a named licensed human in the SOW. Astra-class tools assist. They do not sign. Sonnet and Opus switches never override that clause.

Announce the switch

Model switches get a Slack or email note to humans who review diffs: what changed, why the eval won, what to watch. Silent Opus-to-Sonnet swaps create silent review debt. Price cuts are not a reason for secrecy.

Exception note in PR

Using a non-default model requires a one-line exception note. Habit beats memory. Opus, Sonnet, Sol, Luna — defaults live in AGENTS.md; exceptions live in the PR where reviewers can see them.

Golden set freeze before switch

Before any Opus or Sonnet switch, freeze the golden set so nobody quietly adds easier cases. Price cuts trigger tests against the frozen set. Luna and Sol candidates face the same freeze. Moving goalposts is how vibes re-enter.