Which AI Models Actually Matter This Month

There's a new model or feature almost every week, and trying to follow all of it is a great way to feel busy and learn nothing. As an EM, I don't track releases for sport. I run each one through a single question:

Does this change a task I already do, or is it just a bigger number on a benchmark?

If it doesn't change a task, it's news, not signal. But a handful of models have earned a permanent place in how I work — so most of this guide is the practical part: which ones matter right now, when to reach for each, and the small switches that cut your bill without you noticing.

The lineup that actually matters

Anthropic ships a whole family, but for day-to-day work you're really choosing between three, with a fourth for the frontier:

• Haiku 4.5 — the fast, cheap one. Near-frontier quality at the lowest latency and price (about $1 per million input tokens, $5 output). Reach for it on high-volume, low-stakes work.
• Sonnet 5 — the daily driver. Anthropic positions it as the best combination of speed and intelligence, and it's where I start unless I have a reason not to.
• Opus 4.8 — the deep-reasoning specialist, built for complex agentic coding and enterprise work. Anthropic reports it's roughly four times less likely than the previous Opus to let a code flaw slip past unnoticed.
• Fable 5 — the top of the range, aimed at long-running agents. Overkill for most chat, but it's there when you genuinely need the ceiling.

The mistake almost everyone makes is running one model for everything — burning the expensive one on trivial lookups, or asking the cheap one to carry work it was never built for. Fit the model to the task in front of you and you get better answers and a smaller bill.

When to reach for each

Sonnet 5 — the default. Start here when you're not sure. It's fast enough for real-time back-and-forth and strong enough that most tasks never outgrow it:

• Drafting emails, status updates, and docs
• Normal coding — features, fixes, refactors inside a file or two
• Multi-step work where it needs to think and act
• Summarizing a thread, a PR, or a meeting
• Research that needs reasoning, not just lookup

"Draft a Slack update to my team about this morning's incident — what broke, what we did, what's next. Keep it calm and specific."

Opus 4.8 — when it's genuinely hard. Switch up when correctness matters more than latency, or when Sonnet's first pass comes back shallow:

• A refactor that spans several files or systems
• A long research or planning task that has to hold a lot in its head at once
• An architecture call where you want a real second opinion, not a summary
• Debugging something subtle that survived the obvious fixes

"Here's an 800-line service that's grown unmaintainable. Plan how you'd split it into testable pieces first, then walk me through the changes one at a time."

Haiku 4.5 — for volume. Point it at the cheap, repetitive work that doesn't need a heavy model chewing on it:

• Quick lookups and definitions
• Classifying or labeling in bulk
• Pulling specific fields out of a document
• One-shot answers you'd otherwise Google

"Sort these 40 support tickets into 'bug', 'billing', or 'how-to'. Return a table, nothing else."

How to switch (the three-second move)

On claude.ai the switch takes about three seconds, and most people never touch it:

1. Open any chat. The name of the current model sits at the very top (for example, Claude Sonnet 5).
2. Click it — a dropdown lists every model your plan gives you.
3. Pick one. From your next message on, the conversation runs on whichever you chose.

That's the whole skill. The hard part isn't switching; it's remembering that you can.

Turn on extended thinking for the hard ones

For genuinely hard prompts, let the model reason before it answers. On the current lineup — Sonnet 5, Opus 4.8, Fable 5 — this is adaptive: the model decides how much to think, so easy questions stay fast and the hard ones get the full treatment. On Opus 4.8 the effort level even defaults to high. In the API you enable it with thinking: {"type": "adaptive"} and dial depth with an effort parameter; in the Claude apps it's the reasoning toggle tucked into the chat input's tools.

Rule of thumb: leave it on for Opus work — math, multi-step logic, code analysis, research synthesis — and skip it for quick chat, where the reasoning adds latency without adding much.

What each plan actually gets you

Model access is broadly the same across plans now — the real difference is how much you can run:

• Free — genuine access to Claude across web and apps, including extended thinking for complex work. Enough to see whether it fits your workflow.
• Pro (around $17/month billed annually) — a lot more usage, plus Claude Code and Cowork. This is the pricing tier most working devs actually want.
• Max (from $100/month) — 5x or 20x the Pro usage, higher output limits, and priority access when traffic is heavy.

If you're constantly hitting the wall on Pro, that's your signal to move up — not the model list.

Five habits that cut your token bill

These pay off on any model. Small, and they compound:

1. Edit your last message instead of piling on a follow-up. A correction sent as a fresh message drags the wrong answer forward as context. Editing the original forks the thread clean.
2. Batch related asks into one message. Three tasks in one numbered prompt costs far less than three separate turns.
3. End with "just the output, no preamble." Kills the throat-clearing and trims length — and cost — on simple tasks.
4. New topic, new chat. Long threads re-process everything on every turn. Starting fresh when you switch subjects is free and saves a lot.
5. Feed long PDFs as text. PDFs carry formatting overhead that eats tokens; paste the text (or a .md) and you get the same content for less.

The release triage template

When the next model drops, don't spiral. Run it through this — three questions, sixty seconds:

RELEASE TRIAGE — [model / feature name] — [date]

1. Does it change a task I already do?
   (name the task, or write "none")  →

2. What did it actually save?
   (time / fewer passes / a thing I couldn't do before)  →

3. Keep it in my workflow?  →  Y / N
   If Y: the one task I'll use it for this week →

This week

Open the last three "big AI news" items you saved and never acted on. Run each through the one question. I'd bet at least two quietly fail it — and now you can stop feeling behind on things that were never going to change your work. Then open Claude, and the next time a task feels heavy, switch up to Opus before you switch off.

Sources

• Anthropic — Claude Opus 4.8
• Claude Docs — Models overview
• Claude Docs — Extended thinking
• Claude — Pricing plans