Claude Has Four Models Now, Plus a Dial Nobody Explained
A plain-English breakdown of Haiku, Sonnet, Opus, and Fable, and what that effort slider actually changes.
I write every article on this blog inside Claude. Every draft, every edit, every paragraph I delete at 1am because it read too much like a LinkedIn post, all of it happens in the same chat window. So when that window rearranged itself this week into four different model names and a dial marked effort, I did not feel like a confident daily user. I felt like a first year SOC analyst staring at a SIEM dashboard, watching alerts fire that I did not have names for yet.
If you have opened Claude recently and stared at Haiku, Sonnet, Opus, and Fable wondering which one is actually for you, or what that effort slider does to your bill, this is the explanation I wish someone had handed me. By the end you will know exactly what to pick and why.
Four models is really one triage system with four tiers
Anthropic did not build four separate products so much as it built one system for routing work to the right level of attention, the same way a SOC splits incoming alerts across analyst tiers instead of throwing everything at the most senior person on shift.
Haiku 4.5: the one that never overthinks
This is your Tier 1 analyst. Fast, cheap, and built for the alerts that do not need a second opinion, quick fact lookups, short summaries, cleaning up formatting on something you already wrote. You would not hand Tier 1 a suspected breach, and you should not hand Haiku your hardest problem either. It is not built to sit and think, and it is not trying to.
Sonnet 5: the one doing most of the work right now
Sonnet 5 is Tier 2, and since June 30, 2026 it has owned the daily queue by default. It became the standard model for every Free and Pro user overnight, built for the writing, coding, research, and analysis that fills an actual workday. Early testers had it complete two-part jobs, updating records and sending follow-ups, start to finish without stalling halfway the way earlier versions sometimes did. For most people on most days, whether they know its name or not, this is the model doing the work.
Opus 4.8: the one you call when something does not add up
This is your senior threat hunter. You do not page them for a phishing email that explains itself. You page them when the evidence contradicts itself and someone needs to sit with the problem until it resolves. Opus 4.8 is Anthropic’s flagship reasoning model, built for the hardest coding and multi-step agent work, and it is priced like the specialist it is.
Fable 5: the one you book, not the one on shift
Fable 5 is the outside incident response consultant you do not keep on retainer, because most weeks do not need one. You call Fable in for the genuinely large job, the kind that runs for hours or days mostly unsupervised, plans its own stages, and checks its own output before handing it back. It is Anthropic’s most capable public model, and unlike the other three, it is never selected automatically anywhere. You have to go and choose it.
The effort dial is just asking Claude how long to think
Effort is not a second model picker sitting next to the first one. It is closer to something you already do without noticing.
Someone asks you what two plus two is and you answer instantly. Someone asks you to find the actual root cause of an outage and you sit down and work through it properly. Effort tells Claude which kind of question it is looking at.
Anthropic ships five levels: Low, Medium, High, Extra (developers call it xhigh), and Max. High is the default everywhere, on claude.ai and through the API, which means unless you deliberately change it, Claude is already spending a reasonable amount of thought on whatever you send it.
Mapped onto the same triage instinct: Low is the alert you clear in ten seconds because you have seen the pattern a hundred times before. Medium is the ticket you actually read properly before closing. High, the default, is your normal caseload, thorough enough that you are not quietly cutting corners.
Extra is what you reach for on a long coding session or an agent run with a lot of moving parts, where you would rather Claude overthink than underthink. Max is the incident bridge at 2am, where cost stops being the conversation and getting it right is the only thing left that matters.
Get Vinit Rami’s stories in your inbox
Join Medium for free to get updates from this writer.
Effort does not force Claude to pad a simple answer. Ask something trivial even on Max, and it still answers like it is trivial. The dial changes how much room Claude gives itself for hard questions, not how much it pads easy ones.
What you pay depends on which door you walk through
There are two completely different ways to pay for Claude, and mixing them up is the most common confusion I run into.
If you use claude.ai like everyone else
You are on a subscription. Free costs nothing and puts Sonnet 5 in your hands with daily limits. Pro is 20 dollars a month, or 17 if you pay for a year upfront, and it raises those limits considerably while adding Claude Code and deeper research tools. Max exists in two sizes: 100 dollars a month for five times Pro’s usage, or 200 dollars a month for twenty times it, aimed at people who live inside Claude rather than visiting it a few times a week.
If you are building on the API
None of the above applies. You pay per million tokens, split between what you send in and what Claude sends back, and the four models sit on genuinely different shelves. Haiku 4.5 runs 1 dollar in, 5 dollars out.
Sonnet 5 launched on June 30 at an introductory 2 dollars in, 10 dollars out, a rate that holds until the end of August before settling at 3 and 15. Opus 4.8 sits at 5 and 25, with a Fast mode at double that price for roughly two and a half times the speed. Fable 5 tops the shelf at 10 dollars in, 50 dollars out.
You are not paying for what Claude reads. You are paying for what it writes.
Output almost always costs more than input, sometimes five times more, and that is not arbitrary. Reading a sentence takes far less computation than generating one, word by word, while checking itself as it goes.
Prices and model lineups move fast in this industry. These figures are accurate as of July 2, 2026. If you are reading this months later, check claude.com/pricing before assuming they still hold.
The fun detail: Fable sometimes quietly becomes Opus
Here is the part that made me sit up a little. Fable 5 carries its own safety layer for cybersecurity and biology questions, and when a query trips that layer, Fable does not simply refuse it. It quietly hands the question to Opus 4.8 instead, answers from there, and Anthropic does not charge you Fable’s higher price for that particular reply.
You get a small note telling you the handoff happened. Most people will never see it, because most questions never come near that layer. But if you are the type who tests boundaries out of curiosity, cybersecurity questions especially, it helps to know you are not always talking to the model whose name is on the tab.
It is also not a random precaution. This exact routing system made real news barely two weeks before I wrote this, when Fable briefly went dark worldwide over a government export control order, then returned on July 1 with these same safeguards tightened further. The dial you never actually see is doing more work than most of us assumed.
The rule I actually use
If you remember one thing, make it this. Start on Sonnet 5 and stay there until it visibly struggles. It is the default for a reason, and for nearly everything a normal week throws at you, from writing to research to a stubborn bug, it will not struggle.
Reach for Opus when the task is genuinely hard and getting it wrong is expensive. Save Fable for the rare week you have a problem large enough to hand off entirely and walk away from.
I still catch myself reaching for the most expensive option out of habit, the same instinct that once had me stacking certifications I did not need because more felt safer than fewer. It rarely was. The model is not the hard part anymore. Knowing your own task well enough to choose correctly still is.
So here is my actual question, not the polite kind. Which of the four have you never once tried, and what is stopping you?
