Zkitszo - The Real
October 26, 2026

The Intern, the Senior Dev, the Lead and the Specialist

Edited - October 27, 2026

Whether it's product design, a website that looks like it cost money or a mobile app with the art, the colours and a back end that holds together, every one of these jobs starts in the same place... a model picker with a dozen names on it, and every one of those names comes with its own little "effort" slider running from low through medium, high and xhigh up to max. Top of the list, top of the dial- that's got to be the best result. I've found the higher levels of the same model can actually produce worse results.

So which one is good at what, what does each one cost us, and how fast does each one empty the bucket?

(Dollar figures are Canadian unless they're marked US$- the API prices are listed in US dollars.)

The Staff, Hiring by the Hour

It helps to think of the lineup as staff. Haiku is the intern, fast and cheap and great at the boring stuff, but not the one we'd hand a judgement call to. Sonnet is the senior developer who does most of the real work, Opus is the lead who plans, reviews and makes the calls, and Fable is the outside specialist we bring in when the problem is ugly and everyone else is stuck.

The Staff, and What Each One BillsPrice per token relative to Haiku 5.5 (input and output scale the same) · API list prices in US dollarsHaiku 5.5the intern1× · $0.10 in / $0.50 outSonnet 5.5the senior developer20× · $2.00 in / $10.00 outOpus 5.5the lead40× · $4.00 in / $20.00 outFable 5.1the outside specialist100× · $10.00 in / $50.00 out

Per token, Sonnet 5.5 costs 20 times what Haiku 5.5 does, Opus 5.5 costs 40 times as much and Fable 5.1 costs 100 times as much. Those are API list prices, what a developer pays per million tokens (a "token" being roughly three quarters of a word). The intern's rate comes with a catch, too- it only holds on prompts up to 100,000 tokens. Past that line, Haiku costs five times its headline price.

A subscription doesn't bill in dollars at all, and Anthropic doesn't publish how much faster one model eats through the allowance than another. Their help page puts it this way:

Opus costs several times more per turn than Sonnet, and Sonnet more than Haiku. Spending Opus on routine work is the fastest way to drain a daily limit.

That's "several times more per turn", even though per token Opus 5.5 only costs double what Sonnet 5.5 does. The same help page goes on to say "Sonnet is the right choice for the large majority of coding work"... the people selling the expensive model are telling us to use the cheaper one most of the time.

The Leaderboard Says Otherwise

Arena runs a head-to-head board for building web pages, where two models get the same prompt, two pages come back and people vote for the one they like better (the WebDev board runs on their Code Arena platform these days). The scores work like chess Elo ratings, where a 40-point lead works out to winning roughly 56% of the match-ups, by my arithmetic.

Better at Building Web Pages vs What It CostsArena WebDev rating (up is better) · output price, US$ per 1M tokens, log scale (left is cheaper)167517001725175017751800$10$20$50OUTPUT PRICE PER 1M TOKENS (US$)Opus 5.5 (max) · 1807GPT-6 Astra (max) · 1774Sonnet 5.5 (xhigh) · 1769GPT-6.1 Sol (max) · 1749Fable 5.1 (max) · 1743Sonnet 5.5 (high) · 1709GPT-6 Sol (max) · 1681Gemini 4 Argon (high) · 1674AnthropicOpenAIGoogle

Opus 5.5 sits in first place at 1807, and Sonnet 5.5 on xhigh comes in third, 38 points back, at half the price per token. Fable 5.1, the most expensive Claude and the one built for the hardest problems, lands in fifth, 64 points behind Opus at two and a half times the price. In Elo terms, Fable would win only about 41% of its match-ups against Opus 5.5.

Every Claude and OpenAI dot on that chart is the model cranked up to max or xhigh, and Opus 5.5 on medium or high doesn't show up on the board at all. "Half the price" is also per token rather than per page, because a model on max writes far more thinking tokens for every page it builds. Sonnet's 1769 against GPT-6 Astra's 1774 is close enough to call a tie, and the whole board is made of one-shot web pages judged in a few seconds, so it can't tell us anything about back ends, UX reviews or copy.

Fable's home turf is the long, dig-until-you-find-it work like root-cause hunts, architecture calls and outages. Google's Gemini 4 Argon sits further down the board at Sonnet's price, at least until December 1, 2026, when that price doubles.

Pixel-art boxing match in 8-bit arcade style: a huge heavyweight in gold trunks lands two haymakers on a small boxer in an orange headband, runs out of gas, then gets dropped by quick jabs and an uppercut for the K.O.
Fable vs Sonnet- the underdog takes it.

Who Gets Which Job

Product Design

Opus 5.5 handles the "what are we even building" part- the concepts, the user flows, the list of screens and the trade-offs between them. Medium effort is its default and a fine place to start, since Anthropic's own testing has Opus 5.5 on medium matching or beating the previous Opus on high. When the brief is messy and half-formed, with everything pulling against everything else, that's when it's worth turning it up to high.

A Professional, Modern Website

Opus designs the site and Sonnet builds it. Claude Code even has that split built in: /model opusplan runs Opus while planning and hands over to Sonnet for the typing, and since switching models doesn't clear the conversation, Sonnet still sees everything Opus produced.

Looking "professional" also comes down to keeping the model away from its defaults... the same purple gradients, the same fonts and the same rounded cards that show up on every page on the internet. Design skills that fight that look do more for a site than a bigger model does- covered in Same Same, but Impeccable.

Mobile App- Front End and Back End

On the front end, Opus makes the calls on layout, colours, type and design tokens (the named colours and sizes an app reuses everywhere), and Sonnet writes the screen code in SwiftUI, React Native or Flutter. The back end, meaning the APIs, the database, the login and all the plumbing nobody sees, goes to Sonnet 5.5 on high. When a bug has everyone stumped, where the logs look fine and the app still falls over, that's when Opus earns its rate, and Fable only comes in once Opus has had a real go at it and failed.

Haiku 5.5 gets the grunt work underneath- test scaffolding, sample data, boilerplate and digging through a giant log for the one error that matters. Mind that 100,000-token line with the intern, though, because a long coding session blows past it quickly!

User Experience

UX review goes to Opus 5.5 on high, fed real screenshots of the onboarding, the empty states, the error messages and anything else that makes us ask "why would anyone tap that?"

Copy, Text, the Words

Copy goes to Sonnet on medium, or to Haiku for the bulk stuff like alt text, button labels and ten versions of a headline. What makes copy sound like a person is the voice handed to it- real writing samples, a list of words it must never use and a pass afterwards that takes the robot out.

Pictures

None of these text models can paint a picture, so the icons, illustrations, hero images and splash screen all come from an image model that's billed separately, outside any Claude plan.

  • GPT Image 2.5 from OpenAI is the one for readable words in an image, text-heavy layouts like posters and complicated prompts, at US$30 per million image tokens out.
  • Nano Banana 2.1 from Google is the cheap one for fast iteration, at about 3.4 US cents an image at 1K and 11 cents at 4K, and the Nano Banana family can edit by conversation too ("now make the sky orange").
  • Midjourney V8.2 is for mood and art direction, though it isn't the pick for text-heavy images and has no official public API.
  • Recraft V4.1 puts out actual SVG vectors for icon sets, and Ideogram 4.0 handles text-heavy logos and wordmarks.
  • Qwen-Image has open weights under Apache-2.0, for running on our own hardware.

Pick Two, Compare

Pick a job and two models, and the bars will show what that job costs each one at API list prices, with a line underneath saying how many times over one costs the other. That's roughly how much faster one empties a usage cap as well, so go ahead and pit Fable against Haiku.

Pick Two, Compare
API list prices in US dollars, converted to CA$ at 1.4025 · token counts per job are my estimates

Turning the Effort Dial

Effort is how long the model gets to think before it answers, and more thinking helps when there's a right answer to find, like a bug, a security hole or a calculation. On the web board, both of the models listed at two effort levels did better at the higher one: Sonnet 5.5 gains 60 points going from high to xhigh (about 59% of match-ups against its own high setting), and Opus 5 gains 32 going from high to max.

My experience and the board don't line up there... A side-by-side vote on a single page may well reward the busier result, with more sections, more animation and more "wow", but that isn't necessarily the page anyone wants to ship and maintain. Anthropic's own docs say this about the top of the dial:

max may show diminishing returns and is prone to overthinking, so test before adopting it broadly

The same level name also means a different amount of thinking from one model to the next, because Anthropic calibrates each model on its own. On top of that, the thinking bills as output tokens even when it's hidden, and output costs five times what input does on every one of these models. Slide the dial around and see how the levels compare:

Slide the Effort Dial
What each level is for, who defaults to it, where it shows up on the web board- and how often I actually use it
lowmediumhighxhighmax

When a result comes back bloated or generic, try dropping the effort a level before swapping out the model.

Splitting the Work Between Models

Within the Claude family, splitting the work makes a lot of sense, with Opus taking the decisions, Sonnet doing the typing and Haiku handling the chores, all in one conversation with the same files and skills. Claude Code can even run Haiku or Sonnet as sub-agents (helpers the main session hands a piece of work to) while Opus runs the show.

There's one catch, though- prompt caches belong to a single model. Swap models in the middle of a conversation and the new one has to read the whole thing at full input price the first time around, so it pays to switch at a natural break, like after the plan and before the build.

Across vendors it's rarely worth it, because the new model has to be told everything all over again: the project, the decisions, the style... The two exceptions are images and a second opinion on a design when Claude's taste is stuck in a rut, where one mock from GPT-6 Astra or GPT-6.1 Sol (fourth on the web board at Sonnet's price) can shake things loose.

The Bucket and the Drain

A Claude Max plan comes with three limits- a session limit that resets every five hours, a weekly limit shared across all models and a separate weekly limit just for Fable- and the Claude apps, Claude Code and the IDE plugins all drink from the same pool.

Napkin math, my arithmetic: take a multi-hour website build where the model re-reads the project every turn and call it 2.5 million tokens in and 300,000 out. At API prices with no caching, that comes to US$8 on Sonnet 5.5, US$16 on Opus 5.5 and US$40 on Fable 5.1.

Claude Code caches, though, and a cache hit costs 5% of the input price on Opus 5.5 and Sonnet 5.5, or 2.5% on Fable 5.1. If nine tenths of the input comes back as cache hits (leaving out the cost of writing the cache in the first place), Sonnet drops to about US$3.70, Opus to US$7.50 and Fable to US$18. Output made up US$3 of Sonnet's US$8 before caching and makes up US$3 of the US$3.70 after it.

The fine print on caching is in Cache Me If You Can, and there are more ways to stretch a cap in Usage Limit Reached? Not So Fast..

What the Subscription Actually Saved Me

OpenAI recently put out some bangers of LLMs, Astra being one of them- wholly! One capable model... except... see, I have my Anthropic Max plan, and at present it's a superb offering. A good portion of the tasks I run regularly end with a report back of what the spend would've been on the per-token API route instead of the subscription.

I wish I'd snagged the level up- even more superlative with its multiplier and usage increase. It's been discontinued since, but those who have it can keep it as long as their account is in good standing. When I hit those usage limits... I can't help but kick myself for not going that route.


At the time... looking at the price tiers- my cuurewnt subscription at roughlly $200usd... that next level up subscription, now discontinued, with the caveat that anyone that has it can keep it... I was starring at a monthly $500usd price tag- it's wild that at the time was a big price to pay... now, that $500/ month after learning how much the subscription route has provided value- comparring vs. API token use... the usage multipliers. I kick myself for not sub'ing to that tier.

Many a time the report says the API route would've cost me over $10-15-20... subscription... $0. There's the odd time I pay when the limit's been reached, because I don't want to break my cadence when I'm on a roll- what's an extra $50 thrown into my buffer? Except! It adds up... quick. Viewing those API per-token figures... wowie... this game ain't cheap!

The report sweeps every Claude Code session since my first one on May 21 and prices each call at the public API rates. Up to October 26 that's 26,458 calls, worth US$6,227 at API prices, or CA$8,734 at the Bank of Canada's rate of 1.4025.

What I've actually paid Anthropic since March 20 comes to eight Max plan invoices at CA$313.60 each (CA$280 plus GST and PST) and two one-time top-ups of US$22.40 and US$112, for CA$2,697.30 all in.

All In- March 20 to October 2626,458 Claude Code calls (May 21 on) at API rates vs every dollar paid since March 20 · Canadian dollarsAPI, no cachingCA$44,810API, as cachedCA$8,734Actually paidCA$2,697

That leaves CA$6,036 in my pocket, 69% less than paying per token- or, flipped around, the API would have cost 3.2 times what I paid. Without caching, the same calls would have run CA$44,810!

That 3.2 is the conservative number, too, because the report only starts on May 21 while the March and April invoices and the first top-up paid for months it never sees.



Going the subscription route was 69% less than the token way- token aka API usage- I repeat, had I paid per use, per token, API... his would have cost me over $44,000- $44,000 dollars! Over! With the subscription I spend 3.2x less.



Every Billing Month- API Price vs What I PaidClaude Code calls priced at API rates vs the Max plan invoice (plus top-ups) · Canadian dollarsAPI-equivalentactually paidJun 22 – Jul 22CA$1,279 (4.1×)CA$314Jul 22 – Aug 22CA$2,364 (7.5×)CA$314Aug 22 – Sep 22CA$3,305 (7.0×)CA$471Sep 22 – Oct 22CA$1,566 (5.0×)CA$314

Every full billing month since late June, the API price landed somewhere between four and seven and a half times the CA$313.60 invoice. The busiest one, August 22 to September 22, ran CA$3,305 worth of API calls against CA$470.68 paid with the top-up included, and Fable 5 and 5.1 alone make up CA$5,566 of the total, 64% of it.

Setting Fable Aside

The truth is, Fable- the latest- is such a heavy hitter. So capable, effortlessly producing work that talented, skilled people, even whole teams, would have spent countless hours on- the learning, the screen time, the schooling... the trailblazers of their craft, the ones who set the trends rather than follow them. Fable rivals that level of work, that quality of product.

Still, I have to set it aside- maybe use it to scrutinize a lower-tier model's work. Opus 5.5 on high is my #1, Sonnet 5.5 on xhigh my #2. I just get crushed when I hit a wall with my usage, and with Fable that's a common occurrence. It destroys the flow, I lose sight of key items... then getting back into things, just recalling where we'd left off burns through most of the usage quota by itself. Opus 5.5 on high- a capable model- and watching that little usage wheel barely budge a sliver... mighty pleasant.

My report backs that up. Averaged over every main-session call, a Fable 5.1 turn ran CA$0.77 at API prices while an Opus 5.5 turn ran CA$0.23, under a third of that, and Fable 5.1 writes more per turn as well- 2,806 output tokens on average against Opus's 1,277. The first call of a resumed session, which reloads a median of 85,000 to 112,000 tokens of earlier conversation, averaged CA$1.44 on Fable 5.1 and CA$0.96 on Opus 5.5. Give the wheel a spin and watch how fast it fills:

The Usage Wheel
Average cost of one main-session turn, and of the first turn of a resumed session, from my own Claude Code calls · the wheel is a CA$10 slice, for scale
0%

The OpenAI Side- Astra and Codex

Fanboy praise aside- I also have a subscription with OpenAI. For someone just starting to get serious with LLM work, the pricing is so good for what the different ChatGPT models can do, how well they play with Codex and the results that prove it. Well worth it- with- a- caveat... word to the wise...

I've seen incredible results... wicked games- the graphics, the style, one-shot results that run for some time, sometimes. Just like Fable when it first came out, pumping out the hits- the jams! (IDK- for me, Fable has regressed since... Opus has been producing the hits these days.) Astra too. Team Astra up with Codex and you've got yourself an incredibly capable, exquisite model- I'm blown away almost every time. Coding, design, planning... research I delegate to more cost-effective models, verifying their results.. all of it- so good.

What we don't see behind those games reviewed on YouTube and the one-shot mobile apps is the serious token usage that went into them... many tokens equals many dollars. Astra has been a thorn in my behind on exactly that (I've had it happen on Fable too). The last handful of times I started Codex to get cracking on a project with "let's continue where we left off"... I didn't even get to start. Astra burned through the tokens, and before it had even recollected our last actions I got the dreaded message: usage used up, come back in 5 hours.. I'm going to have to drop down a model with OpenAI.

Starting Where We Left Off- Without the Bill

There's a lesson to be had there. Astra is capable, but in times of low usage capacity it may not be the practical pick.

Resuming the old session with codex resume in Codex, or claude --continue and --resume in Claude Code, reloads the whole earlier conversation before any new work happens, and my report puts that at 85,000 to 112,000 tokens every time we come back. There are a few ways around it:

  • Before calling it a day, have the model write a short handoff file that says what's done, where things live, what failed and what the next step is- a PROGRESS.md in the project, a page or two at most.
  • Come back in a fresh session (/new in Codex, or a new claude session) and tell it to read PROGRESS.md and carry on from the next step- that costs a couple of thousand tokens instead of a hundred thousand.
  • Put the standing rules in the file each tool reads on its own at startup, AGENTS.md for Codex and CLAUDE.md for Claude Code, so the project never has to be explained twice.
  • If the old session really has to be resumed, run /compact first (both tools have it) so the long history gets swapped for a summary, and only that summary gets reloaded.
  • Do the comeback on a cheaper model, since reading a handoff file and picking up the thread doesn't need Astra or Fable.

Price Tags, Side by Side

These are per million tokens, standard tier, in US dollars as of late October 2026. Anthropic and OpenAI match at US$50 out for the top tier and 50 US cents for the smallest, and all three vendors meet at US$10 out in the middle.

Model               Vendor      Input    Output
Claude Fable 5.1    Anthropic   $10.00   $50.00
GPT-6 Astra         OpenAI      $10.00   $50.00
Claude Opus 5.5     Anthropic    $4.00   $20.00
Claude Sonnet 5.5   Anthropic    $2.00   $10.00
GPT-6.1 Sol         OpenAI       $2.00   $10.00
Gemini 4 Argon      Google       $2.00   $10.00   ($4 / $20 from Dec 1, 2026)
Gemini 3.8 Flash    Google       $0.75    $3.75   ($1.50 / $7.50 from Jan 1, 2027)
Claude Haiku 5.5    Anthropic    $0.10    $0.50   (prompts up to 100K tokens)
GPT-6 Luna          OpenAI       $0.10    $0.50

GPT-6 Astra also charges more once the context passes 272,000 tokens.

Which model does your heavy lifting these days, and have you ever turned the effort up only to get something worse back? Leave a comment- I'd love to compare notes.

https://docs.claude.com/en/docs/about-claude/pricing
https://code.claude.com/docs/en/model-config
https://www.benchleader.com/benchmarks/lmarena_webdev

Glossary

  • LLM — large language model; the text-generating AI behind Claude, GPT and Gemini.
  • Token — the unit these models read and write in, roughly three quarters of an English word.
  • Input / output tokens — what the model reads (the prompt, the files, the conversation so far) versus what it writes back, including hidden thinking.
  • Effort — a setting that controls how much the model thinks before answering: low, medium, high, xhigh, max.
  • Thinking tokens — the model's internal reasoning; billed as output even when it isn't shown.
  • Prompt caching — reusing text the model already read, billed at a small fraction of the normal input price.
  • API-equivalent — what a subscription's usage would have cost at pay-per-token API prices.
  • Code Arena / WebDev board — Arena's crowd-voted leaderboard where models build web pages head to head.
  • Elo rating — a chess-style score from head-to-head wins and losses; higher is better.
  • opusplan — a Claude Code model setting that uses Opus while planning and Sonnet while executing.
  • Sub-agent — a helper model the main session hands a piece of work to.
  • Handoff file — a short note the model writes before a session ends, so the next session can pick up without re-reading everything.
  • /compact — a command in Claude Code and Codex that replaces a long conversation with a summary.
  • Codex — OpenAI's coding agent, run from the terminal or an app.
  • Max plan — Anthropic's higher-tier Claude subscription, with five-hour and weekly usage limits.

Sources

Comments