Part 2 of The AI Model Ladder. Part 1 covered ChatGPT’s GPT-6 Luna, Sol, and Astra.
Last week I spent over $250 in two days and didn’t get the thing I sat down to build. I told that story in Part 1. Short version: I ran the top model at the highest reasoning setting for everything, including work a model a fifth the price would have done just as well, and every “fix this one more thing” was another prompt at flagship rates.
Here’s what I’ve come to understand since. That wasn’t a ChatGPT mistake. It was a model-selection mistake, and I could have made it just as easily on the Claude side — easier, actually, because Claude gives you four rungs on the ladder instead of three. More rungs, more ways to pay for the wrong one.
I use both platforms. ChatGPT and Claude work together on most of what I publish. This is the Claude half of the map I built after that bill.
I’m not going to re-explain “one prompt, one shot” in full — Part 1 does that. The short version: you hand the model one complete instruction, and it plans, reads, runs tools, makes changes, tests, and reports back without you managing each step. It is not one API request. It’s one job, one bill.
That framing changes the question. Not “what does a million tokens cost.” Instead: how far does a single prompt go on each model, and what does that single prompt cost?
Think of them as four people you can hire for a job.
Claude Haiku 4.5 is the fast, inexpensive worker. Give it a clear process and organized inputs and it runs. This is Claude’s equivalent of GPT-6 Luna — though at $1 / $5 per million tokens, it’s ten times Luna’s sticker price.
Claude Sonnet 5 is the everyday professional. It investigates, picks between normal options, writes and edits code, builds content, and finishes multi-step work. It sits in Sol’s seat at the same $2 / $10 price. (Verify before you budget: Anthropic announced a move to $3 / $15 on September 1, then reportedly held at $2 / $10. Check the pricing page the day you plan.)
Claude Opus 5.5 is the senior professional. Released September 22, 2026, it’s now the default on Claude’s paid plans. It has no ChatGPT twin — it’s priced between Sol and Astra, and Anthropic says it performs at Fable’s level on most tasks while costing about 40% less to run than the Opus 5 it replaced. Announcement: Claude Opus 5.5.
Claude Fable 5.1 is the top of the ladder for general users — the $10 / $50 model and the exact price twin of GPT-6 Astra. You bring it in when being wrong is expensive. Details: Claude Fable and Mythos 5.1.
Per one million tokens, standard API rates, as of September 24, 2026:
| Model | Input | Output | ChatGPT price peer |
|---|---|---|---|
| Claude Haiku 4.5 | $1.00 | $5.00 | Luna (Luna is 10x cheaper) |
| Claude Sonnet 5 | $2.00 | $10.00 | Sol (same price) |
| Claude Opus 5.5 | $4.00 | $20.00 | None — between Sol and Astra |
| Claude Fable 5.1 | $10.00 | $50.00 | Astra (same price) |
Two things to notice. Claude’s floor is higher than ChatGPT’s. But Claude’s middle is where the value moved: Opus 5.5 is roughly Astra-class judgment at 40% of Astra’s price.
Opus 5 lasted two months. If you built your habits around it, here’s the delta:
Practical translation: everything you ran on Opus 5, run on 5.5 and expect a lower bill for the same result. There’s no reason to stay.
Take the realistic agent workload from Part 1 — about 2.5 million input tokens and 500,000 output tokens across many internal calls. A real job: a site migration, a plugin build, a content-cluster refresh.
| Model | Input cost | Output cost | One-shot task price |
|---|---|---|---|
| Haiku 4.5 | $2.50 | $2.50 | $5 |
| Sonnet 5 | $5.00 | $5.00 | $10 |
| Opus 5.5 | $10.00 | $10.00 | $20 |
| Fable 5.1 | $25.00 | $25.00 | $50 |
That’s the ladder: $5 / $10 / $20 / $50. Same volume of work, four prices for four levels of judgment. ChatGPT’s ladder for the same job is 50¢ / $10 / $50.
Real bills will land somewhere else — caching, batch, retries, and reasoning tokens all move the number, and I’ll get to those. But this is the frame that stopped me from overspending.
Give it a process and a checklist. Don’t ask it to decide how to solve an ambiguous problem.
Its limit isn’t volume. It’s ambiguity.
Your working professional. “Here’s the outcome. Investigate, decide, do it safely, test it, keep going until the acceptance criteria pass.”
Sonnet 5 at medium effort is where I start most days now.
Same shape as Sonnet, but for work where the evidence is messier, the codebase is bigger, or the fix has to be right the first time.
Because 5.5 finishes with fewer tokens, the $20 job sometimes lands under Sonnet’s price. It doesn’t circle.
Reserve this for work where the decision is the expensive part.
With Opus 5.5 matching Fable on most benchmarks, Fable’s honest use case is narrower than it was a month ago: the hardest, most consequential judgment calls, where paying 2.5x for the top of the ladder is cheap insurance.
Start with Haiku when the process is known, the inputs are organized, “done” is a checklist, and mistakes are easy to catch and reverse.
Start with Sonnet when investigation is required, the model must choose among normal solutions, it needs to touch code or production config, and you’d normally hire a competent developer or specialist.
Move to Opus when the codebase is large, the evidence is messy, the job is long-running, or a wrong answer costs you a second full run.
Move to Fable when evidence conflicts, two sensible approaches have already failed, the change risks orders, customer data, payments, or security, rollback is destructive, or the output drives a large investment.
The economical workflow:
Haiku gathers → Sonnet diagnoses and builds → Opus or Fable resolves the hard call → Sonnet finishes → Haiku verifies at scale.
Last week I started at the top and never handed anything down. That’s the whole mistake.
In Part 1 I explained that ChatGPT’s reasoning level doesn’t change the per-token price — it changes how many tokens the model burns thinking. Claude works the same way, and the dial is why you can stop switching models so often.
Sonnet 5, Opus 5.5, and Fable 5.1 expose an effort setting with five levels, and thinking is on by default. The per-token rate is identical at every level. What changes is how much the model reasons before and between actions — which changes your output-token bill and your wait time.
| Effort | Use it for | Cost tendency |
|---|---|---|
| Low | Known edits, formatting, following a checklist | Lowest |
| Medium | Normal features, ordinary bugs, most content work | Moderate (Opus 5.5’s default) |
| High | Hard debugging, multi-step work, migrations | Higher |
| Xhigh | Deep investigation, intermittent bugs, thorough reviews | Much higher |
| Max | The hardest single line of investigation | Potentially highest |
The trick: try turning the dial before switching models. Sonnet 5 at high effort often lands close to Opus-tier quality on agentic work. And the reverse holds: Opus at low effort can be cheaper per finished task than Sonnet at high, because it wastes fewer steps.
Fast mode is a separate control, same as on ChatGPT. It doesn’t make the model smarter; it serves the same model faster. On Opus 5.5 it’s $8 / $40 — double the standard rate — for up to 2.5x speed. Pay for it when the clock matters, not by default.
Subagents are Claude’s version of Ultra. In Claude Code, Opus can orchestrate a team of subagents that split a job into tracks — architecture, security, content, testing — and reconcile the results. Same warning as Ultra: it pays off when the work genuinely splits into independent pieces and burns money when the job is small or sequential.
1. Prompt caching. Agentic work rereads the same context over and over. A cache hit bills at a fraction of the input rate — $0.10 on Haiku, $0.20 on Sonnet 5 and Opus 5.5, $0.25 on Fable 5.1. That last number is why a Fable job can come in far under list price on a long session.
2. Batch processing. Non-urgent work — overnight content generation, bulk classification — runs at 50% off across the lineup. Haiku in batch is $0.50 / $2.50. That’s Luna territory.
3. Tokens per task. Two models at the same per-token price can bill wildly different amounts for the same job, because one thinks longer or circles more. Sonnet 5’s newer tokenizer counts the same text as more tokens than older Sonnets, and its default effort is higher. Opus 5.5 goes the other direction. Price per token is not price per finished task. Measure the second one.
When Opus 5.5 is the right default: almost always now. Anthropic’s own positioning is that 5.5 matches Fable 5.1 on most work, and it’s 60% cheaper per token. For coding, content systems, migrations, plugin builds, and long agentic runs, start here.
Where Fable 5.1 still earns 2.5x:
The second-opinion pattern: let Opus do the work, then hand Fable the one decision that matters — “here’s what Opus concluded and why; find what it missed.” You pay Fable prices for a few thousand tokens of judgment instead of a whole build.
I don’t pick one. I run both, and they’re better as a pair than either is alone. The two ladders line up cleanly — Haiku beside Luna, Sonnet beside Sol, Fable beside Astra — with Opus 5.5 as the odd rung out, which is exactly why it’s worth knowing.
The lesson from both articles is the same. Model choice buys judgment. Effort level buys thinking time. Subagents buy a temporary team. Fast mode buys speed. Match the worker to the job. Stop paying flagship prices for intern work, and stop handing intern-priced models the decisions that cost you money when they’re wrong.
That’s the mistake I made last week. You don’t have to.
Opus 5.5 for complex or long-running work, Sonnet 5 for everyday professional tasks. Anthropic made Opus 5.5 the default on paid plans, and it matches Fable 5.1 on most tasks at 60% less per token.
Only for the hardest judgment calls: conflicting evidence, payments or security, architecture you’ll live with, and pre-production reviews. For most builds, Opus 5.5 delivers Fable-level results at $4 / $20 instead of $10 / $50.
For a realistic agent job of about 2.5M input and 500K output tokens: roughly $5 on Haiku 4.5, $10 on Sonnet 5, $20 on Opus 5.5, and $50 on Fable 5.1 at standard rates. Caching, batch, and effort level move the real number.
No. The per-token rate is identical at low, medium, high, xhigh, and max. Higher effort makes the model reason longer, which uses more output tokens and takes more time. Cost per task rises; price per token doesn’t.
Price dropped 20% to $4 / $20 per million tokens, cache reads dropped 60% to $0.20, output is about 30% faster, and the model uses fewer tokens per task. Default effort is medium.
Subagents in Claude Code. Opus can orchestrate several agents working parallel tracks of one job. Like Ultra, it’s worth it only when the work splits into independent pieces.
Both, if you can. The ladders line up almost model for model, and using each where it’s strongest costs less than forcing one platform to do everything.
Anthropic’s September 22, 2026 Claude Opus 5.5 announcement covers the release, the $4 / $20 pricing, the $0.20 cache-read rate, fast mode, and the comparison with Fable 5.1 and Opus 5. Anthropic — Introducing Claude Opus 5.5
Anthropic’s pricing documentation lists the standard, cache-read, batch, and fast mode rates for Haiku 4.5, Sonnet 5, Opus 5.5, and Fable 5.1, and confirms Sonnet 5’s $2 / $10 pricing. Claude API pricing
Anthropic’s effort documentation explains the five effort levels and each model’s default setting. Claude API — Effort
Anthropic’s Claude Fable and Mythos 5.1 page describes the top of the Claude model lineup. Anthropic — Claude Fable and Mythos 5.1
173 W. Wieuca Road Northeast #104
customersupport@advertisingbusiness.org
(770) 874-6085 
(770) 874-6085
770-874-6085