The Best Claude Model for Every Job (2026)

The best Claude model isn't the smartest one you can reach, it's the one that fits the job in front of you. Anthropic now runs four production models, Haiku 4.5, Sonnet 5, Opus 5 and Fable 5, and in 2026 they're a toolkit, not a ladder to climb. Point the frontier model at a quick summary and you've overpaid many times over; point a cheap model at a days-long build and it falls apart. This guide explains which Claude model to use for what, what each costs, and the one rule that saves the most money.

The old strategy of picking the flagship and routing everything through it is finished. The models are now priced and tuned for a world of agents doing multi-step work, where the smart move is matching the model to the difficulty of the task, not reaching for maximum horsepower every time. Get that right and you can cut your costs by more than half with no real drop in quality.

Watch: Our walkthrough of eight powerful use cases for Fable 5—and if you want all the prompts, skills and resources featured in the video, join our free community today and never fall behind the curve.

The 2026 Claude line-up at a glance

Those are API pay-per-token prices; if you're on a subscription, the maths is different, and we cover that below. The short version of the table: each step up roughly doubles the cost, so you only climb when the job truly needs it.

The one rule: plan with the strong model, execute with the cheap one

If you take one thing from this, take this. The expensive models are for the hard thinking, planning a project, breaking it into steps, checking the work, untangling something ambiguous. The cheaper models do the high-volume execution once the thinking is done. Using Fable 5 to write a summary or format a document isn't just wasteful, it's a design mistake.

"You have to be deliberate about what earns that spend."

— Jan, 9x co-founder

The sharpest version of this is a workflow you can copy. Use the top model to figure the hard thing out once, then run the finished thing on a cheaper model.

"Fable 5 is the model you want when there's a lot to figure out over a long-running task... Now that it's built, everything is figured out... So you can run it on any model in your subscription. I'd start with Opus, and definitely test whether Sonnet can handle it too."

— Jan, 9x co-founder

That single habit, do the expensive thinking once, run the cheap execution many times, is where most of the savings live.

Model by model: which Claude to reach for

Haiku 4.5, the fast, cheap workhorse

Haiku 4.5 is the cheapest model in the line-up and built for scale and speed. It handles a 200k-token context and is the right choice for high-volume, low-complexity jobs: classifying support tickets, extracting fields from documents, tidying formatting, running simple checks. It isn't for deep reasoning or production-grade code, and it lacks the reasoning controls the bigger models have. But when you're doing the same small job thousands of times and speed and cost matter more than nuance, nothing else comes close.

Sonnet 5, the default for real work

Sonnet 5 is the workhorse, and for most people, most of the time, it's the right answer. Released on 30 June, it's the default for Free and Pro users and comes close to flagship performance at a fraction of the cost, with a 1-million-token context window. It's the model doing the bulk of everyday coding, drafting, and agentic work in Claude Cowork and Claude Code. At its standard rate it's around 40% cheaper than Opus per token, which makes it the natural engine for anything that involves generating a lot of text or code.

One thing to watch, covered in full further down: Sonnet 5 thinks harder than older models by default, and pushing that effort to the maximum can quietly make it cost more than Opus.

Opus 5, the new flagship

Opus 5, launched in late July, is the most consequential release of the year for how you should choose. It delivers near-frontier intelligence while keeping the old Opus price of $5 / $25 per million tokens, so its capability jumped without its cost moving. It holds a 1-million-token context, has the freshest knowledge of the line-up, and tops Anthropic's standard coding benchmarks. The practical effect is huge: Opus 5 absorbed most of the jobs that used to need Fable 5, at half the price. For "maximum horsepower" on a hard problem, Opus 5 is now what you reach for, and only on the very hardest, longest work does the frontier model still earn its premium.

Fable 5, the ceiling

Fable 5 sits at the top and costs exactly double Opus 5: $10 / $50 per million tokens. It's built for days-long, highly autonomous jobs where errors compound across long chains, the kind of work where the cost of the API call is trivial next to the cost of the human labour it replaces: a huge legacy code migration, deep multi-source analysis, a system built end to end from one brief. We walk through eight jobs where it earns that cost in Claude Fable 5 use cases. It holds state across enormous context, plans across stages, checks its own intermediate work, and can orchestrate other models. Because it's so capable and sensitive, Fable 5 was briefly pulled from the market under a US export-control order days after its June launch and restored on 1 July, a first for a commercial frontier model.

If you're going to use it, run it inside an agent, not a chat window.

"If you are going to use Fable 5, do it in Claude Cowork or Claude Code, not in the regular Claude chat. You're just going to be using way too much usage for what you get back."

— Jan, 9x co-founder

The cost decision: subscription vs pay-per-use

How you pay changes the calculation entirely.

On a subscription (Pro, Max, Team), you're spending against weekly limits, not per token. On a Max plan, Opus 5 is the default and draws normally, so there's no reason to ration it in favour of Sonnet, use the best model your plan gives you. Fable 5 is the exception: it carries double the quota weight of an Opus 5 query and is capped at 50% of your weekly usage, after which you're switched back to Opus 5 until the week resets. On Pro, Fable 5 isn't included at all and has to be funded with pay-as-you-go credits.

On the API (pay-per-token), the table above is the whole story: Haiku for volume, Sonnet for execution, Opus for the hard reasoning, and Fable only when a job is too big and too ambiguous for anything else.

The trap worth knowing: "max effort" can cost more than the flagship

Here's the counterintuitive one that catches teams out. Sonnet 5 reasons before it answers by default, and you can dial that effort up. Crank it to the maximum to try to squeeze flagship-level intelligence out of the cheaper model, and it writes long, verbose internal reasoning for even simple prompts, and the output cost balloons. In testing, Sonnet 5 at maximum effort has come out more expensive per task than Opus 5, which reaches the right answer efficiently because its baseline intelligence is higher.

The rule: keep Sonnet's effort low or standard for most work, and if a task really needs maximum effort, don't crank Sonnet, escalate to Opus 5. You get better reasoning at a more predictable cost.

For developers: two routing patterns that cut the bill

If you're building agentic pipelines, two patterns get most of the frontier's quality for a fraction of the spend.

The advisor pattern suits focused coding work. Sonnet 5 runs the main loop, reading files, writing patches, running commands, and only calls Fable 5 as an advisor when it hits something ambiguous it can't resolve. Fable returns a short strategic instruction, not reams of code, so the expensive model barely generates any tokens. Anthropic reports this reaches roughly 92% of Fable's solo quality at about 63% of the cost.

The orchestrator pattern suits broad research and analysis. Fable 5 is the brain: it takes the big goal, breaks it into sub-tasks, and dispatches them to a swarm of Sonnet or Haiku workers that do the token-heavy reading and return compressed summaries. That reaches around 96% of Fable's solo quality at roughly 46% of the cost.

The other big lever is prompt caching, which can cut input costs by up to 90% on repeated context. Put your stable content, tool definitions, system instructions, knowledge, at the very start of the prompt so the cache holds, and keep variable user input after it.

How to choose, quickly

• A quick, high-volume, repetitive job (classify, extract, format) → Haiku 4.5

• Most real work (coding, agents, drafting, execution) → Sonnet 5

• Hard reasoning (architecture, tricky multi-file bugs, novel problems) → Opus 5

• A huge, multi-day, high-stakes autonomous build → Fable 5, and even then, plan with it and execute on something cheaper

A living guide

The models change every few months, so treat the version numbers as a snapshot, not scripture. Opus 5 replaced Opus 4.8 without changing price; Sonnet became the default; the frontier keeps moving. We track the specifics in our running guide to what's changed in Claude. The durable skill isn't memorising today's line-up, it's the principle: match the model to the job, plan with the strong one, execute with the cheap one. That's the same instinct behind knowing where you sit on the AI skills ladder.

The bottom line

There is no single best Claude model, there's a best model for each job. Haiku 4.5 is cheapest and fastest for high-volume work; Sonnet 5 is the everyday default and the best value for real execution; Opus 5 is the flagship for hard reasoning and now handles most of what used to need the frontier; Fable 5 is the ceiling, worth its premium only on the biggest, longest, most autonomous jobs. Match the model to the difficulty, use the strong ones to plan and the cheap ones to execute, and you get frontier results without frontier bills.

If you want your team fluent in this, choosing the right model, running the big ones inside Cowork and Code, and building the workflows that do the work, our hands-on Claude training for teams takes people from their first task to a working system.

Common questions

What is the best Claude model? It depends on the job. Opus 5 is the best all-round flagship for hard reasoning, Sonnet 5 is the best default for everyday work and value, Haiku 4.5 is cheapest for high-volume tasks, and Fable 5 is the ceiling for multi-day autonomous jobs. Matching the model to the task beats always reaching for the smartest.

Is Claude Fable 5 worth it? For most people, rarely. Opus 5 delivers near-frontier intelligence at half the price and has absorbed most of the jobs Fable 5 used to be needed for. Fable 5 earns its cost only on the biggest, multi-day, high-stakes autonomous work, and even then you plan with it and execute on a cheaper model.

Fable 5 vs Opus 5: which should I use? Opus 5 for almost everything hard, it's near-frontier at half the cost. Fable 5 only when a single job is too large and too ambiguous for Opus to hold, such as a multi-day migration or an end-to-end build from one brief.

Which Claude model is cheapest? Haiku 4.5, at $1 / $5 per million tokens, roughly a fifth of Opus 5's cost. For everyday work that still needs to be good, Sonnet 5 is the best value.

Which Claude model should I use for coding? Sonnet 5 for day-to-day coding and feature work, it's fast and cost-efficient. Escalate to Opus 5 for complex, multi-file architectural changes and hard bugs.