Why Anthropic Offers Four Different Claude Models
Anthropic does not ship one Claude.
It ships four: Claude Fable 5, Claude Opus 4.8, Claude Sonnet 5, and Claude Haiku 4.5.
Each one is a genuinely different model, not a marketing label slapped on the same weights.
Understanding why a single AI lab maintains four separate tiers - rather than one model everyone uses - is the fastest way to stop overpaying for simple work or underpowering hard work.
Summary
- Core Idea: Anthropic segments Claude into tiers because no single model is simultaneously the cheapest, the fastest, and the most capable - those three properties trade off against each other.
- Why It Matters: Picking the wrong tier either wastes budget on a task that didn't need deep reasoning, or produces weak results on a task that did.
- Key Concepts: model tier, context window, latency, cost per million tokens (MTok), adaptive thinking, default model.
- When to Use: Any time you're choosing which Claude model to call for a task, set as an org default, or budget for at scale.
- Limitations / Trade-offs: More tiers means more decisions - teams that never think about model choice tend to either overspend on Opus/Fable for everything or underuse Sonnet's default capability.
- Related Topics: context windows and tokens, prompting Claude, team and enterprise administration.
Foundations
Every large language model has to balance three things that pull against each other: how capable it is, how fast it responds, and how much it costs to run.
A model that reasons through a genuinely hard problem takes longer and costs more compute than a model giving a quick, confident answer to a simple question.
Anthropic's answer to that tension is not to pick one point on the curve.
It's to offer four fixed points and let the person or the application choose which one fits the task in front of them.
As of roughly June 2026, those four points are, from lightest to heaviest:
Claude Haiku 4.5, the fastest and cheapest tier, priced around $1 input and $5 output per MTok, with a 200K token context window.
Claude Sonnet 5, the balanced, everyday tier and the current default model for Free and Pro plans, released June 30, 2026, with introductory pricing around $2 input and $10 output per MTok.
Claude Opus 4.8, the flagship reasoning model, with a 1M token context window by default, priced around $5 input and $25 output per MTok.
Claude Fable 5, the top Mythos-class tier, Anthropic's most capable widely released model, with a 1M token context window, a 128K token maximum output, and always-on adaptive thinking, priced around $10 input and $50 output per MTok.
This is not a new idea for Anthropic.
Since the Claude 3 family launched in March 2024, Claude has shipped in Opus, Sonnet, and Haiku size tiers.
What's new as of June 2026 is Fable, a Mythos-class tier that sits above Opus for the small slice of work that genuinely needs the most capability and the largest context Anthropic offers.
A useful analogy: think of the four tiers less like "better" and "worse" versions of the same product, and more like different vehicles for different jobs.
You wouldn't rent a moving truck to run to the corner store, and you wouldn't try to move a house in a compact car.
Mechanics & Interactions
The tiering logic rests on one underlying trade-off: capability and reasoning depth versus speed and cost.
A bigger, more capable model generally has to do more internal work to produce an answer - more parameters activated, more careful reasoning, sometimes explicit "thinking" steps before it responds.
That extra work is what makes it better at ambiguous, multi-step, or high-stakes problems.
It's also exactly what makes it slower and more expensive per token than a model built to answer quickly.
Haiku sits at one end of that curve: optimized for speed and volume, at the cost of some raw reasoning depth on the hardest problems.
Fable sits at the other end: optimized for maximum capability and context, at the cost of being the slowest and most expensive option to run.
Sonnet and Opus occupy the middle, with Sonnet 5 tuned as the practical daily driver and Opus 4.8 tuned as the model you reach for when a task genuinely needs deep, extended reasoning.
Context window size follows a related but distinct logic.
A larger context window (Fable and Opus both default to 1M tokens, versus Haiku's 200K) lets a model hold more of a long document, codebase, or conversation history in view at once.
That's valuable for research, large-codebase work, or long threads, but it isn't free - larger context tends to correlate with the more expensive tiers because processing more tokens costs more compute regardless of how "smart" the model is being on any single token.
Fable's always-on adaptive thinking is a further mechanical distinction, not just a marketing term.
Rather than answering in one pass, Fable is built to adjust how much internal reasoning it applies based on the difficulty of what's being asked, every time, without the user having to request it.
That's part of why Fable costs roughly five times what Sonnet costs per output token during Sonnet's intro pricing window: the adaptive thinking is doing real, metered work on every response.
Advanced Considerations & Applications
The practical upshot of this tiering is that most everyday work should default to Sonnet 5, and only a minority of tasks should reach for Opus 4.8 or Fable 5.
Anthropic made this explicit by setting Sonnet 5 as the default model for both Free and Pro plans starting at its June 30, 2026 launch, rather than defaulting to a cheaper Haiku-class model or a pricier Opus-class one.
That default reflects Sonnet 5's own benchmark strength: it scores strongly on coding benchmarks like SWE-bench and on agentic/browser-use tasks, and on some knowledge-work tasks it edges past Opus 4.8 despite costing roughly a third to a fifth as much per MTok.
At scale, the tiering also becomes a cost-control mechanism rather than just a capability choice.
An application that routes thousands of simple classification or extraction calls through Haiku 4.5, reserves Sonnet 5 for its main conversational or coding workload, and escalates only the hardest edge cases to Opus 4.8 or Fable 5 will spend an order of magnitude less than one that calls Fable for everything.
| Tier | Strength | Weakness | Best Fit |
|---|---|---|---|
| Claude Haiku 4.5 | Fastest, cheapest, high-volume friendly | Less depth on ambiguous, multi-step reasoning | Quick chat, high-volume/low-latency automation |
| Claude Sonnet 5 | Strong default balance of capability and cost; now the default tier | Not the absolute ceiling on hardest reasoning tasks | Everyday work, coding, agentic/browser tasks |
| Claude Opus 4.8 | Deep reasoning, 1M context by default | Costs more and responds slower than Sonnet | Hard reasoning, complex analysis |
| Claude Fable 5 | Most capable widely released tier, 1M context, 128K output, always-on adaptive thinking | Highest per-MTok cost of the four | The most demanding tasks needing maximum capability and context |
The tiering also has to keep evolving as models improve.
Sonnet 5's ability to edge past Opus 4.8 on some tasks shows that "newer, cheaper tier" can sometimes outperform an older, pricier one on specific benchmarks - which is why teams should periodically re-check their default model choice rather than assuming a past decision still holds.
Common Misconceptions
- "The most expensive model is always the best choice." - Fable 5 and Opus 4.8 aren't strictly "better" at every task; Sonnet 5 outperforms Opus 4.8 on some knowledge-work benchmarks despite costing far less, so the right tier depends on the task, not the price tag.
- "More tiers means Anthropic is fragmenting quality." - Each tier is a deliberate point on the capability/speed/cost curve, not a degraded version of the top model; Haiku 4.5 is built for its job, not a stripped-down Fable.
- "Context window size and reasoning capability are the same thing." - A 1M token context window (Fable, Opus) means the model can hold more information in view, not that it necessarily reasons more deeply than a smaller-context model on any single response.
- "You should pick one tier and always use it." - Most effective usage mixes tiers: Sonnet 5 as the default, with occasional escalation to Opus 4.8 or Fable 5, and Haiku 4.5 for high-volume or latency-sensitive work.
FAQs
Why doesn't Anthropic just make one really good Claude model?
Because capability, speed, and cost trade off against each other.
- A model tuned to reason as deeply as possible on every query would be too slow and expensive for quick chat or high-volume automation.
- A model tuned purely for speed would underperform on hard, ambiguous, or high-stakes tasks.
- Four tiers let each use case get a model actually built for it.
What are the four current Claude model tiers?
From lightest to heaviest: Claude Haiku 4.5, Claude Sonnet 5, Claude Opus 4.8, and Claude Fable 5.
Which model is the default one on Claude.ai?
Claude Sonnet 5, since its June 30, 2026 launch, is the default model for both Free and Pro plans.
Is Fable 5 always better than Opus 4.8?
Not on every task.
Fable 5 has more capability, a larger practical output ceiling, and always-on adaptive thinking, but Sonnet 5 (a cheaper, lighter tier than Opus) already edges past Opus 4.8 on some knowledge-work benchmarks - showing that the newest mid-tier model can outperform an older top-tier one on specific tasks.
What does "Mythos-class" mean for Fable 5?
It's the name Anthropic gave the new tier introduced in June 2026 that sits above Opus in the lineup - the most capable widely released model Anthropic offers, distinguished by its 1M token context, 128K max output, and always-on adaptive thinking.
How long has Claude had multiple size tiers?
Since the Claude 3 family launched in March 2024, Claude has shipped in Opus, Sonnet, and Haiku tiers; Fable was added in June 2026 as a new tier above Opus.
What's the real difference between Sonnet 5 and Haiku 4.5?
Sonnet 5 is the balanced, default tier for general everyday use, including coding and agentic tasks.
Haiku 4.5 is built specifically for speed and volume, with a smaller 200K context window and roughly a fifth of Sonnet's per-MTok cost, trading some reasoning depth for latency and price.
Why does output cost more than input on every tier?
Generating new tokens requires the model to do active reasoning and generation work for each one, while reading input tokens is comparatively cheaper to process - this pattern holds across all four tiers, with output priced roughly 4-5x input.
Does a bigger context window mean a smarter model?
No.
Context window size determines how much text (documents, code, conversation history) the model can hold in view at once; it's a separate property from reasoning capability, though the largest-context tiers (Fable and Opus, both at 1M tokens by default) also happen to be the more capable ones.
Should a team standardize on a single Claude tier?
Usually not.
Most effective setups default to Sonnet 5 for general work, escalate to Opus 4.8 or Fable 5 only for tasks that genuinely need deeper reasoning or maximum context, and route high-volume or latency-sensitive work to Haiku 4.5.
What is "always-on adaptive thinking" in Fable 5?
It means Fable 5 automatically adjusts how much internal reasoning it applies to a given query based on that query's difficulty, on every response, without the user needing to request extended thinking explicitly.
Will these four tiers stay the same forever?
No - Anthropic has already changed the lineup once, adding Fable above Opus in June 2026, and pricing itself shifts over time (for example, Sonnet 5's introductory pricing was set to change after August 31, 2026).
Model names, pricing, and exact positioning should be verified periodically rather than assumed to be permanent.
What's the single biggest mistake teams make with model tiers?
Defaulting to the most expensive tier "just to be safe," or conversely defaulting to the cheapest tier to save money regardless of task difficulty - both waste either budget or quality, when matching tier to task is usually straightforward once the trade-off is understood.
Related
- Claude Model Family Basics - a quick-start guide to picking a first model based on task and budget
- Claude Model Pricing Compared: Fable, Opus, Sonnet, and Haiku - the full per-MTok pricing table across all four tiers
- Choosing the Right Claude Model for Your Team's Task - a decision list matching everyday tasks to model tier
- How Claude Sonnet 5 Became the New Default Model - the story behind Sonnet 5's June 2026 launch as the new default
Stack versions: Written against the Claude model lineup current as of ~June 2026 - Claude Fable 5, Claude Opus 4.8, Claude Sonnet 5 (the default), and Claude Haiku 4.5. Model names, pricing, and product features move quickly - verify current specifics at platform.claude.com/docs before relying on them.