Claude Models Explained: Fable, Opus, Sonnet, Haiku and Mythos

by David Porter

Thursday, 06 August 2026 at 06:12

claude-models

Anthropic’s current public Claude models are Fable 5, Opus 5, Sonnet 5 and Haiku 4.5. Claude Mythos 5 uses the same underlying specifications and price as Fable 5 but is offered only to approved defensive cybersecurity partners. For most users, the practical decision is between Sonnet for balance, Opus for demanding professional work, Fable for the hardest long-running tasks and Haiku for speed and cost.

The model name is only one layer. Claude subscriptions determine which models and allowances appear in the Claude app. The API exposes model IDs and metered rates. Cloud platforms can use different identifiers and feature sets. This guide separates those layers.

For a tour of the complete product, see our Claude AI cornerstone. For subscriptions and full cost details, see Claude pricing.

Current Claude models at a glance

The following specifications come from Anthropic’s current model documentation on July 27, 2026.

Model	Public position	API context	Maximum output	Base API price per MTok	Relative latency
Claude Fable 5	Highest-capability widely released model for long-running agents	1 million	128,000	$10 input / $50 output	Slower
Claude Opus 5	Complex agentic coding and professional work	1 million	128,000	$5 input / $25 output	Moderate
Claude Sonnet 5	Best balance of speed and intelligence	1 million	128,000	$3 input / $15 output standard	Fast
Claude Haiku 4.5	Fastest and most economical current model	200,000	64,000	$1 input / $5 output	Fastest
Claude Mythos 5	Restricted defensive cybersecurity model	1 million	128,000	Same as Fable 5	Slower

Sonnet 5 has introductory API pricing of $2 per million input tokens and $10 per million output tokens through August 31, 2026. Its published standard rate is $3 and $15 afterward.

The context figures above are API specifications. Anthropic’s individual Claude plan table currently shows a 200,000-token context window, while Enterprise lists 500,000 on the default model. A one-million-token API context therefore does not promise a one-million-token chat in every subscription.

What do Claude model names mean?

Historically, Anthropic’s main family used three tier names:

Haiku: fast and inexpensive;
Sonnet: balanced;
Opus: most capable and expensive.

In 2026, Anthropic added Fable and Mythos above or beside that familiar hierarchy. Fable is the most capable model generally available. Mythos is a restricted variant intended for approved defensive cybersecurity work.

The names are product labels, not scientific units. An Opus generation can be overtaken by a newer Sonnet, and price does not map perfectly to every benchmark. Compare the exact model version on the exact task.

Claude Fable 5

Claude Fable 5 is Anthropic’s highest-capability model available to the broad market. It is designed for long-running agents and work that must maintain a plan, use tools and recover from problems over an extended trajectory.

Anthropic released Fable 5 in June 2026 alongside Mythos 5. Both use the same underlying model, but Fable carries stronger cybersecurity safeguards for general release. Access was briefly suspended after a US export-control intervention and later restored globally with updated classifiers.

Best uses for Fable 5

Fable is most defensible when:

a task is genuinely at the frontier of what other models complete;
an agent must sustain work over many steps;
failures are expensive enough to justify a higher model bill;
or testing shows a meaningful success-rate improvement over Opus.

Examples include difficult repository-scale engineering, complex tool-using agents, advanced research synthesis and high-value planning where the model must preserve constraints over a long execution.

Fable 5 limitations

Fable is the slowest and most expensive generally available current model. It also has enhanced cyber safeguards that can block legitimate security or programming requests. Anthropic says the updated classifier can create more false positives in routine coding and debugging.

Fable 5 is also a designated Covered Model. It requires 30-day data retention and is not available under a zero-data-retention arrangement. That can rule it out even when its capability fits the task; our Claude privacy and security guide explains the policy boundary.

Do not default every task to Fable. A model that costs five times as much as Sonnet input at the standard rate needs to reduce errors, retries or staff time enough to justify that difference.

Claude Opus 5

Claude Opus 5 is Anthropic’s advanced everyday model for complex coding, agentic work and professional analysis. Anthropic introduced it on July 24, 2026 and describes it as approaching Fable-level intelligence at half Fable’s API price.

Opus 5 costs the same as its predecessor Opus 4.8: $5 per million input tokens and $25 per million output tokens. It is available across Anthropic’s platforms and through supported distribution channels.

Best uses for Opus 5

Choose Opus for:

difficult software engineering;
financial, legal or technical analysis that has human review;
open-ended work where judgment matters;
multi-tool agents;
planning and critique;
and tasks where Sonnet completes the work but needs too many corrections.

Opus is often the practical “strong model” choice. It is less costly and faster than Fable while retaining a large context and output capacity.

Opus 5 and Fast mode

Anthropic offers a research-preview Fast mode for Opus 5. It uses the same model but can produce output at up to 2.5 times the standard speed. The rate doubles to $10 per million input tokens and $50 per million output tokens.

Fast mode mainly improves output throughput, not necessarily time to the first token. It should be purchased when latency has business value, not because it sounds like a more intelligent model.

Opus 5 and cybersecurity

Anthropic says it intentionally avoided training Opus 5 specifically on cyber tasks. General capability still improved its security performance, but Mythos remains the specialized restricted model and Fable’s public safeguards are designed around a different risk boundary.

That distinction matters when reading benchmarks. “Can find vulnerabilities” and “can operationalize an exploit” are different capabilities.

Claude Sonnet 5

Claude Sonnet 5 is the general-purpose balance of capability, speed and price. It is the sensible starting model for most API applications and everyday advanced work.

Anthropic made Sonnet 5 available across all Claude plans in June 2026. It is positioned for coding, agents and professional tasks at scale rather than only lightweight chat.

Best uses for Sonnet 5

Use Sonnet for:

standard writing and document work;
customer-facing applications with guardrails;
extraction and structured output where the task has nuance;
most coding assistance;
research with tools;
and production agents whose economics matter.

Start with Sonnet, create an evaluation set and escalate only the cases it cannot meet. This produces a more rational architecture than using the most expensive model everywhere.

Sonnet 5 pricing transition

Through August 31, 2026, Anthropic lists Sonnet 5 at:

$2 per million input tokens;
$10 per million output tokens;
$2.50 per million cache-write tokens;
$0.20 per million cache-read tokens.

The announced standard rates after the introductory period are:

$3 input;
$15 output.

Budget with the standard rate if the application will operate after August. An introductory price is a launch discount, not a permanent unit-economics assumption.

Claude Haiku 4.5

Claude Haiku 4.5 is the fastest and least expensive current model. Its API context is 200,000 tokens and its maximum output is 64,000 tokens.

Haiku costs:

$1 per million input tokens;
$5 per million output tokens;
$1.25 per million cache-write tokens;
$0.10 per million cache-read tokens.

Best uses for Haiku 4.5

Haiku is a strong choice for:

classification;
routing;
field extraction;
moderation layers;
short summaries;
simple transformations;
high-volume support triage;
and low-latency steps inside a larger agent.

An architecture can use Haiku to decide which requests need Sonnet or Opus. That often matters more than debating one universal default.

When not to use Haiku

Do not choose Haiku solely because it is cheap when the task requires subtle judgment, deep planning or reliable long-horizon tool use. A cheap failed request followed by an expensive retry can cost more than a successful Sonnet request.

Measure cost per accepted result, not price per token.

Claude Mythos 5

Claude Mythos 5 is not a normal consumer or self-serve API model. It is offered in limited availability through Project Glasswing to approved customers for defensive cybersecurity.

Anthropic says Mythos shares Fable 5’s core specifications and pricing. The difference is access and safeguards: Mythos exposes advanced cyber capabilities within a restricted program, while Fable is the general-release model with stronger protections against potentially harmful use.

There is no public self-serve sign-up. A model menu, third-party wrapper or social-media claim that offers “unrestricted Mythos” should be treated skeptically.

Opus versus Sonnet

This is the most common practical comparison.

Decision factor	Opus 5	Sonnet 5
Capability target	Hard professional and agentic work	Broad high-performance production work
Speed	Moderate	Fast
Standard API input	$5/MTok	$3/MTok
Standard API output	$25/MTok	$15/MTok
Context	1M API tokens	1M API tokens
Best default	Difficult cases	Most applications

Choose Sonnet first. Move a task to Opus when evaluation shows one of these patterns:

Sonnet’s success rate is too low;
the human correction time is expensive;
the task depends on nuanced judgment;
or an agent fails to preserve the plan across steps.

Stay with Sonnet when the output already passes review. Paying for unused intelligence does not improve the product.

Fable versus Opus

Fable is more capable at the top end and designed for long-running agents. Opus is faster, costs half as much per token and is intended as a strong everyday professional model.

Use Fable when Opus has been tested and still fails a valuable task. Use Opus when it reaches the acceptance threshold.

The decision is not philosophical. Run a controlled evaluation:

Collect 20–100 representative difficult tasks.
Hide model identity from reviewers where possible.
Score success, correction time, latency and token use.
Calculate total cost per accepted outcome.
Route only the difficult tail to Fable if that is where the benefit appears.

Sonnet versus Haiku

Haiku is cheaper and faster. Sonnet is more capable and has a larger API context.

Use Haiku for predictable, narrow transformations and routing. Use Sonnet when language ambiguity, multi-step reasoning or tool decisions materially affect the answer.

A tiered pipeline can:

classify and validate the request with Haiku;
handle ordinary requests with Sonnet;
escalate difficult cases to Opus;
reserve Fable for the hardest verified need.

This gives model choice an operational purpose.

Context windows explained

A context window is the amount of tokenized information a model can consider in one request, including system instructions, messages, tool definitions, tool results, images and documents.

Anthropic documents:

1 million tokens for Fable 5, Opus 5 and Sonnet 5 through the API;
200,000 for Haiku 4.5;
128,000 maximum output for the 5-series models;
64,000 maximum output for Haiku 4.5.

One million tokens is not perfect memory

A large context window can hold a very large source set, but it does not guarantee:

every fact will be retrieved;
attention is equal across the whole input;
contradictions will be resolved;
or the answer will be correct.

Structure the context. Put authoritative instructions in the system layer, label sources, remove irrelevant material and ask for references. For a large persistent collection, retrieval can be cheaper and more controllable than sending the entire corpus each time.

App context versus API context

Subscription pages may expose smaller context allowances than the underlying model supports through the API. The product must manage cost, latency, attachments, conversation history and other features.

Always cite the surface:

“Sonnet 5 supports a 1M API context” is precise.
“Every Claude Pro chat holds 1M tokens” is not supported by Anthropic’s current plan table.

Knowledge cutoffs

Anthropic distinguishes a reliable knowledge cutoff from the broader training-data cutoff.

Model	Reliable knowledge cutoff	Training-data cutoff
Fable 5	January 2026	January 2026
Opus 5	May 2026	May 2026
Sonnet 5	January 2026	January 2026
Haiku 4.5	February 2025	July 2025

A cutoff does not mean the model knows every event before that date or nothing after it. It describes the training boundary and expected reliability. Use web search or supplied data for anything current.

Adaptive thinking and effort

Fable 5, Opus 5 and Sonnet 5 use adaptive thinking. Haiku 4.5 supports the older extended-thinking configuration rather than adaptive thinking.

The effort parameter lets a developer trade response depth against token use and latency. Higher effort can help on difficult agentic or analytical work, but it can also increase output and cost.

Do not set maximum effort universally. Evaluate:

low effort for routine transformations;
medium or high for reasoning-heavy tasks;
and the highest settings only where they improve acceptance.

The Claude API guide shows how model and request choices fit into an application.

Model IDs

Current first-party API identifiers include:

claude-fable-5 claude-opus-5 claude-sonnet-5 claude-haiku-4-5-20251001

The Haiku convenience alias is:

claude-haiku-4-5

Anthropic says the dateless identifiers used from the 4.6 generation onward are still pinned snapshots rather than pointers that silently become a new model. Older convenience aliases can resolve to a dated model ID.

This is important for production change control. Do not assume a dateless-looking ID means automatic upgrades. Read the deprecation policy and test a new model explicitly.

AWS Bedrock and Google Cloud can use provider-specific identifiers. Claude Platform on AWS uses Anthropic-style model IDs rather than Bedrock IDs. Copy the identifier from the documentation for the exact service.

How to choose the best Claude model

Use this sequence:

1. Define acceptance

What must be correct? What format is required? How much review is allowed? How fast must it respond?

2. Start with Sonnet

Sonnet is the balanced baseline for most non-trivial work.

3. Test Haiku

If the task is narrow or high-volume, see whether Haiku meets the same acceptance threshold.

4. Escalate failures

Send difficult cases to Opus. Test Fable only where Opus still fails or where long-horizon performance is the core need.

5. Measure the full outcome

Include:

model cost;
retries;
latency;
tool calls;
reviewer time;
and the cost of an incorrect result.

6. Route dynamically

One application can use several models. A fixed “best model” is often less efficient than a routing policy.

Which Claude model is best for each task?

Task	Recommended starting point	Why
Everyday writing and editing	Sonnet 5	Strong quality without Opus cost
Simple classification at scale	Haiku 4.5	Fast and inexpensive
Difficult repository refactor	Opus 5	Strong agentic coding and judgment
Long-running frontier agent	Fable 5	Highest generally available capability
Standard customer-support response	Haiku 4.5 or Sonnet 5	Choose by ambiguity and risk
Complex financial analysis	Opus 5	Better fit for difficult professional reasoning
Web research synthesis	Sonnet 5	Balance of speed, quality and tool use
High-stakes final decision	No model alone	Use a suitable model plus independent human and source review
Defensive frontier cybersecurity	Mythos 5 if approved	Restricted specialist access

These are starting points, not guarantees. The prompt, evidence, tools and evaluation standard can change the result more than the tier name.

Legacy Claude models

Earlier models do not become fake or useless when a new generation launches. They may remain in production because:

behavior is already validated;
migration would change tokenization or output;
a cloud provider has a different schedule;
or a particular feature still depends on the older model.

Legacy models can also be deprecated. Maintain an inventory of model IDs, owners and planned migration dates. Run regression tests before switching.

Do not build a new article section for every past version. Historical model releases belong in news coverage or a version timeline; the evergreen page should prioritize currently supported choices.

Frequently asked questions

What is the newest Claude model?

Claude Opus 5 is Anthropic’s newest public model as of July 27, 2026. Fable 5 remains the highest-capability model broadly available.

Is Fable better than Opus?

Fable is positioned above Opus for maximum capability and long-running agents. Opus is faster and costs half as much per token. “Better” depends on whether Fable improves the actual task enough to justify the difference.

Is Opus better than Sonnet?

Opus targets more difficult professional and agentic work. Sonnet is the better default for most applications because it balances quality, speed and cost.

What is Claude Mythos?

Mythos 5 is a restricted model for approved defensive cybersecurity partners in Project Glasswing. It is not a normal Claude plan or public self-serve API option.

Which Claude model is free?

The Free plan currently uses available default models within plan limits; model access can change. A consumer plan is not the same as purchasing one API model.

Which Claude model has the largest context window?

Fable 5, Opus 5 and Sonnet 5 each have a documented one-million-token API context. Consumer and workplace plans can expose different context limits.

Which Claude model is cheapest?

Haiku 4.5 has the lowest current base API price at $1 per million input tokens and $5 per million output tokens.

Can I use several Claude models in one application?

Yes. Many efficient systems route simple work to Haiku, standard work to Sonnet and difficult cases to Opus or Fable.

Bottom line

Sonnet 5 is the best starting model for most work. Haiku 4.5 is the speed-and-cost option, Opus 5 is the advanced professional model, and Fable 5 is the highest-capability public choice for difficult long-running agents. Mythos 5 is restricted, not a consumer upgrade.

Choose with an evaluation, not a hierarchy. The right model is the least expensive and fastest one that reliably passes the acceptance standard for the task.

byDavid Porter

Apple asks judge to immediately restrict OpenAI; ChatGPT maker hits back hard

EU AI Act: Article 4 and AI Literacy—What You Need to Know

Write a comment