How to Choose Anthropic’s Current Claude Models: Fable, Opus, Sonnet, and Haiku Compared

Thu, 02 Jul 2026 07:27:14 +0800

Anthropic’s current Claude lineup is no longer as simple as “Haiku is fast, Sonnet is balanced, Opus is strongest.” As of July 2026, the main models include Claude Fable 5, Claude Opus 4.8, Claude Sonnet 5, Claude Haiku 4.5, and the limited-availability Claude Mythos 5.

If you want a quick rule: start with Sonnet 5 for daily development and most Agent tasks; use Opus 4.8 for complex enterprise Agents and hard coding; consider Fable 5 when you need the highest capability; use Haiku 4.5 for low-latency, cost-sensitive work; do not treat Mythos 5 as a general-purpose option because it is only available to approved customers.

Current main models

Model	API ID / alias	Positioning	Context	Max output	Latency tendency	Standard price
Claude Fable 5	`claude-fable-5`	Highest capability among broadly released Anthropic models, suited to long-running Agents	1M tokens	128k tokens	Slower	$10 / MTok input, $50 / MTok output
Claude Opus 4.8	`claude-opus-4-8`	Complex Agent coding, enterprise workflows, browser/computer use	1M tokens	128k tokens	Medium	$5 / MTok input, $25 / MTok output
Claude Sonnet 5	`claude-sonnet-5`	Best balance of speed, capability, and price; a strong default	1M tokens	128k tokens	Fast	$2 / $10 before 2026-08-31; then $3 / $15
Claude Haiku 4.5	`claude-haiku-4-5`	Fastest and lowest-cost model for high-throughput light tasks	200k tokens	64k tokens	Fastest	$1 / MTok input, $5 / MTok output
Claude Mythos 5	`claude-mythos-5`	Same specs and price as Fable 5, but limited availability	1M tokens	128k tokens	Slower	$10 / MTok input, $50 / MTok output

MTok means million tokens. Pricing may also be affected by prompt caching, batch processing, data residency, cloud platform, and region, so the table only includes common base input/output prices.

Fable 5: highest capability, but not the default

Claude Fable 5 is the highest-capability broadly released model in Anthropic’s current lineup. Its official positioning is “next-generation intelligence for long-running agents.”

It fits:

Long-running, multi-step Agent workflows.
Complex research, code migration, and cross-system planning.
Enterprise tasks that need the strongest reasoning and large context.
High-value tasks that are less sensitive to cost and latency.

But Fable 5 is not necessarily the default. It is the most expensive and slower. Unless the task truly needs the highest ceiling, starting with Sonnet 5 or Opus 4.8 is usually more practical.

One more detail: Fable 5 uses adaptive thinking and it is always on. It decides when and how much to reason automatically. That helps complex tasks, but it also makes cost and response time more dependent on the task itself.

Opus 4.8: a solid choice for complex coding and enterprise Agents

Claude Opus 4.8 sits between Fable 5 and Sonnet 5. Anthropic suggests starting with Opus 4.8 when you are unsure which model to use for complex Agent coding and enterprise work.

Its strengths include:

1M token context.
128k token max output.
Strong performance on complex coding, browser Agents, computer use, and enterprise workflows.
Half the price of Fable 5.
Support for adaptive thinking.

Opus 4.8 is well suited as the “hard-task default.” Codebase-level refactors, complex PR fixes, enterprise data analysis, multi-tool Agents, and long-document reasoning can all start here.

If a task is extremely hard, upgrade from Opus 4.8 to Fable 5. If task volume is high and cost pressure is obvious, move down to Sonnet 5.

Sonnet 5: the best daily default

Claude Sonnet 5 is the most important default candidate today. Its positioning is the best combination of speed and intelligence.

It fits:

Daily coding and code review.
Documentation, research assistance, and knowledge work.
Medium-complexity Agents.
Internal enterprise automation.
API applications that need cost control without losing too much quality.

The biggest change in Sonnet 5 is that many Agent capabilities that used to feel closer to Opus have moved into the Sonnet price tier. It also supports 1M token context and 128k token output, with lower latency than Opus.

Pricing has an introductory discount until August 31, 2026: $2 / MTok input and $10 / MTok output. From September 1, 2026, it returns to $3 / MTok input and $15 / MTok output. Even at the standard price, it remains significantly cheaper than Opus 4.8.

For most teams, I would start with Sonnet 5: let it cover 70% to 80% of tasks, then escalate truly difficult work to Opus 4.8 or Fable 5.

Haiku 4.5: high throughput, low latency, low cost

Claude Haiku 4.5 is the fastest model in the current main Claude line. Anthropic positions it as the fastest model with near-frontier intelligence.

It fits:

Classification, extraction, summarization, and format conversion.
Batch processing of short text.
Customer support, tickets, moderation, and other high-throughput scenarios.
Latency-sensitive interactive products.
Light tasks that do not need 1M context.

Its limits are clear: 200k token context and 64k token max output, below the 1M / 128k of Fable, Opus, and Sonnet. It should not be the first choice for long codebases, complex multi-document analysis, or long-running Agents.

But for large volumes of simple, fast tasks, Haiku 4.5 is straightforwardly cost-effective: $1 / MTok input and $5 / MTok output.

Mythos 5: not a normal option

Claude Mythos 5 shares the same specs and price as Fable 5, but it is not generally available. Anthropic documentation marks it as limited availability, only for approved Project Glasswing customers.

In other words, for ordinary API model selection, you usually do not need to include Mythos 5. Unless you are already approved or obtain access through Anthropic, AWS, or Google Cloud account teams, it is not a direct replacement for Fable 5.

How to choose: tier by task complexity

Use this order:

Start with Sonnet 5 by default
Good for most coding, documentation, Agent, and enterprise automation tasks.
Move to Opus 4.8 for clearly complex tasks
Long codebases, multiple tools, multi-step tasks, and stronger reasoning requirements.
Try Fable 5 when you need the highest capability
High-value, long-running, high-failure-cost tasks where price matters less.
Use Haiku 4.5 for high-throughput light work
Classification, extraction, summaries, support, batch processing, and low-latency interaction.
Consider Mythos 5 only if you have access
It is not a default option for ordinary developers.

Two migration and cost details

First, newer Claude models use a new tokenizer. Anthropic’s documentation says Opus 4.7 and later Opus models, Fable 5, Mythos 5, Mythos Preview, and Sonnet 5 may produce about 30% more tokens for the same text. Cost estimates should not rely on per-million-token price alone.

Second, 1M context does not mean every request should fill 1M context. Fable 5, Opus 4.8, and Sonnet 5 all support 1M tokens, but tool calls, caching, output, and multi-turn Agents add cost. A better deployment approach is:

Use prompt caching for common system prompts and long background context.
Chunk long documents first, then use stronger models for synthesis.
Send simple steps to Haiku or Sonnet, and escalate key decisions to Opus / Fable.
Run real task samples instead of relying only on official benchmarks.

A simple conclusion

Claude’s current model line is fairly clear:

Fable 5: highest capability for the hardest and highest-value tasks.
Opus 4.8: strong choice for complex Agent coding and enterprise work.
Sonnet 5: best daily default, balancing capability, speed, and price.
Haiku 4.5: fastest and cheapest for large-scale light tasks.
Mythos 5: limited availability, not a normal option.

If you are choosing Claude models for a product or internal workflow, the practical strategy is not to chase the highest tier. Split tasks: Haiku for lightweight batch work, Sonnet 5 as the default execution layer, Opus 4.8 for complex Agents and hard coding, and Fable 5 for the small set of tasks that are hardest, most expensive, and most worth it.

Model routing advice

When selecting Claude models, avoid having only one default model. A more useful design is routing: light batch jobs go to Haiku; daily coding and knowledge work go to Sonnet; complex repository tasks and multi-step Agents go to Opus; the highest-value and hardest tasks escalate to Fable.

The routing can start simple. Summaries, classification, and field extraction prefer Haiku. PR review, documentation generation, and ordinary code changes prefer Sonnet. Cross-module refactors, incident reviews, and complex planning prefer Opus. If Opus fails repeatedly or the task is very valuable, use Fable.

Every tier should also have exit conditions. Uncertain output, tool-call failures, repeated test failures, context over threshold, or tasks involving permissions or production data should trigger human confirmation instead of continued automation.

Cost evaluation method

The price table is only a rough estimate. Real cost depends on context length, cache hit rate, retries, output length, and human rework time. A more expensive model that completes a task once may be cheaper than multiple retries with a cheaper model.

For each task type, record three metrics: average token cost, average human review time, and the share of failures that require escalation. After two weeks, it is usually clear which tasks belong on Sonnet and which deserve Opus or Fable.

References:

Model Selection on KnightLi Blog

How to Choose Anthropic’s Current Claude Models: Fable, Opus, Sonnet, and Haiku Compared

Current main models

Fable 5: highest capability, but not the default

Opus 4.8: a solid choice for complex coding and enterprise Agents

Sonnet 5: the best daily default

Haiku 4.5: high throughput, low latency, low cost

Mythos 5: not a normal option

How to choose: tier by task complexity

Two migration and cost details

A simple conclusion

Model routing advice

Cost evaluation method