Claude 5 Field Guide: Fable, Mythos, Opus, Sonnet, and Haiku

Anthropic’s current model lineup looks confusing because it encodes more than intelligence, speed, and price. Fable and Mythos share an underlying model but expose different capabilities to different users. Opus 5 competes with Fable on many everyday frontier tasks at half the price. Sonnet 5 brings agentic behavior to the scaled production tier. Haiku 4.5 remains the speed and volume option.

The result is not one clean ladder. It is a matrix of capability, effort, safeguards, access, and deployment surface.

This guide reflects the model and platform documentation available on August 4, 2026. Benchmark claims are Anthropic’s unless stated otherwise; test every candidate on your own workload before changing a production route.

The current lineup

Anthropic maintains the canonical specifications in the models overview and recommends an eval-driven choice in its model selection guide. Prices and availability change faster than prose does, so treat those pages as authoritative at implementation time.

Fable 5: maximum capability with an active safety layer

Anthropic released Claude Fable 5 on June 9 as its most capable widely available model. The launch emphasized long-running software work, scientific research, vision, persistent file-based memory, and tool use. Anthropic says the capability advantage grows as tasks become longer and more complex.

Fable’s distinguishing feature is not only its benchmark position. It runs additional safety classifiers for cybersecurity, biology and chemistry, and model distillation. If a request crosses a classifier boundary, Anthropic may route it to another Claude model. The product informs the user when this happens; the API can use server-side fallbacks.

Anthropic initially said the classifiers triggered in fewer than 5% of sessions on average, but that average is not a guarantee for any workload. A security team, biology lab, or low-level systems developer may encounter a very different rate. Evaluate:

• how often your real prompts are flagged;
• whether fallback responses preserve the required quality and tool behavior;
• whether a model switch is visible in traces and usage data;
• whether the fallback model remains compliant with your data and retention requirements.

Fable is appropriate when the expected value of a stronger answer exceeds the 2× Opus price and the workload is not systematically disrupted by its classifiers. Do not route all traffic to it just because it sits at the top of the table.

Mythos 5: the same intelligence under a different trust contract

Claude Mythos 5 uses the same underlying model as Fable 5, but some safeguards are lifted for approved use. Anthropic launched it to a small set of cyberdefenders and critical-infrastructure partners through Project Glasswing, with plans for separate trusted access in biology.

This makes Mythos a deployment category, not a normal upgrade. The relevant questions are organizational:

• Who is eligible and verified?
• Which safeguards are relaxed for this program?
• Which personnel may invoke the model?
• What monitoring, disclosure, and incident duties apply?
• Where do generated exploits, biological designs, and research artifacts go?

Anthropic says Mythos 5 is stronger than Opus 5 at vulnerability exploitation, while Opus can approach it at vulnerability discovery. The Opus 5 launch uses that distinction to explain why Opus receives a less restrictive cyber policy than Fable. It is a useful reminder that “cyber capability” is not one scalar: code review, vulnerability discovery, binary analysis, penetration testing, and exploit construction carry different risk.

Opus 5: the daily frontier model

Claude Opus 5, released July 24, is the most consequential option for many teams. Anthropic positions it close to Fable on coding and professional work at half Fable’s token price, and made it the default on Claude Max and the strongest model on Claude Pro.

Its API behavior also changes important defaults:

• adaptive thinking is on unless explicitly disabled;
• effort ranges from low through max;
• thinking cannot be disabled at xhigh or max effort;
• the context window is 1 million tokens by default, not a separate long-context variant;
• mid-conversation tool changes and automatic fallback modes are available in beta;
• the minimum cacheable prompt is smaller than on Opus 4.8.

Anthropic’s Opus 5 migration guide warns about behavioral changes, too. The model narrates agentic work more readily, delegates to subagents more often, and verifies its own work without needing older “double-check” instructions. Those are not cosmetic differences. They change token use, latency, tool concurrency, and the shape of user-visible progress.

When upgrading, remove inherited scaffolding only after comparing traces. A prompt that mandates a verifier subagent may compound Opus 5’s own verification behavior. Conversely, a production workflow with compliance-required independent review should retain that control regardless of what the model tends to do voluntarily.

Sonnet 5: agentic execution at the production tier

Claude Sonnet 5 is the default for Free and Pro plans and the likely starting point for high-volume agent applications. Anthropic says it narrows the gap with Opus 4.8 and can match Opus-class capability on some tasks at higher effort, with substantially better cost efficiency at medium effort.

Three migration details deserve attention:

1. Adaptive thinking is on by default.
2. Manual fixed thinking budgets are removed.
3. Non-default sampling parameters return an error.

Sonnet 5 also uses the tokenizer introduced in newer Claude models. Anthropic says the same text can produce roughly 1.0–1.35× as many tokens depending on content. The promotional $2/$10 price through August 31 is intended to make migration roughly cost-neutral, but the standard price begins September 1. Recount prompts and replay real traces before forecasting spend.

The best Sonnet use cases are not simply “easy tasks.” They are tasks where the model can sustain the required tool loop and quality at a better cost-performance point: data analysis, business automation, routine software changes, extraction with judgment, and scaled subagent lanes.

Haiku 4.5: use speed as an architectural advantage

Haiku 4.5 remains Anthropic’s current low-latency tier. It is a reasonable starting point for classification, routing, high-volume transformation, narrow tool calls, and cheap subagents.

The important design move is to avoid making one model carry every stage. A production system can use Haiku for intent classification or schema repair, Sonnet for routine execution, Opus for ambiguous or high-value cases, and Fable only when the capability lift justifies the cost and classifier behavior. That is a routing policy, not a benchmark ranking.

Effort is now a first-class model-selection dimension

For Opus 5 and Sonnet 5, the model name no longer fully describes the operating point. Effort trades more inference time and tokens for capability. Anthropic recommends starting Opus 5 at high, stepping up for demanding long-horizon work, and stepping down where evals show quality holds.

This creates a two-dimensional frontier:

An effort sweep should be part of every model bake-off. Compare at least:

• pass rate and severity-weighted failures;
• total tokens, not only listed price;
• wall-clock latency and time to first useful artifact;
• number and cost of tool calls;
• variance across repeated runs;
• rate of escalations, refusals, and fallbacks;
• human review time after the run.

The cheapest successful run often comes from a stronger model at lower effort, not a weaker model at its maximum. That is workload-specific; measure it.

Safeguards make routing observable—or dangerous

Anthropic’s server-side fallback can turn a refusal on one model into a response from another. That improves continuity, but it can also make the application’s effective model nondeterministic from the caller’s point of view.

A safe integration should record:

{
  "requested_model": "claude-fable-5",
  "served_model": "claude-opus-5",
  "route_reason": "safety_fallback",
  "refusal_category": "cyber",
  "effort": "high",
  "policy_version": "model-route-2026-08-04",
  "toolset_version": "sec-review-v7"
}

The exact API fields differ from this conceptual receipt. The principle is durable: if policy can change the model, the route belongs in the decision record.

A model-selection playbook

1. Define task slices. Separate routine, ambiguous, long-horizon, high-risk, and latency-critical work.
2. Start with the cheapest plausible candidate. Use Anthropic’s efficiency-first approach where failures are inexpensive.
3. Run an effort sweep. Do not compare only default settings.
4. Replay production-like tools and context. Agent performance is model plus harness.
5. Test classifier-heavy slices separately. Average fallback rates hide domain concentration.
6. Score total outcome cost. Include retries, tools, reviewers, and remediation.
7. Pin the route. Version the model, effort, prompt, tool schema, fallback policy, and eval set together.

The durable takeaway

Claude’s lineup is becoming a capability-and-trust architecture. Fable offers maximum general access under strong runtime safeguards. Mythos exposes more dual-use capability to verified users. Opus 5 is the practical frontier workhorse. Sonnet 5 is the scaled agent model. Haiku remains the fast execution tier.

Choose the route that produces a trusted outcome at the lowest total cost. The model name is only one input.

What to read next

• Anthropic in 2026: the full-stack map
• From Claude Code to Cowork and Claude Tag
• Inside the Claude Platform and Managed Agents
• Tokenomics: cost per trusted outcome