Anthropic · June 2026

Claude Model Guide

Which model for which task — with real output examples.

Claude Haiku 4.5 Fast · Cheap · High-volume ▼
Speed
Fastest
Cost
$1 / $5 per MTok
Context
200K tokens
Best for
Scale & routing

Use Cases

Lead classification Bulk summarisation Customer support bots Content tagging Multi-agent subtasks Rapid prototyping High-volume extraction

Example Outputs

↳ Prompt: Classify this lead — "We're a 500-person fintech, need AI compliance tools, budget TBD, timeline Q3."
Segment: Enterprise B2B Intent: High (specific use case + timeline) Industry: Fintech / Regulated Priority: Tier 1 — route to senior sales Action: Schedule discovery call within 48hrs
↳ Prompt: Summarise this 800-word article about MAS regulations in 2 sentences.
MAS has tightened AI governance requirements for financial institutions, mandating explainability documentation and human oversight for high-risk decisioning systems. Firms must submit compliance frameworks by Q4 2026 or face operational restrictions.
Claude Sonnet 4.6 Balanced · Default · Production-ready ▼
Speed
Fast
Cost
$3 / $15 per MTok
Context
1M tokens
Best for
Most everything

Use Cases

Thought leadership Client deliverables Strategy documents LinkedIn content Workshop materials Framework development Email campaigns Long-form analysis Code generation SEO audits

Example Outputs

↳ Prompt: Write a LinkedIn hook about why most AI pilots fail.
Most AI pilots don't fail because the technology doesn't work. They fail because no one asked the harder question first: "What problem are we actually solving — and who owns it when the pilot ends?" Pilot purgatory isn't a tech problem. It's a governance problem dressed up as innovation theatre. Three things I see missing every time:
↳ Prompt: Draft an executive summary for an AI readiness assessment for a financial services client.
Executive Summary This assessment evaluates [Client]'s AI adoption maturity across People, Process, and Platform dimensions. Scored against the P-P-P framework, the organisation presents strong infrastructure foundations (Platform: 72/100) but faces material gaps in change readiness and cross-functional AI literacy (People: 44/100) that will constrain deployment at scale. Priority recommendations centre on three areas: structured upskilling for mid-management, formalising a human-in-the-loop governance model before scaling, and establishing measurable success criteria tied to business outcomes rather than model metrics. Without addressing the People gap, technology investment risk remains high.
Claude Opus 4.8 Deep reasoning · Long-horizon · High-stakes ▼
Speed
Deliberate
Cost
$5 / $25 per MTok
Context
1M tokens
Best for
Complex systems

Use Cases

Multi-file architecture Compliance systems Deep research synthesis Agentic workflows Complex debugging Long refactors High-stakes analysis Self-verifying outputs

Example Outputs

↳ Prompt: Design the architecture for a MAS advertising compliance decisioning system with human-in-the-loop review.
Architecture: ADNA Compliance Decisioning System Layer 1 — Ingestion • Asset intake API accepts creative (image, copy, video frame) • Pre-processing: OCR extraction, metadata tagging, campaign context binding • Risk signal detection: flagged keywords, claim types, product categories Layer 2 — AI Decisioning Engine • Primary classifier: MAS CMCG rule-set mapping (37 guideline nodes) • Confidence scoring: HIGH (≥0.92) → auto-approve; LOW (<0.72) → escalate • Explainability output: rule citation + confidence rationale per decision Layer 3 — Human-in-the-Loop Gate • Review queue for borderline cases (0.72–0.92 confidence band) • Reviewer interface: decision + annotated rule breach highlighted • Override logging: all human decisions feed back to training pipeline Layer 4 — Audit & Compliance Reporting • Immutable decision log per asset version • Regulator-ready export: timestamp, model version, rule applied, reviewer ID • Escalation analytics: drift detection on override frequency Failure mode safeguard: If model confidence degrades >15% week-on-week, system defaults all decisions to human review until recalibration.
↳ Prompt: Find the bug — this React state update isn't triggering a re-render.
// Bug: mutating state directly instead of returning new reference // This won't trigger re-render: state.items.push(newItem) setState(state) // Fix: return a new array reference setState(prev => ({ ...prev, items: [...prev.items, newItem] })) // Root cause: React uses shallow comparison for state updates. // Mutating the existing array keeps the same reference, // so React sees no change and skips the re-render cycle. // Always return a new object/array when updating nested state.
Claude Fable 5 Frontier · Maximum power · Long-horizon ▼
Speed
Adaptive
Cost
$10 / $50 per MTok
Context
1M tokens
Best for
Frontier tasks

Use Cases

Week-long autonomous runs Frontier research Complex multi-agent systems Scientific reasoning Cross-domain synthesis High-autonomy coding

Example Outputs

↳ Prompt: Synthesise competing research on AI governance frameworks across EU, US, and Singapore — identify gaps and recommend a hybrid model.
Cross-Jurisdictional AI Governance Synthesis The EU AI Act (risk-tiered, compliance-first) and US Executive Order approach (sector-led, innovation-permissive) represent opposing poles: the former mandates pre-market conformity assessment for high-risk AI; the latter delegates to agencies with limited cross-sector coherence. Singapore's Model AI Governance Framework occupies a distinct third position — principles-based, voluntary, and explicitly designed for ASEAN adoption friction. Its human-in-the-loop emphasis and sectoral flexibility make it the most implementation-ready of the three, but it lacks enforcement teeth. Key gaps across all three: 1. No jurisdiction has resolved liability attribution in multi-agent systems where no single model makes the final decision 2. Explainability requirements remain undefined for proprietary foundation models 3. Cross-border data flow rules create compliance conflicts for APAC-headquartered global firms Recommended hybrid: Adopt Singapore's principles-based structure as the operational layer, with EU Act-style risk-tiering applied only to deployment contexts (not model development), and US sector-agency accountability as the enforcement mechanism by industry vertical. This allows speed-to-market while maintaining audit readiness for regulated sectors.
⚠ Note: As of June 2026, Fable 5 access may be affected by a US export-control directive. Check availability before building workflows that depend on it.

Quick Comparison

Dimension Haiku 4.5 Sonnet 4.6 Opus 4.8 Fable 5
Speed ⚡⚡⚡⚡ ⚡⚡⚡ ⚡⚡ ⚡⚡ (adaptive)
Reasoning depth Basic Strong Deep Frontier
Context window 200K 1M 1M 1M
Cost (input/output) $1 / $5 $3 / $15 $5 / $25 $10 / $50
Writing & strategy Adequate Excellent Excellent Excellent
Complex coding Simple only Strong Best Best
Agentic tasks Subtasks only Good Very strong State of art
Self-verification Minimal Moderate Strong (4× Opus 4.7) Strongest
Default recommendation Bulk/scale work ✓ Start here Escalate to this Frontier tasks only