A complete guide to today’s Claude models (Autumn 2026 Edition)
Claude is a family of generative AI models developed by Anthropic, designed for content creation and analysis, software development, and business process automation. Individual models offer different trade-offs between capability, speed, and cost. The right choice is therefore not necessarily the most advanced model, but the one best suited to the specific task and your business objectives.
Choosing among Haiku, Sonnet, Opus, and Fable is no longer simply a technical question. It is a business decision. In this guide, we explain how to approach that decision thoughtfully and select the tool that genuinely meets your needs.
When choosing an AI model, two questions often come up:
– “Why not simply choose the most capable one?”
– “Why would anyone deliberately opt for a less capable model?”
Under real-world workloads, the answer quickly becomes clear. The most capable model is also the slowest and most expensive. For a large share of business tasks, the difference in output quality is negligible, while the differences in cost and response time are substantial.
Anthropic’s Claude ecosystem now comprises four publicly available models – Haiku 4.5, Sonnet 5, Opus 5, and Fable 5 – along with a restricted-access tier called Mythos 5, available to verified organizations. Each model offers a distinct balance of speed, intelligence, and cost. Understanding these differences is not merely a technical exercise; it is a strategic priority. Companies that choose the right model develop faster, spend less, and achieve better results than those that automatically default to the most advanced model for every interaction.
The Claude Family in 2026
Anthropic’s naming convention does not follow a simple linear progression. The current lineup includes Claude Haiku 4.5 alongside Claude Sonnet 5, Opus 5, and Fable 5 – the numbers reflect generational position within each model class, not a single unified scale. Fable and Mythos represent a new capability category above Opus, introduced with the 5-series. Sonnet 5, Opus 5, and Fable 5 all support a one-million-token context window with up to 128,000 output tokens. Haiku 4.5 operates within a 200,000-token context and supports up to 64,000 output tokens. All four models handle text and images, support tool use, and offer multilingual and vision capabilities. At a glance:
– Claude Haiku 4.5: $1/$5 per million tokens. Speed and efficiency at scale. Built for high-volume, predictable work.
– Claude Sonnet 5: $2/$10 per million tokens. The balanced workhorse for everyday professional tasks. Strong enough for almost everything.
– Claude Opus 5: $5/$25 per million tokens. Deep reasoning and autonomous workflows. When quality matters more than speed or cost.
– Claude Fable 5.1: $10/$50 per million tokens. Maximum widely available capability for demanding, long-horizon work.
Claude Haiku 4.5: Speed and Efficiency at Scale
Haiku is the model built for the long tail of enterprise workloads. The tasks that run thousands of times a day with consistent quality, where latency directly affects user experience and unit cost determines viability.
Customer support triage, document classification, extraction pipelines, automated content tagging, workflow routing, lightweight internal assistants – these are Haiku’s natural territory. According to Anthropic’s launch reporting, Haiku 4.5 delivered coding performance comparable to the earlier Sonnet 4 at one-third the cost and more than twice the speed, which tells you something about how far the efficiency tier has come.
It is also not a simple chatbot. Haiku supports vision, tool use, and extended thinking. What it lacks is the broader reasoning capacity needed for ambiguous, multi-stage, or genuinely open-ended tasks.
One underappreciated use case: Haiku works well as a sub-agent within a multi-model architecture, handling straightforward subtasks efficiently while a more capable model handles complex decisions and oversight. That design pattern – routing by task complexity rather than defaulting to a single model for everything – is increasingly how sophisticated teams build.
Choose Haiku 4.5 when response latency is a product requirement, requests are high-volume and relatively predictable, the instructions are clear, and the task has limited ambiguity, or cost per call is a primary constraint.
Claude Sonnet 5: The Balanced Default
For most businesses evaluating Claude, Sonnet 5 is the right place to start. It offers the best combination of intelligence, speed, and price in the family, and its capabilities cover the vast majority of professional use cases without the cost overhead of the frontier tiers.
Anthropic introduced Sonnet 5 in June 2026 as its most agentic Sonnet model to date – capable of planning, working with browsers and terminals, using tools, and managing workflows that previously required a larger model. It is the default model for the Free and Pro plans, reflecting where Anthropic itself expects most usage to land.
The strongest argument for Sonnet is practical rather than technical. A team can run everyday software development, content production, research synthesis, business analysis, and internal automation on Sonnet without paying frontier prices on every interaction. That efficiency compounds over time. A business spending fifty interactions a day on tasks that Sonnet handles as well as Opus is leaving meaningful money on the table if it defaults to Opus across the board.
Sonnet also has a one-million-token context window – enough to work across large documents, repositories, and datasets – and, at higher-effort settings, its performance on complex tasks can approach Opus 4.8. Opus 5 has since raised that ceiling, but the overlap is real for many workloads.
Choose Sonnet 5 when you need a capable everyday model across several kinds of professional work, coding and tool use matter, but maximum reasoning depth is not required, or you are still establishing performance benchmarks for your application. Sonnet is also the logical starting point before deciding whether Opus or Fable creates enough additional value to justify the step up.
Claude Opus 5: Complex Reasoning and Serious Agentic Work
Opus 5 is Anthropic’s high-end workhorse – released on 24 July 2026 as a step-change improvement over Opus 4.8. The gains are most pronounced in deep reasoning, long-horizon tasks, and agentic work where the model must plan, execute, verify, and recover across many stages. It is the default on Claude Max and the strongest model available on Claude Pro.
The practical case for Opus is straightforward: some tasks are simply too consequential, too ambiguous, or too structurally complex for a mid-tier model. Software architecture decisions, complex debugging across large codebases, financial and legal analysis, scientific reasoning, computer-use workflows, and autonomous coding agents that must verify their own outputs all fall into this category. When a failure is expensive – in time, in money, or in downstream consequences – the marginal cost of Opus becomes easy to justify.
Opus 5 also introduces configurable effort levels, which make model selection less binary. Running Opus at reduced effort can be faster and less expensive than the default, while increasing effort extracts more from its reasoning capacity on the hardest problems. In practice, this means you can tune the model’s behavior to the task rather than treating tier selection as an all-or-nothing decision.
Choose Opus 5 when the task requires deep reasoning or nuanced judgment, you are working across a large codebase or complex system, accuracy matters more than minimizing cost, the workflow involves scientific or mathematical analysis, an autonomous agent must plan and verify its own work, or Sonnet performs well on routine cases but fails on your hardest evaluation examples.
Claude Fable 5: Maximum Widely Available Capability
Fable 5.1 is Anthropic’s most capable generally available model, designed for ambitious coding, complex knowledge work, and long-running agentic tasks. It sits above the traditional Opus tier and brings Mythos-level underlying capabilities to a broadly available model, with additional safeguards for sensitive cybersecurity, biology, and chemistry use cases.
Where Fable 5 was built to sustain unusually long and difficult projects, Fable 5.1 pushes further on execution quality and autonomy. It is designed for work that can take hours or days and span multiple tools and applications: large codebase changes, deep research, complex analysis, multi-stage enterprise workflows, browser-based tasks, and autonomous agents that need to recover from failures and continue working with minimal supervision. Anthropic also highlights improved vision capabilities for understanding charts, diagrams, tables, PDFs, and for visually verifying the results of coding work.
For software engineering, Fable 5.1 is positioned as Anthropic’s strongest model for ambitious coding projects, including code review, performance optimisation, features that span an entire codebase, and multi-day autonomous development sessions. It can create its own tests, verify its work, and focus on solving underlying problems rather than taking shortcuts that merely address symptoms. Anthropic reports new best results across coding, knowledge work, and long-running problem-solving evaluations.
The business case has also become more interesting. Fable 5.1 retains Fable’s headline API pricing of $10 per million input tokens and $50 per million output tokens, compared with $5/$25 for Opus 5. However, cache reads now cost just $0.25 per million tokens – 75% less than with Fable 5. Anthropic estimates this can reduce the cost of typical Fable workloads by roughly 25% and highly agentic workloads by up to 45%. This makes the comparison with Opus 5 less straightforward than headline token pricing alone suggests.
The right question, therefore, is not simply whether Fable 5.1 is more capable than Opus 5. It is whether its stronger reasoning, autonomy, verification, and ability to sustain long-running work produce enough additional value for your workload to justify the higher base price. For routine knowledge work and many everyday coding tasks, Opus 5 may still offer the better cost-performance balance. For difficult tasks where failure, repeated supervision, or multiple retries are expensive, Fable 5.1 can be the stronger economic choice.
One important limitation remains. Fable 5.1 uses additional safeguards because its underlying capabilities are advanced enough to create elevated risks in cybersecurity, biology, and chemistry. These safeguards are more precise than in Fable 5: Anthropic says benign biology requests trigger interventions 85% less often, and Fable 5.1 can now identify vulnerabilities directly from source code. However, dual-use biology and chemistry requests may still be routed to Opus 5, while cybersecurity requests caught by the safeguards typically fall back to Opus 4.8. Certain activities, including penetration testing, exploit generation, and binary-based vulnerability scanning, remain restricted. API customers need to configure fallback behaviour through Anthropic’s Fallback API.
Choose Fable 5.1 when you need Anthropic’s highest generally available capability; the task is unusually complex, long-running, or open-ended; an agent needs to operate across many stages or tools with limited supervision; or the cost of mistakes and repeated intervention outweighs the higher model price. For organisations considering it at scale, the deciding factor should be evaluation on real workloads: if Fable 5.1 consistently completes difficult tasks that require retries, escalation, or human intervention with Opus 5, the premium can be justified.
A Note on Claude Mythos 5
Mythos 5 uses the same underlying model as Fable 5. The difference is in safeguards and access. Where Fable makes most Mythos-class intelligence broadly available while restricting or rerouting sensitive requests, Mythos provides vetted organizations with fewer restrictions for approved work – initially deployed for cyber defenders and critical infrastructure providers, with plans for trusted access in biological research.
Mythos is not a conventional fifth option in a model picker. It is a specialized deployment for organizations with validated requirements and an established relationship with Anthropic. If your business needs it, the access process makes that clear.
Extended Thinking and Effort Controls
Extended thinking allows Claude to perform additional computation woto break through a problem before answering – breaking it down, plan an approach, evaluate alternatives. The implementation varies by model: Fable 5 uses always-on adaptive thinking; Opus 5 and Sonnet 5 use adaptive thinking with a configurable effort parameter; Haiku 4.5 supports extended thinking but not the newer effort parameter.
For practical purposes: extended thinking adds real value for mathematics, debugging, architecture, detailed document analysis, and correctness-critical work. For straightforward drafting or classification, the additional computation typically adds cost without improving the result. Calibrating effort to the task is itself a form of model selection.
The Multi-Model Architecture
Most businesses eventually discover that a single model for all tasks is a blunt instrument. The more durable design is a tiered architecture that matches model capability to task complexity across your workload.
Haiku handles high-volume, predictable requests efficiently. Sonnet handles the broad middle of professional tasks – coding, writing, analysis, communication – at practical speed and cost. Opus or Fable handle the small proportion of requests that genuinely require more: the complex debugging session, the nuanced research synthesis, the autonomous agent running an extended workflow.
An efficiency-first team starts with Haiku 4.5 and upgrades only where evaluations reveal capability gaps. A capability-first team starts with Opus 5, refines the workflow, then moves suitable tasks to less expensive models as the workload becomes better understood. Both approaches reach the same destination: a tiered system where cost tracks complexity rather than defaulting to the same tier for everything.
The Bottom Line
For most professional teams, Claude Sonnet 5 is the right everyday starting point. It combines strong coding, analysis, writing, vision, and tool use with practical speed and pricing.
When the work involves deep reasoning, scientific analysis, complex software engineering, or autonomous workflows where failures are expensive, Claude Opus 5 is the stronger choice. It replaces Opus 4.8 as Anthropic’s current high-end workhorse and offers a compelling balance between frontier intelligence and cost.
When evaluations show that Opus 5 is not enough – and when latency and price are secondary concerns – Claude Fable 5 offers maximum, widely available capability for long-running agents and demanding knowledge work. Validate the advantage before committing to the cost differential.
For fast, high-volume, predictable workloads, Claude Haiku 4.5 remains the economical standard.
The right way to choose is ultimately empirical. Build a representative evaluation set from your actual workload. Test the least expensive plausible model first. Measure quality and failure modes. Move upward only where the added capability creates measurable value. That discipline – rather than instinct or marketing – is what separates teams that scale AI efficiently from those that spend heavily and measure little.
Model availability, pricing, safeguards, and plan limits can change. Consult Anthropic’s official documentation before making production or purchasing decisions.