The best Claude model is not automatically the most capable one. In 2026, Anthropic’s lineup gives buyers at least three meaningfully different decisions: Sonnet 5 for efficient everyday agentic work, Opus 5 for stronger professional reasoning at a mid-tier price, and Fable 5 for the hardest work where the remaining capability margin justifies a frontier premium.
That distinction matters because model cost is only one part of the bill. A cheaper model can become expensive when it needs repeated prompts, human repair or a second model to finish the task. A frontier model can also waste money when the job is simple. The right comparison therefore measures cost per accepted outcome, latency, reliability and risk—not only benchmark position.
Anthropic’s own pages for Sonnet 5, Opus 5 and Fable 5 provide the current first-party descriptions. Prices, plan limits and safeguards can change, so verify them at the point of purchase.
The short answer
Choose Sonnet 5 when speed, volume and predictable cost matter more than extracting the final few percentage points of reasoning quality. Choose Opus 5 when difficult coding, research or multi-step professional work is common enough that fewer failures can repay the higher token rate. Choose Fable 5 only when you can identify tasks where it reliably succeeds and the other models do not.
| Workload | Recommended starting model | Why | Escalate when |
|---|---|---|---|
| High-volume extraction and classification | Sonnet 5 | Low cost, strong tool use, fast responses | Validation failures exceed the accepted threshold |
| Everyday coding and repository work | Sonnet 5 or Opus 5 | Sonnet for routine changes; Opus for complex plans and debugging | Tests repeatedly fail or cross-file reasoning breaks down |
| Research across conflicting documents | Opus 5 | Better fit for synthesis and constraint tracking | Evidence remains contradictory or consequences are high |
| Long-running autonomous agent | Opus 5 with strict controls | Stronger planning can reduce compounding errors | The task reaches a high-risk tool or decision |
| Frontier scientific or strategic analysis | Fable 5 | Highest capability tier may matter at the difficult edge | Human expert review is still mandatory |
| Customer-facing low-latency assistant | Sonnet 5 | Responsiveness and consistency often dominate | Confidence or policy triggers require escalation |
| One-off executive brief | Opus 5 | Quality and reduced rework can outweigh modest extra cost | The subject requires specialist judgement |
This matrix is a starting policy, not a universal ranking. The same organisation may route document extraction to Sonnet, code migration to Opus and a small number of frontier research questions to Fable.
Sonnet 5: the throughput choice
Anthropic describes Sonnet 5 as a more agentic Sonnet model, able to plan, use browsers and terminals, and operate autonomously at a level that previously required larger models. Its role is broad: it is designed to be capable enough for serious work while remaining efficient enough for frequent use.
That makes Sonnet the natural default for workflows with a large number of moderately difficult requests. Examples include classifying support cases, transforming documents, generating routine code, checking known policies and preparing first drafts from structured evidence.
The risk is assuming that “default” means “safe to automate without controls”. A faster model can make mistakes faster. Tool permissions, data access, test coverage and approval gates still determine how much damage a failed action can cause.
Sonnet is also a strong baseline for evaluation. If Sonnet already meets the quality threshold on 95% of a workload, a more expensive model must create value in the remaining 5% or reduce supervision enough to justify the change.
Opus 5: the professional-work choice
Opus 5 occupies the middle in economic terms, but it is not a compromise model. Anthropic positions it close to Fable on many coding and knowledge-work evaluations, with a lower price and a Fast mode for latency-sensitive use.
Its most compelling use case is work where failures compound. A weak first plan can cause an agent to inspect the wrong files, call unnecessary tools and produce a plausible but unusable result. Better reasoning at the beginning can reduce the total number of steps. That is why cost per token can be misleading.
Opus also supports an effort control. A team can use moderate effort for normal cases and reserve higher effort for difficult ones. This helps only when the application can recognise difficulty or failure. Running every task at maximum effort merely creates a new form of over-provisioning.
For many professional teams, Opus 5 should be the first escalation tier rather than the universal default. It can handle difficult cases while leaving routine volume on Sonnet.
Fable 5: the frontier-margin choice
Fable 5 is Anthropic’s highest capability tier in this comparison. The rational reason to use it is not prestige. It is a measurable capability margin on tasks that matter.
Those tasks may include difficult scientific reasoning, unusually complex software architecture, deep analysis of conflicting evidence or long-horizon agents where small planning improvements prevent expensive failure. The organisation should be able to show a set of cases where Fable’s accepted-outcome rate is higher than Opus’s.
Fable also carries stricter safeguards in some sensitive domains. That can reduce misuse risk while creating refusal or fallback behaviour that legitimate specialist users must understand. A buyer should test not only whether the model can solve the task, but whether the safety layer permits the intended workflow and leaves a clear audit trail.
If Fable cannot demonstrate a reliable advantage on your own evaluation set, it should remain an exception rather than a default.
Price is not the same as cost
Published API prices provide a useful baseline. The total workflow cost adds several other components:
input tokens, including repeated context and tool descriptions;
output tokens and reasoning effort;
prompt caching or batch discounts when available;
failed tool calls and retries;
human review and correction time;
external APIs, browser sessions and compute;
the cost of an incorrect or delayed outcome.
Consider two models processing 10,000 monthly tasks. Model A costs £0.08 per request and passes 70% on the first attempt. Model B costs £0.16 and passes 92%. If each repair takes five minutes of staff time, Model B can be cheaper despite doubling the API bill. If both pass 99%, Model A probably wins.
This is why a model-selection spreadsheet needs acceptance rate and rework time. Token price alone cannot make the decision.
Coding: which Claude model is best?
For isolated functions, standard tests and well-scoped edits, Sonnet 5 is likely to offer the best balance. It can plan and use tools while keeping latency and cost under control.
Opus 5 becomes more attractive when a task crosses architectural boundaries: tracing an intermittent failure, migrating a large dependency, understanding undocumented behaviour or coordinating changes across many files. The expected advantage is fewer wrong turns, not simply more verbose code.
Fable 5 should be reserved for a test set of genuinely hard problems. A model can appear brilliant on a benchmark and still misunderstand a company’s conventions. Repository-specific evaluation remains decisive.
Whichever model is used, run it in a constrained environment. Limit credentials, isolate untrusted code, require tests and ask for approval before production or external actions. Stronger coding ability increases the need for dependable controls.
Research and executive analysis
Sonnet can handle structured summaries, extraction and known-format comparison. It is well suited to the first pass over a large document set.
Opus is better positioned for synthesis: identifying contradictions, maintaining qualifiers, comparing scenarios and drafting an argument that stays within the evidence. It still needs reader-visible sources. A confident paragraph without a traceable source is not improved research.
Fable may add value when the problem is genuinely novel or requires deeper reasoning across disciplines. The responsible pattern is to pair it with specialist review rather than treating it as a substitute for expertise.
An executive brief can use a tiered workflow: Sonnet extracts, Opus synthesises, and a human owner approves. Fable enters only when the difficult cases justify it.
Agents and tool use
Agentic tasks expose the largest difference between a chat response and completed work. The model must decide what to do, call tools, inspect results and recover when the environment disagrees.
Sonnet is appropriate when the action space is narrow and reversible. Opus is preferable when planning complexity is high or failures are expensive. Fable may be useful for frontier research agents, but its capability does not remove operational limits.
Set tool permissions independently of model choice. The agent should have the minimum scope required. A model upgrade must not silently inherit broader production access. Approval gates should follow the action’s consequence, not the provider’s benchmark score.
The recent attention around agent containment is a useful reminder: intelligence and authority are different variables. Our AI agent permissions checklist provides a practical control set.
A five-step selection test
Define the business threshold: accuracy, latency, cost and acceptable failure.
Build a representative test set from real tasks, including difficult and adversarial cases.
Run Sonnet, Opus and Fable with the same instructions and tools.
Score blind where possible, then calculate cost per accepted task.
Assign a default and explicit escalation rules.
Do not choose from one impressive answer. Repeat tests because agentic performance can vary. Measure the median and the failure tail. A model that usually excels but occasionally takes an unsafe action may be less suitable than a slightly weaker but more predictable option.
Recommended routing policy
A sensible first policy is:
Sonnet 5 for low- and medium-consequence routine work;
Opus 5 for difficult professional tasks, complex coding and evidence synthesis;
Fable 5 for a small, measured set of frontier cases;
human approval for high-consequence decisions and external actions.
Review the policy monthly while the release cadence remains rapid. A new model can change the cost frontier, but the evaluation method should remain stable.
The verdict
Sonnet 5 is the best default for throughput. Opus 5 is the best general choice when quality and reduced rework matter more than the lowest token price. Fable 5 is the best specialist option when its additional capability is proven on tasks where failure is expensive.
The most mature answer to “Which Claude model is best?” is therefore a routing table. Use the least expensive model that reliably clears the threshold, escalate based on evidence, and keep tool authority constrained at every tier. For the release details behind the middle option, read our Claude Opus 5 explainer.
