Which Claude model should your business use?
Anthropic's current Claude API comparison spans four models, while availability in Claude apps, Claude Code and partner platforms varies. This guide explains what changed, how we route Claude alongside other systems, and how to add it to a real business without granting more access than the job needs.
Direct answer: As of 7 October 2026, Anthropic's current Claude API comparison lists Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 4.5. Access inside Claude apps, Claude Code and partner platforms depends on the account, plan, rollout and platform. Start from the job, then test the exact available surface.
That answer is useful, but it does not settle the setup. A model can be right for a task and still be the wrong way to deliver it. One person may need Claude chat for document work, a developer may need Claude Code inside a repository, and a company may need the API behind a controlled workflow. Those are different decisions.
Find the answer you need.
Claude models and Claude products are different choices
A model is the reasoning and generation engine. Fable, Opus, Sonnet and Haiku are Claude model lines. Versions such as Fable 5.1 and Opus 5.5 are individual models; Anthropic also uses family language for a cross-line release set such as Claude 5.5. An assistant wraps a model in a user interface, account, memory, file access and product controls. A coding agent adds a repository, shell, editing tools, tests and permissions. A platform lets a team build repeatable applications with APIs, tools, logs, workspaces and spending controls.
Anthropic offers Claude through web, mobile and desktop apps, Claude Code, the Claude API and Claude Platform. Cowork capabilities can add connected apps, files, browser work and computer use, depending on the plan and administrator settings. Do not assume that a feature available in one surface exists in another. Check the account, plan and organisation settings before designing the workflow. (Anthropic, Claude Code for Team and Enterprise, checked 7 October 2026) (Anthropic, Claude Cowork and chat, checked 7 October 2026)
The current Claude model line
Anthropic's current comparison page recommends starting most workloads with Opus 5.5, moving to Fable 5.1 for the hardest long-running work, using Sonnet 5.5 when speed and capability need a better balance, and using Haiku 4.5 for the fastest current option. All four accept text and images and return text; their context, thinking controls, output limits, latency and price differ. (Anthropic model overview, checked 7 October 2026)
| Model | Claude API ID | Best starting use | Context and output | API list price per million tokens | Important detail |
|---|---|---|---|---|---|
| Claude Fable 5.1 | claude-fable-5-1 | Demanding reasoning and long-horizon agent work | 1M input context, 128K output | USD 10 input, USD 50 output | Always-on adaptive thinking, high default effort. Public release: 1 September 2026. |
| Claude Opus 5.5 | claude-opus-5-5 | Long-running coding and knowledge work | 1M input context, 128K output | USD 4 input, USD 20 output | Always-on adaptive thinking, medium default effort. Public release: 22 September 2026. |
| Claude Sonnet 5.5 | claude-sonnet-5-5 | Well-scoped production work where speed and cost matter | 1M input context, 128K output | USD 2 input, USD 10 output | Adaptive thinking, fast relative latency. Public release: 28 September 2026. |
| Claude Haiku 4.5 | claude-haiku-4-5-20251001; alias claude-haiku-4-5 | Fast, high-volume, lower-complexity work | 200K input context, 64K output | USD 1 input, USD 5 output | Extended thinking; no effort parameter. Public release: 15 October 2025. |
Prices above cover Claude API input and output tokens. Prompt caching, Batch API and any separately priced tools use their own published rates. Fable, Opus and Sonnet list a June 2026 reliable knowledge cutoff; Haiku 4.5 lists February 2025. A large context window also does not supply current facts. Current work still needs search, retrieval or a connected source. (Anthropic model overview) (Anthropic, Sonnet 5.5 pricing details) (Anthropic, training-data recency)
Claude Mythos 5.1 shares Fable 5.1's specifications and pricing but is limited to verified access programmes. It is not part of the normal public choice for most businesses. Anthropic also announced Haiku 5.5 on 28 September as coming in the following weeks, but no official model page or release entry existed by this guide's source date. We therefore treat it as announced, not released. (Anthropic, Fable 5.1 and Mythos 5.1) (Anthropic, Sonnet 5.5 announcement)
What changed in September 2026
Three top-end releases arrived within four weeks: Fable 5.1 on 1 September, Opus 5.5 on 22 September and Sonnet 5.5 on 28 September. The successors did not increase the 1M default context or 128K standard output limits versus Fable 5, Opus 5 and Sonnet 5. The material changes were model behaviour, effort and thinking constraints, a lower Opus list price, a lower Fable cache-read price, and vendor-reported speed or token-efficiency gains. (Anthropic, Fable 5) (Anthropic, Opus 5) (Anthropic, Sonnet 5) (Anthropic Platform release notes)
Anthropic says Opus 5.5 reaches Fable 5.1's level on most work at a lower typical task cost than Opus 5, and says Sonnet 5.5 produces output faster and can use fewer tokens than Sonnet 5. Those are Anthropic's measurements, not DLVX benchmarks. We use them as a reason to test, never as proof that a model will win on a client's work. (Anthropic, Opus 5.5 announcement) (Anthropic, Sonnet 5.5 announcement)
Production migrations need more care than swapping a model ID. Fable 5.1 and Opus 5.5 require adaptive thinking and reject some forced tool-choice patterns. Opus applications should control depth through effort; Sonnet 5.5 changes how disabled thinking and between-tool thinking work. Preserved thinking blocks can also depend on the originating model, conversation and account. Test the model ID, thinking settings, tool calls, browser or computer tool version, output shape, latency, cost and stop reasons together. (Anthropic Platform release notes)
Claude Sonnet 4.5 was deprecated on 30 September 2026 and is scheduled to retire on the Claude API on 30 November 2026. Anthropic's model lifecycle policy states that publicly released models on Anthropic-operated platforms receive at least 60 days of notice before retirement; partner platforms can follow different schedules. (Anthropic model deprecations)
Which Claude model fits which job?
Use the least expensive model that passes a real task test, then escalate for complexity or risk. Do not choose by job title. One marketer may use Haiku for classification, Sonnet for a first production draft and Opus for a difficult exception. A developer can plan with Opus, implement with Sonnet and use deterministic tests before accepting either result. Anthropic itself describes planning with Opus and executing with Sonnet as a common Claude Code pattern. (Anthropic, models and limits in Claude Code)
| Workflow | Start here | Escalate when | Proof before acceptance |
|---|---|---|---|
| Everyday coding and implementation | Sonnet 5.5 | Use Opus for cross-cutting architecture or difficult debugging. Use Fable for unusually long sessions. | Tests, reviewed diff, security checks and a working build. |
| Ambiguous strategy or architecture | Opus 5.5 | Use Fable if a higher-effort Opus run still fails the task rubric. | Decision record, constraints, counter-case and accountable owner. |
| Large document extraction | Sonnet 5.5 | Use Opus for judgement-heavy synthesis and Fable for long research that must become a finished artefact. | Citations, source checks and sampled extraction accuracy. |
| Research | Opus 5.5 | Use Fable for long multi-step work; use Sonnet for bounded retrieval and synthesis. | Dated primary sources and a separate fact check. |
| Browser or computer work | Sonnet 5.5 for bounded tasks | Use Opus when navigation or exception handling needs more judgement. | Site allowlist, action log and confirmation before consequential writes. |
| Operational routines | Sonnet 5.5 | Use Opus for exceptions and Fable only for the hardest long-running cases. | Allowed-action policy, approval points, replay log and fallback. |
| High-volume triage | Haiku 4.5 | Move edge cases to Sonnet when the eval shows judgement failures. | Sampled human review and an edge-case queue. |
| Quality review | Opus 5.5 | Add a specialist or second provider for contested, high-impact decisions. | Independent rubric. Do not let the same run create and certify the work. |
How we actually use Claude at DLVX
We do not send every task to Claude, and we do not send every Claude task to the same model. Our working method routes by task. Claude is most useful to us where context, voice, trade-offs and difficult judgement matter. Kimi or Codex can take implementation or long-context work when they fit it better, while separate checks cover facts, code, security and rendered output.
There is an important date boundary. Our internal routing method was last verified in July 2026, before Fable 5.1, Opus 5.5 and Sonnet 5.5 launched. It supports the routing principle, but it does not prove that DLVX benchmarked those September releases or migrated every internal route to a new pinned model. This guide therefore separates current Anthropic product facts from our established operating method.
In practice, we keep four things portable: the source pack, instructions, acceptance test and output format. That lets us change a model without rebuilding the business process. On client work, we also separate research, implementation and review when the consequence warrants it. A strong builder should not be its only judge.
We have built private AI workflows inside wider business systems, but the model name is not the product. The useful part is the controlled connection between source data, the job, permissions and the person responsible for the result. Our Pro Pick Property case shows that wider system pattern without claiming that Claude powers the private operator described there.
Choose the simplest Claude surface that solves the job
| Surface | Use it when | Do not add it merely because |
|---|---|---|
| Claude chat or Cowork | A person needs to analyse files, compare documents, draft, prepare a meeting or work with approved connected apps. | The task happens more than once. Repetition alone does not require an API. |
| Claude Code | The job lives in a repository and needs file search, edits, shell commands, tests or a reviewed diff. | Someone asked a general technical question that chat can answer safely. |
| Connector or MCP server | Claude repeatedly needs controlled access to a real source system or business tool. | A broad integration sounds impressive. Start with the few read actions the job needs. |
| Claude API | The workflow needs repeatable structured input, an embedded interface, logs, tool calls, access rules or scale. | A manual pilot has not yet shown that the workflow works. |
| Managed agent | A long-running workflow has defined inputs, allowed actions, approvals, budget, stop conditions and escalation. | The underlying process still changes every week or nobody owns failures. |
Claude Code supports model selection, structured output, turn limits, permission modes and MCP configuration. In API tool use, the model proposes a structured tool call, but the application or Anthropic's server executes the action. That separation matters: a confident model request is not authorisation. (Anthropic, Claude Code CLI reference) (Anthropic, tool-use contract)
MCP standardises how applications expose tools and data to models. It does not make every server trustworthy. Anthropic's API MCP connector is still beta, supports remote HTTPS servers rather than local STDIO directly, and lets a developer enable or deny specific tools. Review each provider, protect credentials and start with read access. (Model Context Protocol overview) (Anthropic, MCP connector)
Three practical setup patterns
A solo operator who needs better knowledge work
Start with a paid Claude plan, one bounded project knowledge set and two repeatable tasks such as document comparison and meeting preparation. Write an input checklist and a review checklist. Add Claude Code only if work regularly touches a repository. Skip broad connectors, autonomous sends and a custom API until manual use shows a clear repeated job.
A small team that needs shared workflows
Use a managed team account, agreed project spaces, a short approved-connector list and a named owner for each workflow. Keep source permissions inherited from the original system where possible. Let Claude search, summarise, classify and draft within that boundary. Keep external messages, publishing and changes to customer records behind review until a measured pilot proves a narrower automatic path.
A product or operations team that needs integration
Prototype through the API with versioned prompts, structured outputs, a test set made from real cases and read-only staging tools. Separate development, staging and production. Log model ID, prompt version, tool request, tool result, approval, stop reason, cost and final state. Add write access one action at a time.
An enterprise or regulated setup adds data classification, region and retention review, role controls, audit evidence, contractual review and a tested incident route. A local coding agent does not mean the model inference or connected data remains local.
A safe thirty-day rollout
Days 1 to 3: define one useful job
Name the output, source of truth, data sensitivity, failure impact and person responsible. Write five to twenty real examples, including awkward cases. Decide what counts as a pass before comparing models.
Week 1: run read-only pilots
Test two or three tasks in chat or Claude Code without production writes. Record accepted results, material corrections, time, cost and failure type. Do not measure output volume. Measure whether the result was good enough to use.
Week 2: test routes, not impressions
Compare Sonnet, Opus and Haiku where the task makes sense. Use the same source pack and acceptance test. Fable belongs in the comparison only when the job is genuinely long-running or still fails with Opus at suitable effort.
Week 3: add one scoped connection
Connect one source system with read access and an allowlist. Test documents, messages or web pages containing malicious instructions because prompt injection can arrive through tool output, not only through a user's prompt. (Anthropic, mitigating jailbreaks and prompt injection)
Week 4: automate the proven path
Automate only the workflow that passed. Add spend and turn limits, named failure handling, a fallback policy, approval before consequential writes and a monthly model-change review. Claude Platform workspaces can separate environments and apply spending limits; Managed Agent sessions can also stop at a cost ceiling. (Anthropic, workspaces) (Anthropic, Managed Agent sessions)
Our default rule is simple: draft automatically, act deliberately.
The approval and evidence framework
| Action | Default control | Evidence to keep |
|---|---|---|
| Search, summarise, classify or draft | Automatic inside scoped data | Source references, model ID, prompt version and output. |
| Read confidential internal data | Role-gated and logged | Identity, fields accessed, purpose and retention. |
| Update a low-risk internal record | Review during pilot; automate only after measured success | Before and after value, rule, timestamp and rollback. |
| Send a message or publish content | Human approval unless a narrow transactional flow is already approved | Recipient or destination, approved content and durable receipt. |
| Deploy software | Tests, preview and accountable approval path | Exact source revision, build result, review and live verification. |
| Delete data, make a payment, change access or create a legal commitment | Explicit human confirmation and provider-native controls | Authoriser, exact action, provider receipt and recovery route. |
Anthropic advises task-specific evals with measurable success criteria and cases that reflect the real distribution and edge conditions. Code-based grading works where outputs can be checked exactly; human or model grading can cover judgement, provided the review standard is explicit. (Anthropic, building evaluations)
Data treatment also depends on the product. Anthropic says commercial chats and coding sessions are not used for model training unless the customer opts into its Development Partner Program. Standard API inputs and outputs are generally deleted from Anthropic's systems within 30 days, subject to product, safety, legal and contracted-retention exceptions. Anthropic also documents covered-model retention rules for specified organisations and models, including some organisations that otherwise use zero data retention. Confirm whether the rule applies to the exact model, organisation and contract. (Anthropic Privacy Centre, commercial data use) (Anthropic Privacy Centre, retention) (Anthropic Privacy Centre, covered models)
Common mistakes
- Buying the highest-priced model before defining a pass condition.
- Treating a one-million-token context as proof that every detail will be used correctly.
- Connecting email, storage, CRM and production tools before a read-only pilot.
- Letting one model research, implement and approve its own work.
- Moving a model ID without retesting thinking settings, tools, output shape, stop reasons, cost and latency.
- Calling a consumer chat setting an enterprise privacy guarantee.
- Building a general autonomous agent around a business process that nobody has made explicit.
Five questions businesses ask about Claude
What are the newest Claude models as of 7 October 2026?
As of 7 October 2026, Anthropic's current Claude API comparison lists Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 4.5. Access inside Claude apps, Claude Code and partner platforms depends on the account, plan, rollout and platform. Mythos 5.1 is limited to verified programmes, and Haiku 5.5 was announced but not released on the source date.
Which Claude model should most businesses start with?
Anthropic recommends Opus 5.5 for most workloads. We would still test the real job. Sonnet 5.5 often fits well-scoped production work at lower cost, Fable 5.1 suits the hardest long-running work, and Haiku 4.5 suits fast, simpler volume.
Should I use Claude chat, Claude Code or the API?
Use chat or Cowork for individual knowledge work, Claude Code when the task lives in a repository, and the API when a repeated workflow needs structured inputs, tools, logging, access rules or a place inside your product. Start with the simplest surface that solves the job.
Is MCP safe for company data?
MCP is a connection standard, not a security guarantee. Use trusted servers, begin with read-only tools, allow only what the job needs, keep credentials outside prompts and require confirmation before writes or destructive actions.
Does DLVX use only Claude?
No. We route different work to Claude, Kimi, Codex and specialist checking tools. Claude is most useful where context, voice, trade-offs or difficult judgement matter. The route depends on the workflow, client environment, data rules and current product behaviour.
Download the one-workflow checklist
Use the Claude model selection checklist to define one job, its source, surface, permissions, acceptance test, cost ceiling and review date before choosing a model.
Update policy
This guide was checked on 7 October 2026. We review it when Anthropic changes a named model's release status, price, context or output limit, thinking controls, tool behaviour, plan availability, retention terms or deprecation date. dateModified changes only when the advice or product facts change materially.
Sources
- Anthropic model overview, checked 7 October 2026.
- Choosing a Claude model, checked 7 October 2026.
- Claude Platform release notes, checked 7 October 2026.
- Claude model deprecations, checked 7 October 2026.
- Claude Fable 5.1 and Mythos 5.1, checked 7 October 2026.
- Claude Opus 5.5, checked 7 October 2026.
- Claude Sonnet 5.5, checked 7 October 2026.
- Claude Haiku 4.5, checked 7 October 2026.
- Claude Fable 5, checked 7 October 2026.
- Claude Opus 5, checked 7 October 2026.
- Claude Sonnet 5, checked 7 October 2026.
- Claude Fable 5.1 announcement, checked 7 October 2026.
- Claude Opus 5.5 announcement, checked 7 October 2026.
- Claude Sonnet 5.5 announcement, checked 7 October 2026.
- Claude Code CLI reference, checked 7 October 2026.
- Claude Code models, usage and limits, checked 7 October 2026.
- Claude Code for Team and Enterprise, checked 7 October 2026.
- Claude Cowork and chat, checked 7 October 2026.
- Claude training-data recency, checked 7 October 2026.
- Anthropic tool-use contract, checked 7 October 2026.
- Model Context Protocol overview, checked 7 October 2026.
- Anthropic MCP connector, checked 7 October 2026.
- Anthropic guardrail guidance, checked 7 October 2026.
- Anthropic workspaces, checked 7 October 2026.
- Anthropic rate and spend limits, checked 7 October 2026.
- Anthropic Managed Agent sessions, checked 7 October 2026.
- Anthropic evaluation guidance, checked 7 October 2026.
- Anthropic commercial data use, checked 7 October 2026.
- Anthropic organisation data retention, checked 7 October 2026.
- Anthropic covered-model retention, checked 7 October 2026.