Anthropic Claude API Models and Platforms 2026

Which Claude model fits which job?

Anthropic’s current models overview lists four general-access options: Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5. They all accept text and images, return text, support tool use and list a 1-million-token context window. The differences are mainly capability, speed and the type of workload Anthropic recommends for each.

• Fable 5.1: the most capable model open to all customers, aimed at demanding reasoning and long-running agent work. It is the slowest option in the lineup, so reserve it for tasks where extra capability can justify the wait.
• Opus 5.5: designed for extended agentic coding and knowledge work. Anthropic suggests starting with it for most workloads; its comparative latency is moderate.
• Sonnet 5.5: the speed-and-intelligence balance for everyday use. It is fast, making it a sensible starting point for coding assistants and general-purpose product features.
• Haiku 5.5: the fastest tier, intended for high-volume, latency-sensitive tasks such as ticket classification, extraction and routing. Start here when efficiency matters more than handling the hardest cases.

Mythos 5.1 is not a fifth general option. Anthropic limits it to organisations verified through programmes such as its Cyber Verification Program. The overview also lists older models, including Opus 4.6 and Sonnet 5, as available legacy choices; do not confuse those with the current four-model lineup.

Match the model to your workload

For a daily coding assistant or IDE agent, try Sonnet 5.5 first and check the results against your own codebase. Move to Opus 5.5 when sessions need to plan and act over many steps, such as a multi-hour refactor or a complicated investigation. Fable 5.1 is the option for especially demanding reasoning or long-horizon agents, but test cost and rate limits before making it your default.

For bulk work, route routine tasks to Haiku 5.5 and escalate uncertain or consequential decisions to Opus 5.5 or Fable 5.1. This executor pattern avoids spending the strongest model’s capacity on every simple classification. An orchestrator can also divide a large job among cheaper workers and use a stronger model to review or combine the results.

Before switching models, test the effort setting. Anthropic’s selection guidance says adjusting effort can be a better lever than changing models. Build a small evaluation set from real inputs, then compare accuracy, response time and the amount of review your team must do. A model that appears faster may cost more in staff time if it needs frequent correction.

Direct access or a multi-model platform?

Calling Anthropic directly gives you a relationship with the model maker and access to its Claude-specific API surface. That route is worth considering if your application depends on particular thinking controls, prompt caching, batch processing, server tools or a direct data-retention arrangement. It also removes an additional API provider from your request path.

A multi-model platform can make sense when your team already uses models from several makers. One key, one billing relationship and a common integration may simplify account management, while routing can let you use a different model family as a fallback. Those are operational gains, not a promise of lower Claude costs or identical feature support. Compare the actual usage terms and test the requests your application sends.

Our filtered comparison lists AI/ML API, CometAPI, Eden AI, PiAPI, WaveSpeedAI and BeeSwitch as platforms offering Anthropic models. Across all 7 plans and 7 providers in our full AI comparison, listed monthly prices run from $35.00 to $35.00; those are whole-comparison figures, not prices for this Claude filter. We last synced prices on 31 October 2026. Check each platform’s current model list and billing details before you commit.

Check feature support and model IDs

Do not assume a platform exposes every Claude feature just because it lists a Claude model. Adaptive thinking controls, effort, tool choice, server tools such as web search, prompt caching and the Message Batches API can be absent or behave differently. If your code relies on one of them, send a representative request through the platform and inspect the response and errors before migration.

Model IDs also change between routes. Anthropic’s API IDs for the four current models are claude-fable-5-1, claude-opus-5-5, claude-sonnet-5-5 and claude-haiku-5-5. A platform may use its own alias; Amazon Bedrock, Google Cloud and Microsoft Foundry use their own naming or deployment formats as well. Keep model identifiers in configuration, not scattered through application code, so you can change routes deliberately.

Anthropic’s Models API can return each model’s maximum input and output tokens and a capabilities object. That is useful if you build a router that checks support before choosing a model. Still, test settings such as disabled thinking and tool choice per model: supported options are not necessarily interchangeable.

Context, output and thinking settings

All four current models list a 1-million-token context window and a maximum output of 128,000 tokens on the synchronous Messages API. A large context does not mean every prompt should be filled to the limit. Long histories and retrieved documents increase the amount of input your application handles, so trim irrelevant content and measure the quality difference before sending everything.

All four accept image input as well as text, and all can use tools. Their default thinking and effort settings differ: adaptive thinking is always on for Fable 5.1 and Opus 5.5; default effort is high for Fable 5.1 and Sonnet 5.5, and medium for Opus 5.5 and Haiku 5.5. Check the model-specific documentation when changing settings rather than copying one request body across the lineup.

Fast mode is available on supported models for up to 2.5 times higher output speed at a premium. It may help a user-facing interaction where latency matters, but benchmark it against your normal mode on your own prompts. Faster generation does not automatically mean lower end-to-end latency if tools, network calls or application processing dominate.

Data retention: check both sides of the route

Anthropic says API inputs and outputs are deleted from its backend within 30 days by default, subject to exceptions. Those include Files API use, legal requirements, and Usage Policy enforcement. For flagged chats under that enforcement, the stated retention can be up to two years; trust-and-safety scores can be retained for up to seven years. Zero data retention is not the default: it requires an agreement.

Covered Models have limited data retention and may be reviewed as part of safety work, but Anthropic’s privacy information does not give a retention period for that category. Check the applicable terms for the exact model and account rather than applying the default rule to every situation.

A platform route adds another layer to review. Read its logging and data-collection settings, confirm whether requests can be sent to fallback models, and ask what happens to prompts and outputs on those routes. For sensitive customer data, compare a direct Anthropic zero-data-retention agreement with a platform’s account-level controls and confirm the specific model’s policy in writing.

Ways to control cost without changing providers

Batch processing and prompt caching are potential savings tools on the direct Claude API, but confirm that your chosen route supports them and check current pricing before estimating a budget. Anthropic lists the Message Batches API at 50% off base price. Its overview also gives prompt-cache read rates as a fraction of base input price: 10% for Haiku 5.5, 5% for Opus 5.5 and Sonnet 5.5, and 2.5% for Fable 5.1. These percentages are not a substitute for checking the live pricing page.

Batch is suited to work that does not need an immediate answer, such as overnight extraction or evaluation. Caching can help when repeated requests share a stable prompt prefix. Model routing is another lever: send straightforward work to Haiku, then escalate cases that fail confidence or quality checks. Compare total spend and human review time, not only the model selected for each request.

Model availability and retirement dates can differ by route. Anthropic’s overview gives the current four models retirement commitments no sooner than dates between 1 September and 7 October 2027, while partner platforms set their own dates. If a production workflow depends on a particular ID, confirm its availability and migration notice with the platform you plan to use.