Claude (Anthropic)

Overview

Claude is Anthropic's family of models, offered through claude.ai, the Claude API, Claude Code, and cloud platforms (Amazon Bedrock, Google Vertex AI, Microsoft Foundry, and Claude Platform on AWS). The lineup has tiers trading capability against price and speed. Claude is especially strong at coding, agentic work, and long documents.

Key concepts

Tiers: Fable (most capable), Opus (default flagship), Sonnet (balanced, cheaper), Haiku (smallest, fastest). Mythos models run only in restricted programmes.
Model IDs: short aliases such as claude-opus-5-5; no date suffix on current models.
Adaptive thinking and effort: thinking is on by default; output_config.effort controls depth (low to max). On Fable 5.1 and Opus 5.5 thinking is always on and cannot be disabled. Haiku 5.5 accepts thinking: {type: "disabled"} at effort high or below, and Sonnet 5.5 turns off up-front thinking with {type: "between_tools"} (also at high or below). Default effort is high on Fable 5.1 and Sonnet 5.5, medium on Opus 5.5 and Haiku 5.5. Thinking blocks are tied to the model and conversation that produced them ("preserved thinking").
Tool use: client tools you run, and server tools Anthropic runs (web search, web fetch, code execution, tool search). Fable 5.1, Opus 5.5, and Sonnet 5.5 reject forced tool_choice (any / tool).
No prefill: the newest generation no longer accepts a prefilled assistant turn; use structured outputs or instructions instead.
Prompt caching: reuse of stable prompt prefixes at a large discount.
Batch API: asynchronous processing at about half price.
Structured outputs: JSON responses guaranteed to match a schema.
Agent features: Skills, an MCP connector, memory tool, compaction, and Managed Agents (Anthropic-hosted agent loop and sandbox).
Refusals: safety classifiers can end a response with stop_reason: "refusal"; server-side fallbacks can retry on another model.
Cyber access: cyber safeguards are relaxed for verified security professionals through the Cyber Verification Program, which has three tiers (Defense, Red Team, Specialized) and covers Opus 5.5, Sonnet 5.5, and Mythos 5.1. On 2026-10-06 it absorbed Project Glasswing and became the only route to Claude Mythos 5.1.

How to use it

Need	Model	Price per 1M tokens (input / output)
Hardest reasoning, long-horizon agents	Claude Fable 5.1 (`claude-fable-5-1`)	$10 / $50
Default for most demanding work	Claude Opus 5.5 (`claude-opus-5-5`)	$4 / $20
Everyday coding, agents, high volume	Claude Sonnet 5.5 (`claude-sonnet-5-5`)	$2 / $10
Classification, extraction, routing, high volume	Claude Haiku 5.5 (`claude-haiku-5-5`)	$0.10 / $0.50 (prompts ≤100K tokens); $0.50 / $2.50 above

All four current models have 1M context, 128K output, and a June 2026 knowledge cutoff. Haiku 4.5 (200K context, 64K output, $1 / $5) is superseded but still listed as Active, with retirement "not sooner than 2026-10-15" and no deprecation notice as of 2026-10-13. Sonnet 4.5 retires on 2026-11-30.
Haiku 5.5 (released 2026-10-07) is the first Haiku with effort control and supports computer and browser use. Claude Code 2.1.293 made it the default Haiku.
Haiku 5.5 is the only current model priced by prompt length; the others bill the full 1M window at one rate. Cache reads on Opus 5.5 and Sonnet 5.5 cost 5% of input.
Stream long requests. Use the Models API for live capability data.
Try lower effort on a strong model before adding a cascade of smaller models.

Current state

2026-10-15: Anthropic gave six METR staff "employee-like access" (hardware, office, internal tools) with the right to publish without approval, apart from redactions in three fixed categories. SecureBio is reviewing the CBRN chapter of the August 2026 Risk Report; Accenture and Faculty are negotiating broader evaluation work. See evals for the access tiers.
2026-10-14: Anthropic reports that Opus 5.5, Sonnet 5.5, and Haiku 5.5 score 99% on its political evenhandedness evaluation and refuse fewer political requests. Later in October, US Claude.ai users get state-specific official voting information and election-night results in the chat (data from Democracy Works). External reviewers still flag unsolicited caveats. See security for misuse findings.
2026-10-14: Anthropic opened a Singapore office (its fifth in Asia-Pacific) and signed an MOU with Singapore's MDDI covering joint model evaluation and testing standards. Regulated customers there use in-region inference on Amazon Bedrock.
2026-10-13: The deprecations page still lists Haiku 4.5 as Active with no deprecation notice. Anthropic gives at least 60 days' notice, so it cannot retire before mid-December 2026. Sonnet 4.5 (deprecated 2026-09-30) retires on 2026-11-30; Opus 4.5 is active until at least 2026-11-24.
2026-10-13: The thinking troubleshooting docs confirm that Haiku 5.5 can run with thinking disabled, but only at effort high or below (xhigh and max return a 400), and effort cannot change per message while it is disabled.
2026-10-08: Anthropic published a revised Usage Policy, effective 2026-11-12. It adds a section on fake account networks, narrows the elections rules to voter deception (personalised campaign targeting is now allowed), spells out weapons, surveillance, and law-enforcement limits, and bans sustained cruelty towards models in extreme cases.
2026-10-08: Anthropic launched its Cyber Mission (infrastructure defence programme and free OSS Scanner; see security) and committed $150 million over three years to the US Genesis Mission, giving Claude and Claude Code to 15+ federal agencies and several hundred research projects.
2026-10-07: Claude Haiku 5.5 released, completing the 5.5 family (details above). Anthropic reports 72.4% on OSWorld 2.1 (offline subset) and 39.2% on Terminal-Bench 4.0. Sonnet 5.5 cache reads fell to 0.05x input, matching Opus 5.5.
2026-10-06: With the Glasswing merger, Anthropic said Glasswing partners found over 129,000 verified vulnerabilities from April to July 2026.

Open questions

Simon Willison reports new monthly API credits for Max and Team subscribers ($100 to $500) announced with Haiku 5.5; no Anthropic primary source found on 2026-10-08.
Simon Willison reported that reasoning cannot be disabled on Haiku 5.5, but Anthropic's docs (2026-10-13) say it can at effort high or below. He may have tested at xhigh or max; the docs are treated as authoritative here.
Haiku 4.5 retirement date: not yet announced (checked 2026-10-13). Check the deprecations page for a notice.