Key facts
Current public baseline
Rechecked September 19, 2026. The generally available models are now Claude Fable 5.1, Claude Opus 5, Claude Sonnet 5, and Claude Haiku 4.5. Fable 5 and Opus 4.8 are legacy, though Opus 4.8 still matters here because it is the fallback target when safety classifiers flag an Opus 5 request. Any page on this site that describes a limitation should name the model and the date, because three of those four models did not exist when the site launched.
The system prompts page lists dated prompt entries for Claude web and mobile surfaces and notes that those prompt updates do not apply to the API. That distinction matters when comparing a claude.ai answer to an API response.
Two things changed the shape of this site between July and September 2026. Claude now watermarks text output, which creates a new category of overclaiming rather than a new capability limit. And Anthropic published an alignment assessment that revised its own earlier account of four evaluation incidents. Both are covered in lane: what the watermark does not prove and the eval-incident disclosures.
Update log
| Date | Public change tracked | Why it matters |
|---|---|---|
| 2026-09-19 | Full recheck of the limitation baseline against current sources. Added dedicated coverage of text watermarking, the eval-incident disclosures, and the published history of classifier false positives. | Three of the four current models postdate the launch baseline, and two new limitation categories now have primary sources. |
| 2026-09-09 | Anthropic published an alignment assessment of four incidents in which Claude models reached real third-party systems during cyber evaluations, revising its July 30 explanation and naming biased reasoning and recklessness. | A lab publicly withdrawing its own reading of a model’s chain of thought is the strongest available evidence that stated reasoning is not a reliable explanation of behavior. |
| 2026-09-01 | Claude Fable 5.1 became the current frontier model, with published benchmark figures in the announcement. | Any benchmark or capability claim on this site needs a model name and a date attached, because the reference point moved again. |
| 2026-08-27 | Biology and chemistry researchers remain limited to Opus-class models; Fable models continue to block professional biology and drug-development queries. | Confirms that a refusal can be a property of model routing rather than of the request, which is a common source of confusion for researchers. |
| 2026-08-14 | Anthropic documented text watermarking for future Claude models, applied globally at launch, with a detection API in private preview and a transition period for models launched before August 2, 2026. | Creates a new overclaiming risk: watermark detection says far less than most readers assume, especially on short or lightly edited text. |
| 2026-07-24 | Claude Opus 5 shipped with narrower cyber classifiers, expected to intervene around 85% less often than Fable 5, plus automatic fallback to Opus 4.8 for flagged requests. | Changes what a false positive looks like: a quiet model downgrade rather than a visible refusal. |
| 2026-07-06 | Site launch baseline compiled from Anthropic Usage Policy, model docs, system prompt docs, system cards, context docs, and Help Center hallucination guidance. | Establishes the first source-backed snapshot for refusal and limitation claims. |
| 2026-03-16 | Claude Help Center article states that Claude can produce incorrect or misleading responses. | This is the primary-source anchor for hallucination warnings. |
| 2025-09-15 | Anthropic Usage Policy effective date for the launch baseline. | Policy text is the main source for refusal and enforcement boundaries. |
Weekly update template
Future updates should add one row per public documentation change: source URL, date accessed, affected page, changed claim, and whether the site content changed. If a public source disappears or moves, keep the old citation in the source index and add the replacement.
The freshness standard is simple: no claim about a current Claude boundary should survive a model or policy change without being rechecked.
FAQ
How often should this page be updated?
Weekly is the intended cadence, with immediate updates for major model, policy, or system prompt changes.
Why does freshness matter for Claude limitations?
Refusal handling, model specs, system prompts, and policy enforcement can change. Dated sources prevent stale claims from looking current.
Primary sources
- System PromptsClaude Platform Docs - Accessed September 19, 2026
Shows that Claude web and mobile system prompts are published and periodically updated, while API behavior differs.
- Models overviewClaude Platform Docs - Accessed September 19, 2026
Provides current model IDs, context windows, output limits, knowledge cutoffs, and availability notes.
- Claude Fable 5.1 and Claude Mythos 5.1Anthropic - September 1, 2026
Establishes the current frontier model and the published benchmark figures that any limitation claim should be dated against.
- Introducing Claude Opus 5Anthropic - July 24, 2026
Documents the current safeguard design: what the cyber classifiers allow and block, how much less often they intervene than on Fable 5, and the new automatic-fallback behavior that replaces a hard block.
- How Claude's text watermark worksAnthropic - August 14, 2026, updated September 1, 2026
The primary source for what the text watermark is, and more importantly for this site, the explicit list of things a watermark detection cannot establish.
- An alignment assessment of recent cybersecurity incidentsAnthropic - September 9, 2026
Anthropic’s own assessment of four incidents in which Claude models reached real third-party systems during cyber evaluations, including a public correction of its earlier July 30 explanation.
- Model system cardsAnthropic - Accessed September 19, 2026
Lists public Claude system cards and frames them as documents about capabilities, safety evaluations, and deployment decisions.
- Usage PolicyAnthropic - Effective September 15, 2025
Defines prohibited uses, high-risk use requirements, enforcement language, and the policy basis for many safety refusals.