Anthropic and the Claude models

What the lab is

Anthropic is a closed-weights frontier lab whose public identity rests on safety research as much as on model capability. Its models ship through its own API and through AWS, Google Cloud, and Microsoft Azure. Parameter counts, training compute, and dataset sizes are not disclosed.

Current flagship models

Model	Release	Notable	Source
Claude Fable 5.1	2026-09-01	Flagship; 1M context, 128K output; $10/$50 per million tokens	Anthropic post; MarkTechPost
Claude Mythos 5.1	2026-09-01	Same weights as Fable 5.1, fewer domain safeguards; vetted access only	System card
Claude Opus 5	2026-07-24	Prior top general model	Wikipedia
Claude Sonnet 5	2026-06-30	Mid-tier	Wikipedia
Claude Haiku 4.5	2025-10-15	Small tier	Wikipedia

The release pattern to note is the split. Fable and Mythos are one model with different safeguard layers. Fable is generally available and blocks certain biology and cybersecurity tasks; Mythos relaxes those blocks for participants in the Cyber Verification Program and Life Sciences Verification Program, currently US organizations only. Anthropic's own benchmark table shows the gap between them (55.8% vs 60.9% on Terminal-Bench 4.0) and attributes it to safeguard interventions, not model differences. The knowledge cutoff is June 2026.

How the training approach differs

Data. The system card states the models were trained on "a proprietary mix of publicly available information from the internet, public and private datasets, and synthetic data generated by other models," collected in part by the ClaudeBot crawler. No token counts are given.

Constitutional lineage. Anthropic's 2022 Constitutional AI paper introduced RL from AI feedback: a model critiques and revises its own outputs against written principles, and a preference model trained on AI-labeled comparisons replaces most human harmlessness labels. The current system card says post-training aims to align behavior "with the values described in Claude's constitution." In January 2026 Anthropic replaced its list of standalone principles with a longer document that explains reasoning, on the stated view that models need to understand why in order to generalize. Whether the 2022 RLAIF mechanics are still the primary method is not documented; the lab describes the target, not the recipe.

Reasoning. Fable 5.1 uses "adaptive thinking" that is always on, with Low, Medium, and High effort settings that trade cost for accuracy. Anthropic publishes accuracy-versus-cost curves at each effort level rather than single scores. The system card also reports that Mythos 5.1 can control the contents of its extended thinking more reliably than earlier models, which Anthropic flags as weak evidence that chain-of-thought monitoring may get harder.

Interpretability in the release loop. The alignment assessment used Natural Language Autoencoders, a method that trains one copy of the model to describe an activation in text and another to reconstruct the activation from that text. Anthropic used it to look for unverbalized evaluation awareness. This is the clearest case of interpretability research feeding a deployment decision rather than staying a research artifact.

Anti-distillation. New API accounts can no longer edit prior context while preserving Claude's thinking transcript, a change Anthropic frames as closing a documented distillation route.

Open questions

• Anthropic reports 30-day retention by default and a customer-hosted "Enterprise Frontier Safeguards" system arriving this fall. How much monitoring is required to ship a Critical-adjacent model is not yet clear.
• Third-party parameter estimates circulate but Anthropic has not confirmed any. Omitted here.

Review log

• 2026-09-08: created; sources checked on this date.