Anthropic announced Claude Opus 4.6 and Sonnet 4.6 on February 18, 2026, positioning the release around practical reliability for production AI workloads. The company said both models preserve hybrid behavior, allowing fast responses for low-latency interactions and extended thinking for deeper reasoning tasks. That framing matters for teams deploying AI agents at scale: in enterprise settings, consistency over long task chains is often more valuable than isolated benchmark spikes.

In its release notes, Anthropic described Opus 4.6 as its top coding model and reported state-of-the-art performance on SWE-bench (74.5%) and Terminal-bench (47.4%) as of February 17, 2026. Sonnet 4.6 was presented as a high-efficiency option with strong capability-per-cost for broad deployment. Anthropic also highlighted reduced shortcut behavior in coding workflows, improved handling of large codebases, and better coherence during complex multi-step reasoning. These are meaningful operational claims for engineering teams that care about fewer retries and safer automation handoffs.

Security was a central element of the announcement. Anthropic introduced dynamic threat prevention protections aimed at reducing risks from prompt injection, data exfiltration, and agent misuse patterns. The company also stated that Opus 4.6 remains deployed with ASL-3 safeguards first applied to Opus 4.5. This continuity in safety posture is notable because many enterprise buyers evaluate model updates not only on capability gains, but also on whether security controls and risk governance remain stable across versions.

The release further expanded workflow capacity by raising the extended output limit to 64k tokens, a change relevant for long-form analysis, code refactoring, and document-heavy agent tasks. Anthropic said the new models are available through the Anthropic API, Amazon Bedrock, Google Cloud’s Vertex AI, and Claude products. Taken together, the launch suggests that leading-model competition in 2026 is increasingly defined by a bundle of factors: coding depth, long-context robustness, and deployable security controls across multiple cloud channels.

#model-safety

Anthropic Releases Claude Opus 4.6 and Sonnet 4.6 with Stronger Coding Reliability