DROPGoogle launches Gemini 3.8 Flash TTS and Flash-Lite TTS///DROPGemini Omni brings 1080p generation to Google Vids///BREAKClaude Opus 5.5 is here: 1M context, $4 input / $20 output///BREAKOpenAI adds GPT-6 Sol and Luna to the GPT-6 family///PAPEROpus 5.5's system card documents gains—and evaluation limits///BREAKGrok 4.7 launches at $2 input / $6 output per million tokens///DROPGrok Voice Transcribe 2.0 targets difficult real-world audio///TRENDGoogle steers new Gemini API projects to 3.8 Flash or 3.5 Flash-Lite///DROPGoogle launches Gemini 3.8 Flash TTS and Flash-Lite TTS///DROPGemini Omni brings 1080p generation to Google Vids///BREAKClaude Opus 5.5 is here: 1M context, $4 input / $20 output///BREAKOpenAI adds GPT-6 Sol and Luna to the GPT-6 family///PAPEROpus 5.5's system card documents gains—and evaluation limits///BREAKGrok 4.7 launches at $2 input / $6 output per million tokens///DROPGrok Voice Transcribe 2.0 targets difficult real-world audio///TRENDGoogle steers new Gemini API projects to 3.8 Flash or 3.5 Flash-Lite///
DARC.DEV

[ SIGNAL — PAPER · published 22 Sept 2026 ]

Opus 5.5's system card documents gains—and evaluation limits

The new system card covers agentic coding, computer use, alignment evaluations and deployment safeguards. Anthropic also reports weaknesses, including malicious instructions pasted into user prompts. Reported scores are not guarantees of safety in your application.

← Back to the newsroom
Source-checked Sources & corrections

The practical read

Read the caveats alongside the headline benchmarks. A more robust model does not replace tool isolation and approval gates.

Editorial interpretation, separate from the provider announcement

Who should pay attention

Agent builders reviewing safety and deployment constraints

One useful next step

Read the system-card limitations alongside the capability tables and turn relevant failure modes into application tests.

Before you act on the headline

  • Distinguish automated evaluation outcomes from deployment guarantees.
  • Test malicious content pasted or retrieved into context.
  • Keep application permissions independent of model cooperation.

Original source

Read Anthropic

The date above belongs to the original announcement, not this site's build time. Prices, rollout status and provider claims should be read with the source's conditions. This briefing does not claim an independent product test.