Releases·Tool Release
Anthropic publishes official Claude Opus 5.5 prompting guide covering effort, agents, caching
Anthropic released an official prompting guide for Claude Opus 5.5, and a Reddit user summarized its 12 key points. The guide covers migration from Opus 5, agent orchestration tips, new safety classifiers, and image handling. It matters to practitioners deploying Opus 5.5 because several changes — mandatory thinking, guardrails, progress-update behavior — can break existing agent loops.
1659 engagement1 sourceMon, Sep 28, 2026, 12:28 AM
AI Overview
- Anthropic's Opus 5.5 guide says thinking can't be disabled; users migrating from Opus 5 should start at low effort and measure results
- Time budgets like "elapsed 340s / 1200s" make agent teams pace themselves and finish sooner, per the guide
- Opus 5.5 runs 30%+ faster than Opus 5 with fewer tokens, and existing prompts should run unchanged
- Changing top-level effort between requests invalidates the prompt cache; per-message effort changes (beta) preserve it
- Guide: progress updates now arrive as thinking blocks, and some end the turn, so unattended agents can stop early
- Removing "think carefully" from chat prompts starts replies sooner with no clear quality drop; effort is the main thinking dial
- Safety classifiers cover biology, cybersecurity, and reasoning extraction; pasted text tagged with random IDs resists hidden instructions
Entities
Related
Social·Community
Opus 5.5 seems to be really good at explainer videos.
A developer reports that Claude Opus 5.5 produces explainer videos of notably higher quality than earlier models. Dogfooding their own product, they generated multiple videos from URLs in one shot, with opencode running on a Linux machine producing the underlying HTML. They ask why the model performs so well at this task and request blog posts on its training.
A developer reports that Claude Opus 5.5 produces explainer videos of notably higher quality than earlier models. Dogfooding their own product, they generated multiple videos from URLs in one shot, with opencode running on a Linux machine producing the underlying HTML. They ask why the model performs so well at this task and request blog posts on its training.
22 engagement1 sourceredditFri, Sep 25, 2026, 06:22 PM
Social·Community
Developer details Opus 5.5 Ultracode spawning 28 subagents and 5.4M tokens in one day of work
A developer building with Claude Opus 5.5 on Ultracode reports the model over-orchestrates: every spec change triggers 2-3 subagents running 10-20 minutes each, causing scope creep and slow feature delivery. He compares the behavior to embedded consultants inventing adjacent problems, and says he prefers Opus 4.6 and GPT 5.4.
A developer building with Claude Opus 5.5 on Ultracode reports the model over-orchestrates: every spec change triggers 2-3 subagents running 10-20 minutes each, causing scope creep and slow feature delivery. He compares the behavior to embedded consultants inventing adjacent problems, and says he prefers Opus 4.6 and GPT 5.4.
15 engagement1 sourceredditTue, Sep 29, 2026, 08:19 AM
Social·Community
The 'preserve thinking' feature breaks context management and degrades Claude's performance
A developer of an 8-month context management plugin reports that the 'preserve thinking' feature added in Claude Fable 5.1 and Opus 5.5 makes any prefix change invalidate all existing thinking blocks. The developer argues this breaks pruning and compression workflows, forcing users back to compaction, and ironically harms Claude users more than would-be distillers.
A developer of an 8-month context management plugin reports that the 'preserve thinking' feature added in Claude Fable 5.1 and Opus 5.5 makes any prefix change invalidate all existing thinking blocks. The developer argues this breaks pruning and compression workflows, forcing users back to compaction, and ironically harms Claude users more than would-be distillers.
10 engagement1 sourceredditWed, Sep 23, 2026, 05:45 AM
Social·Community
Reddit users debate how close Anthropic's Opus 5.5 and Astra are to AGI
A Reddit discussion debates how close Anthropic's Opus 5.5 and Astra are to AGI, based on the poster's experience using Opus 5.5 for coding and observations of Astra on long agentic tasks. The thread reflects practitioner interest in whether current frontier models can plan, reason, and operate computers reliably enough to extend into robotics.
A Reddit discussion debates how close Anthropic's Opus 5.5 and Astra are to AGI, based on the poster's experience using Opus 5.5 for coding and observations of Astra on long agentic tasks. The thread reflects practitioner interest in whether current frontier models can plan, reason, and operate computers reliably enough to extend into robotics.
126 engagement1 sourceredditSun, Sep 27, 2026, 11:52 PM
Social·Community
Opus 5.5 has replaced Fable for me
A developer who previously used Claude Fable for codebase audits reports switching to the newly released Opus 5.5 after benchmarking both on their repo. Opus 5.5 identified 156 bugs with 151 valid and 147 suggested fixes applied without manual edits, leading the user to prefer it over Fable. The report is a single first-hand impression, not a systematic benchmark.
A developer who previously used Claude Fable for codebase audits reports switching to the newly released Opus 5.5 after benchmarking both on their repo. Opus 5.5 identified 156 bugs with 151 valid and 147 suggested fixes applied without manual edits, leading the user to prefer it over Fable. The report is a single first-hand impression, not a systematic benchmark.
7 engagement1 sourceredditTue, Sep 29, 2026, 12:34 AM