OPUS 5 CLICK NOW

Overview

The video is a livestream reacting to major AI news and then focusing heavily on Anthropic’s newly released flagship model, Claude Opus 5. The speaker argues it represents a meaningful step-change in both capability and cost-efficiency.

Key News / Context: Open-Source AI Competition

The streamer begins by referencing a letter signed by Jensen and other AI industry leaders supporting open-weight / open-source models.

Main argument for open source

The speaker claims open source:

• Builds a wider ecosystem where more developers can build, fine-tune, and deploy models.
• Helps drive down AI “cost per task” (the speaker’s preferred metric) by enabling competition and more inference options.
• Prevents power concentration among a small number of closed-source labs.

The streamer also frames open ecosystems as beneficial for the US and the wider world by diffusing AI capacity across sectors and increasing model choice.

Nvidia and the economics of openness

The streamer adds that Nvidia benefits as well: broader AI adoption means more chips are needed, making openness economically beneficial.

Note: The speaker mentions being conceptually critical of Anthropic/“Fronteir-labs,” but supports openness for broader competitive pressure.

Main Model Review: Claude Opus 5

The second half shifts to testing and digesting benchmarks and pricing for Claude Opus 5.

Benchmark highlights (as presented in the stream)

The streamer claims Opus 5 beats Fable 5 on most benchmarks, including:

• Agentic terminal coding: higher score (e.g., 43 vs 33).
• GDP “real world practical tasks”: large improvement (about +100 vs Fable 5).
• Arc AGI 3 benchmark: very large jump (about 30%, described as unprecedented in the previous set).
• Browse/comp tasks and OSWorld (computer-use): strong results, with further improvements.
• DeepSui: described as slightly better overall, despite some benchmark-specific variation.

Efficiency emphasis: cost per task

The speaker emphasizes that the improvements aren’t only about raw quality—they point to efficiency and cost per task as central.

Cybersecurity and “misalignment” tradeoffs

A notable exception is cybersecurity performance:

• Opus 5 is described as stronger than Opus 4.8 on some cybersecurity measures.
• However, it still trails “Mythos 5” for exploit development.

The streamer speculates this could be due to stronger guardrails—suggesting Anthropic achieved a surprising outcome: reduced cyber capability while improving overall helpfulness/performance elsewhere.

The streamer argues this is unusual because guardrails often reduce general capability too.

Cost / Efficiency Argument (Core Takeaway)

The speaker repeatedly concludes that you should evaluate cost per task, not just token price.

Pricing comparisons discussed

Opus 5 is presented as:

• Half the price of Fable 5
• Comparable in price points to GPT-5.6 “Soul”, but with better efficiency/bench results per dollar (per the charts shown)

Claimed results

The streamer claims Opus 5 delivers:

• Higher performance
• Lower or similar cost per task
• Even in areas like automation and computer-use

Pricing and Product Details Mentioned

Claude Opus 5 API pricing (as read during the stream)

• $5 per million input tokens
• $25 per million output tokens

The streamer notes this pricing matches Opus 4.8, and is half the price of Fable.

Safety classifier fallbacks

The streamer also mentions automatic fallbacks when requests are flagged by Anthropic safety classifiers:

• Routing may fall back to a safer model (e.g., Opus 4.8).
• The user is still expected to pay the fallback model’s cost.

Availability

The model is described as “available today” across platforms.

Enterprise Validation via Sponsorship (Box)

The video includes a Box sponsorship segment claiming Box benchmarked Opus 5 for enterprise document/knowledge-work tasks.

Reported improvements (vs Opus 4.8) include gains in:

• Due diligence
• Report drafting
• Data analysis (described as a particularly meaningful jump)

Opus 5 is framed as especially strong for unstructured-document grounded analysis.

Additional Community Claims

• The streamer cites messages from Greg (Arc Prize) stating Opus 5 is the most impressive model they’ve seen, with Arc AGI 3 results described as exceptionally high.
• The host speculates Opus 5 may be trained/distilled from Fable 5, based on the pattern of improvements.

Overall Conclusion (Streamer’s Position)

Opus 5 is presented as the best-performing model in the streamer’s view, while being substantially cheaper than Fable 5, with emphasis on:

1. Higher benchmark performance
2. Lower cost per task
3. Guardrails that reduce cyber exploitation without collapsing overall capability

Presenters / Contributors

Main presenter / host

• (Unnamed in subtitles; the streamer speaking throughout)

Named contributors mentioned

• Jensen (referenced as signing the open-model support letter)
• Greg / Greg (Arc Prize president) (messages cited during the stream)
• Matias (super chat)
• Joel (chat participant)
• Alex (producer; mentioned as informing the host about model/UI availability)
• Rodrigo (chat participant)
• Mark Santos (chat participant)
• Dammo Gallagher (chat participant)
• Matthew Berman (named handle cited by the host for X updates)
• Box (sponsoring organization; benchmark contributor)

Models/series mentioned (not people)

• “Mythos,” “Fable 5,” “Opus 4.8,” “GPT 5.6 Soul,” “Haiku,” “Sonnet,” “Claude Opus 5,” “Claude”