Sonnet 5.5: No-Hype Full Review & Testing

Sonnet 5.5: No-Hype Full Review & Testing

by Pat Simmons · View the channel on PSI

Formal & Physical Sciences Computing & Cybersecurity UDDigital Lifestyle and online world: consumer and user guidesUDTMobile phones and smartphones: consumer
🎙 Pat Simmons 👥 26K 📅 September 29, 2026 ⏱ 46 min 👁 353 📄 expert opinion 🧭 2026-09-29
Available in: English (current) Français

Keywords

Sonnet 5.5Opus 5.5GPT-6 Astrabenchmarkscost efficiency

Summary

Pat Simmons reviews Anthropic’s newly released Claude Sonnet 5.5, comparing it against Opus 5.5, Fable 5.1, GPT-6 Astra, and Sonnet 5. He starts by analyzing official benchmarks, highlighting Sonnet 5.5’s strong performance on Terminal-Bench (70.6% vs 66% for Opus) and its closeness to Opus on knowledge-work tasks. The video then presents three real-world builds: a product-launch video for Sonnet 5.5, a 3D website clone of butter.video, and a backrooms horror game. For each build, Simmons uses identical prompts and tracks time, token usage, and cost. He also provides a cost breakdown, noting that Sonnet 5.5 is more efficient than Sonnet 5 and offers similar pricing but better performance. However, he points out that at max effort, Sonnet 5.5 may not be cheaper than Opus 5.5 due to increased token usage. The blind reveal ranks the launch videos, with Opus 5.5 producing the best result, while Sonnet 5.5 comes in third. Overall, Simmons concludes that Sonnet 5.5 is a solid upgrade for daily tasks and knowledge work, but Opus 5.5 remains superior for creative video generation. He emphasizes that the choice depends on the use case and effort level.

188 words

Critical Evaluation

Value of the Information & Strength of the Argument

The video provides substantial value by offering an independent, hands-on evaluation of Sonnet 5.5, backed by quantitative data on cost, tokens, and performance. The argumentation is solid: Simmons systematically compares benchmarks, runs identical prompts across models, and presents raw costs. He also nuances Anthropic’s claims by referencing third-party analysis about token usage at different effort levels. However, the testing is limited to three specific tasks, and the subjective quality rankings could be influenced by personal taste.

Scientific Rigor, Source Quality, Title Accuracy

The scientific rigor is decent; the methodology is reproducible to a degree, with prompts and skills provided in the blog post. Sources cited are Anthropic’s official data and his own experiments, plus a reference to Artificial Analysis. The title accurately reflects the content, promising a no-hype review and delivering detailed testing. The quality of sources is good, though all are from the creator’s own blog or company pages, lacking independent verification.

162 words

Title / Content Match

The title accurately reflects the content: a non-hyped, thorough review with hands-on testing and cost analysis.

Quality & Reliability

8/10

The review is systematic, comparing Sonnet 5.5 against several frontier models on benchmarks and real-world builds, with detailed cost and token tracking. However, the tests are limited to a few scenarios and the evaluation is subjective in parts.

Chapters

Cited Sources

Contribution & Novelties

The video adds original value by conducting independent, real-world testing of Sonnet 5.5 against multiple frontier models, with a focus on cost efficiency and practical applications. It challenges Anthropic’s marketing claims by analyzing token usage at different effort levels, citing third-party measurements. The video also introduces a skill for creating launch videos, which is shared in the blog post.

Pour aller plus loin :

  • Claude (language model) — Context on Anthropic’s Claude family and evolution.
  • Large language model — Overview of LLMs, benchmarks, and evaluation methods.
  • Terminal-Bench — Mentioned as a benchmark for AI agent capabilities; exact address uncertain.

99 words

Radar Profile

The radar chart likely shows high scores on information quantity and quality, with moderate technical depth and reliability, indicating a well-rounded but not flawless review.

Reliability 7/10