AI News | Update from Sonnet 4.6 to Sonnet 5

✦ What you will learn in this article

What has changed in Sonnet 5?
And how should you choose between it and the higher-end Opus 4.8 model?

This article breaks it down with concrete figures.

✦ Breaking News: Sonnet 5 has been released

The announcement was made in the early hours of July 1, 2026, Japan Standard Time.

It has been updated from the previous Sonnet 4.6
to Sonnet 5.

According to the official announcement,
it is positioned as the most autonomous model
among all Sonnet models to date.

Making plans,
using tools,
and completing complex tasks from start to finish.

It feels like we are gradually moving in that direction.

✦ Before / After: A look at 6 key metrics

Here are the evaluation metrics released by the official team.

🔘 Coding (SWE-bench Pro)
　Sonnet 4.6: 58.1%
　Sonnet 5: 63.2%
　Opus 4.8: 69.2% (Reference)

🔘 Terminal Operation (Terminal-Bench 2.1)
　Sonnet 4.6: 67.0%
　Sonnet 5: 80.4%
　Opus 4.8: 82.7% (Reference)

🔘 Interdisciplinary Reasoning - No Tools (Humanity's Last Exam)
　Sonnet 4.6: 34.6%
　Sonnet 5: 43.2%
　Opus 4.8: 49.8% (Reference)

🔘 Interdisciplinary Reasoning - With Tools (Same as above)
　Sonnet 4.6: 46.8%
　Sonnet 5: 57.4%
　Opus 4.8: 57.9% (Reference)

🔘 Computer Operation (OSWorld-Verified)
　Sonnet 4.6: 78.5%
　Sonnet 5: 81.2%
　Opus 4.8: 83.4% (Reference)

🔘 Knowledge Work (GDPval-AA v2)
　Sonnet 4.6: 1395
　Sonnet 5: 1618
　Opus 4.8: 1615 (Reference)

✦ Knowledge work is the only area where it surpassed the higher-end model

Looking at the numbers side-by-side,
there is one point that stands out.

The knowledge work metric (GDPval-AA v2) shows that
Sonnet 5slightly outperformed Opus 4.8.

For tasks such as writing, organizing documents, and summarizing,
it means you can achieve results equal to or better than the higher-tier model.

✦ On the other hand, areas where gaps still remain

It is not the case that everything is on par,
and in some areas, Opus 4.8 still maintains higher scores.

The areas where differences particularly remain are
coding and complex reasoning.

For tasks like investigating the causes of intricate bugs or
organizing large codebases,
my impression is that the higher-tier model is still better.

✦ How has the pricing changed?

While performance has improved,
the pricing remains nearly the same.

🔘 Sonnet 5 (Introductory price, until August 31, 2026)
Input: $2 per 1 million tokens
Output: $10 per 1 million tokens

🔘 Sonnet 5 (Standard price, from September onwards)
Input: $3
Output: $15

🔘 Opus 4.8 (Reference)
Input: $5
Output: $25

Even at the standard price,
it comes to about 60% of the price of Opus 4.8.

✦ In the end, how should you choose between them?

When you organize the numbers,
a rough guideline becomes visible.

✅ Writing, planning, summarizing, and organizing information
→ Sonnet 5 is sufficient (in fact, it is better in some situations)

✅ General coding and automation
→ Try Sonnet 5 first

✅ Intricate bug investigation and large-scale tasks
→ Choose Opus 4.8

✅ Areas where mistakes are problematic (situations requiring expert judgment)
→ Opus 4.8 from the start

✦ Procedure when you are unsure

1. First, try it with Sonnet 5

2. If you feel the accuracy is insufficient or it gets stuck due to complexity, switch to Opus 4.8

3. For tasks where you know you need high precision from the start, choose Opus 4.8

4. For lighter tasks where speed or cost is the top priority, consider a smaller model

✦ Precisely because the top-tier models are currently unavailable

Until now, the top-tier Mythos 5 and Fable 5 have remained unavailable to the general public
due to the impact of export controls.

Given that situation,
the fact that the intermediate model, Sonnet 5, has been improved this much
is something I feel is quietly helpful.

Even while the higher tiers remain out of reach,
the quality of the models at hand is quietly rising.

While keeping credit consumption down,
the range of what can be done is expanding, is the feeling I get.

✦ Update: Restrictions on top-tier models to be lifted

There has been a development just as I am writing this article.

Anthropic has announced that the U.S. Department of Commerce has lifted export restrictions
on Mythos 5 and Fable 5.

Access restoration has begun,
and an update is expected to be announced in the near future.

✦ Today's small step

Without realizing it,
the models you are using may have evolved.

It might be a good idea to check your current version
and try it out with some writing tasks.

Thank you for reading. See you in the next article.

If you only wanted to see the article, you are all set.
I have summarized the site map and related links below.

This page is
a guide page organized so you can reach the articles you are interested in without getting lost.

Please take your time and browse from the sections that interest you.
※Currently being updated as needed.

▼Pickup Articles

Thank you very much.

📱 SNS ⇒ X・Threads・Instagram

💬Related Tags
#AI #AIUtilization #ImproveDailyLifeWithAI #DailyLifeImprovement #OrganizeYourLife #LifestyleImprovement #Efficiency #WorkEfficiency #BusinessImprovement #WorkStyle #WorkTechniques #LifeHack #DigitalUtilization #ToolUtilization #Prompt #PromptDesign #LivingWithAI #LifeWithAI #AIForBeginners #AILearning #LearningRecord #TriedIt #TriedItWithAI #ExperienceReport #Insights #Mindset #HabitFormation #note #LikeIt #LateNightLatte