/Tech8h ago

Anthropic releases Claude Sonnet 5, scoring 63.2% on SWE-bench Pro and 81.2% on OSWorld computer use

Story Overview

Anthropic is releasing Claude Sonnet 5 today as its most agentic Sonnet model, with improved planning, browser and terminal tool use, and autonomous task execution that previously needed larger models. Official details highlight gains over Sonnet 4.6 in reasoning and coding while landing close to but generally below Opus 4.8 on evaluations. Reports circulating on social media mention a 1M context window and January 2026 knowledge cutoff, yet these specifics do not appear in the primary announcement or system card.

9495.7K670542245.4K

Original post

Chubby♨️@kimmonismus#1360inTech

Sonnet 5 incoming today per leaker @synthwavedd

- Knowledge cutoff January 2026 - Launching at $2/$10 per Mtok promotional rate* - 1M context variant available

I assume we at least (!) get also fable 5 news.

leo 🐾@synthwavedd

🚨 SCOOP: Claude Sonnet 5 is releasing later today:

- Knowledge cutoff January 2026 - Launching at $2/$10 per Mtok promotional rate* - 1M context variant available

*Some changes have been made behind the scenes to prepare for a promo price offering, preliminary values

6:45 AM · Jun 30, 2026 · 2.3K Views

Pricing Watch

Intro rates hold expenses steady for now

The model launches at $2 per million input tokens and $10 per million output tokens through August 31 2026 before shifting to $3 and $15. An updated tokenizer raises token counts, but the temporary rates are set to keep overall costs roughly neutral for users during the promo window.

Developer Impact

Agentic upgrades reach more users at lower cost

Sonnet 5 becomes the default model for free and pro plans immediately, with access also rolling out to higher tiers and the API. It shows clear lifts versus Sonnet 4.6 on agentic coding and computer-use benchmarks while safety checks report fewer unwanted behaviors than the prior version.

Sentiment

Positive users praise Claude Sonnet 5 for matching higher-end performance at lower prices with stronger agentic features, while negative users call it a downgrade that feels disappointing or outdated compared to earlier versions.

Pos

61.3%

Neg

38.7%

232 comments with sentiment.

Cluster Engagement

Digg Deeper

Kevin Roseasked

Compare this models SWE-bench Pro vs all other frontier models side-by-side, just give me models and scores

Claude Sonnet 5: 63.2%. Opus 4.8: 69.2%; Sonnet 4.6: 58.1%; GPT-5.4 (xHigh, Scale public): 59.1%; Opus 4.6 (thinking): 51.9%; Gemini 3.1 Pro (thinking): 46.1%. ^¹

Vendor aggregates and other scaffolds report higher peaks (Opus 4.8 at 69.2%, Fable 5/Mythos 5 near 80%).