跪拜 Guibai
← All articles
Frontend · AI Programming · Claude

Anthropic Ships Sonnet 5 and Claude Science on the Same Day, Valuation Hits $965B

By kyriewen ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Anthropic's simultaneous product and research push, backed by a near-trillion-dollar valuation, signals that the leading AI labs are no longer just API vendors—they are becoming full-stack scientific organizations. For developers, Sonnet 5's Opus-competitive performance at Sonnet pricing resets the cost floor for high-capability coding agents.

Summary

Claude Sonnet 5 now ships as the default model for all users at unchanged pricing, delivering coding and agent performance that rivals the far more expensive Opus tier. The same day, Anthropic unveiled Claude Science, a research platform that integrates over 60 scientific databases to automate tasks like protein-structure prediction, alongside an internal drug-discovery program targeting neglected diseases.

OpenAI countered with its first hardware product, a macro keyboard for its Codex assistant, and an internal inference optimization that reportedly halves model serving costs. Meanwhile, Meituan open-sourced its LongCat-2.0 model trained entirely on Chinese-manufactured chips, and UBTECH priced its humanoid U1 robot from $16,500 to $136,000.

Taiwan escalated investigations into AI-hardware smuggling as US export controls tighten, and IBM disclosed progress on a 1-nanometer chip process. The day's announcements trace a widening split between cost-cutting scale plays and premium pricing strategies across the AI industry.

Takeaways
Claude Sonnet 5 is now the default model for Free and Pro users, priced at $2/M input and $10/M output tokens.
Sonnet 5 outperforms Sonnet 4.6 on coding, tool use, and multi-step agent tasks, with some capabilities reaching Opus levels.
Claude Science integrates over 60 scientific databases and automates biological and chemical research workflows.
Anthropic launched an internal drug-discovery project targeting neglected diseases, with applications open until July 15.
Anthropic's valuation reached $965 billion, surpassing OpenAI as the highest-valued AI startup.
OpenAI's first hardware product is the Codex Micro, a 13-key programmable macro keyboard shipping July 15.
OpenAI engineers demonstrated an inference optimization that cuts model serving costs by over 50%.
Meituan open-sourced LongCat-2.0, a large model trained entirely on Chinese-manufactured chips.
UBTECH's humanoid U1 robot ranges from 119,800 yuan to 990,000 yuan depending on the configuration.
Taiwan is intensifying investigations into AI-hardware smuggling as US chip export controls tighten.
Conclusions

Anthropic is executing a 'product plus research plus capital' strategy that no other AI lab currently matches—OpenAI has more users and Google has more infrastructure, but neither runs an internal drug-discovery program.

Sonnet 5's pricing strategy is a deliberate commoditization of high-end capability, compressing the mid-tier market and forcing competitors to respond on either price or performance.

OpenAI's macro keyboard is a low-risk hardware entry point that tests supply chains with its most loyal user base before attempting a consumer device.

Cutting inference costs in half lets OpenAI push free-tier adoption aggressively, while DeepSeek's recent price hike suggests a harvest-phase strategy—two business models that will increasingly diverge.

Meituan's fully domestic-chip-trained open-source model proves that non-Nvidia training pipelines are viable for certain scales, which matters for any organization facing export-control risk.

UBTECH's six-figure pricing for a humanoid robot marks the transition from lab prototype to consumer product, even if the price curve has years to fall.

US export controls on AI chips are creating a predictable gray-market boom, with Taiwan caught between its manufacturing role and compliance pressure.

Concepts & terms
Claude Science
Anthropic's AI research platform that integrates over 60 scientific databases to automate tasks such as protein-structure prediction and chemical analysis, targeting biology and chemistry workflows.
Codex Micro
OpenAI's first hardware product, a 13-key programmable macro keyboard built with Work Louder, designed for quick actions within the Codex programming assistant.
LongCat-2.0
A large language model open-sourced by Meituan, trained entirely on Chinese-manufactured chips as a demonstration of non-Nvidia training viability.
ASIC chip (Jalapeño project)
A custom chip co-developed by OpenAI and Broadcom, designed from scratch specifically for large-model inference rather than repurposing general-purpose GPUs.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗