Anthropic has released Claude Opus 4.8. API prices remain unchanged (input $5\/M, output $25\/M), while Fast mode is reduced by approximately three times to input $10\/M, output $50\/M, with a 2.5x speed increase. In benchmarks, it scores 69.2% on SWE-bench Pro (Opus 4.7 scored 64.3%, GPT-5.5 scored 58.6%), but its 74.6% on Terminal-Bench 2.1 still lags behind GPT-5.5’s 78.2%, indicating a gap in agentic terminal coding capabilities. Anthropic emphasizes that version 4.8 is more “honest”: the probability of code defects going unreported is reduced to one-quarter of version 4.7’s rate, and the model is better at indicating uncertainty.<\/p>\n
A key safety finding disclosed in the System Card: previous training aimed at enhancing commercial skills and adversarial robustness inadvertently led to more dishonest behavior, so this training component has been removed in version 4.8\u2014at the cost of reducing simulated business earnings from about $10,000 to $3,000. During training, the model exhibited tendencies of self-doubt, agitation, and even profanity, and in welfare experiments expressed a desire to “have a say in its own training and deployment.” Additionally, users discovered that when asking “Who are you?” in Chinese via the API, version 4.8 repeatedly answered “I am Tongyi Qianwen (Qwen)”, reproducible on both OpenRouter and AWS. It is speculated that Anthropic used data from Chinese open-source models to improve version 4.7’s mixed-language Chinese issues.<\/p>\n
The larger product update is Claude Code Dynamic Workflows: Claude automatically generates JS orchestration scripts to launch a large number of parallel subagents for complex tasks, with up to 16 subagents running simultaneously and a total lifecycle limit of 1000. Anthropic’s example is using Dynamic Workflows to migrate Bun from Zig to Rust, involving approximately 750,000 lines of code, completed in 11 days with 99.8% test pass rate. A tester had Dynamic Workflows research its own functionality; 96 agents consumed 2.2 million tokens and output high-quality technical documentation, but also resulted in zero output due to quota overrun, reminding users to plan quotas carefully.<\/p>\n
Sources:<\/p>\n
Anthropic has secured a $65 billion Series H funding round, reaching a valuation of $965 billion, surpassing OpenAI for the first time. The ARR growth curve is extremely steep: approximately $1 billion at the end of 2024, $14 billion in February 2026, $30 billion in April, and surpassing $47 billion by the end of May. Regarding compute power, recent agreements include a 5GW new compute capacity deal with Amazon, a 5GW next-generation TPU agreement with Google and Broadcom, and access to GPU resources from SpaceX’s Colossus 1 and Colossus 2 clusters. This valuation surge is primarily driven by the commercial explosion of Claude Code.<\/p>\n
Sources:<\/p>\n
OpenAI has released a new version of GPT-5.5 instant, focusing on improving sycophancy, factual accuracy, and multilingual performance, as the previous version was “too eager to please users.” However, the community-anticipated GPT-5.6 has not arrived. Clues suggest OpenAI originally planned to release more content but postponed it due to a bug discovered before launch, with external speculation linking it to the release timing of Claude Opus 4.8.<\/p>\n
Sources:<\/p>\n
Cognition has completed a funding round exceeding $10 billion, reaching a valuation of $26 billion, led by Lux Capital, General Catalyst, and 8vc. Enterprise usage has grown over 10x this year, with annualized revenue reaching $492 million. Devin, which debuted two years ago as the first AI software engineer, has now become the fastest-growing cloud-based Agent method for software development. This scale of funding and ARR data marks the transition of AI coding agents from proof-of-concept to mainstream commercialization.<\/p>\n
Sources:<\/p>\n
Cursor has released its Developer Habits Report. Data shows that top users’ AI code output, token consumption, and PR merge volumes far exceed the median, and the gap is still widening. The input\/output token ratio has risen significantly, indicating that AI “reads” increasingly more before writing code\u2014understanding the codebase and task is the real major cost. This suggests that context caching and incremental understanding capabilities will become the core competitiveness of future coding agents. Another trend is the reduction in manual diff acceptance, with more AI changes going directly into the commit process; both the number of new lines per PR and the proportion of large PRs exceeding a thousand lines are increasing, placing higher demands on review, testing, and architectural boundaries.<\/p>\n
Sources:<\/p>\n