[{"data":1,"prerenderedAt":12},["ShallowReactive",2],{"$f2lm8ackrz5agl":3},{"slug":4,"title":5,"date":6,"dateDisplay":7,"minuteRead":8,"description":9,"problems":10,"html":11},"2026-08-09_2026-08-15","Alibaba says its new 27B open model beats Claude Opus 4.6","2026-08-16","August 16, 2026",2,"Alibaba says its new 27B open model beats Claude Opus 4.6. Alibaba released Qwen3.8-27B on August 14 under Apache 2.0, the permissive open-source licence…",[],"\u003Csection class=\"section section--paragraph\">\n  \u003Cspan class=\"section__marker\" aria-hidden=\"true\">01\u003C\u002Fspan>\n  \u003Ch2 class=\"section__heading\">Alibaba says its new 27B open model beats Claude Opus 4.6\u003C\u002Fh2>\n  \u003Cp>Alibaba released \u003Ca href=\"https:\u002F\u002Fhuggingface.co\u002FQwen\u002FQwen3.8-27B\" rel=\"noopener\">Qwen3.8-27B\u003C\u002Fa> on August 14 under Apache 2.0, the permissive open-source licence that allows free commercial use. In Alibaba's own testing it beats Claude Opus 4.6 on 9 of 12 text benchmarks and jumps well past April's Qwen3.6-27B on coding and agent tasks. At 4-bit its weights fit in about 18 GB: one 24 GB consumer graphics card, or a Mac with 32 GB of memory.\u003C\u002Fp>\n\u003C\u002Fsection>\n\u003Csection class=\"section section--blurbs\">\n  \u003Cspan class=\"section__marker\" aria-hidden=\"true\">02\u003C\u002Fspan>\n  \u003Ch2 class=\"section__heading\">DeepSeek raised prices; Anthropic opted not to\u003C\u002Fh2>\n  \u003Cp class=\"blurb\">\u003Cstrong class=\"blurb__lead\">DeepSeek's own API got more expensive on August 16.\u003C\u002Fstrong> Three days after releasing V4-Pro under an MIT license, DeepSeek raised what it charges to serve the model: output prices more than doubled off-peak and rose about 4.6x at peak, and cached input — most of an agentic coding bill — up to 12x. \u003Ca href=\"https:\u002F\u002Fapi-docs.deepseek.com\u002Fquick_start\u002Fpricing\" rel=\"noopener\">Prices\u003C\u002Fa>\u003C\u002Fp>\n  \u003Cp class=\"blurb\">\u003Cstrong class=\"blurb__lead\">Anthropic called off a price rise it had already scheduled.\u003C\u002Fstrong> Sonnet 5's $2\u002F$10 per million tokens was due to become $3\u002F$15 on September 1; the pricing docs now say the increase will not occur. It comes less than two weeks after \u003Ca href=\"https:\u002F\u002Fopenai.com\u002Findex\u002Fadvancing-the-price-performance-frontier-with-gpt-5-6\u002F\" rel=\"noopener\">OpenAI's 80% cut\u003C\u002Fa> to GPT-5.6 Luna, to $0.20\u002F$1.20, which OpenAI credited to serving efficiency. \u003Ca href=\"https:\u002F\u002Fplatform.claude.com\u002Fdocs\u002Fen\u002Fabout-claude\u002Fpricing\" rel=\"noopener\">Pricing\u003C\u002Fa>\u003C\u002Fp>\n\u003C\u002Fsection>\n\u003Csection class=\"section section--paragraph\">\n  \u003Cspan class=\"section__marker\" aria-hidden=\"true\">03\u003C\u002Fspan>\n  \u003Ch2 class=\"section__heading\">Claude's text watermark ships before its detector\u003C\u002Fh2>\n  \u003Cp>Per Anthropic, Claude models launched on or after August 2 embed an invisible watermark in the text they generate, its implementation of the EU marking obligation that took effect that day, applied globally. The mechanism is a version of Google DeepMind's SynthID-Text: a key steers which word the model picks next, so nothing is inserted and there are no hidden characters. What does not exist yet is any way to check for it: the detection API is \"soon.\" Generated files instead carry C2PA provenance metadata, which Anthropic itself calls very different from a watermark, and which platforms routinely strip on upload. \u003Ca href=\"https:\u002F\u002Fwww.anthropic.com\u002Fnews\u002Fclaude-text-watermark\" rel=\"noopener\">Anthropic\u003C\u002Fa>\u003C\u002Fp>\n\u003C\u002Fsection>\n\u003Csection class=\"section section--paragraph\">\n  \u003Cspan class=\"section__marker\" aria-hidden=\"true\">04\u003C\u002Fspan>\n  \u003Ch2 class=\"section__heading\">Sony and TSMC sign a $4.8 billion sensor-fab deal\u003C\u002Fh2>\n  \u003Cp>Sony and TSMC signed a deal on August 11 to make image sensors for smartphones together in Kumamoto, Japan, in a new company Sony will control. Sony contributes about $3 billion, including a factory it has already built there; TSMC adds up to $1.8 billion in cash plus its manufacturing process. \u003Ca href=\"https:\u002F\u002Fwww.sony-semicon.com\u002Fen\u002Fnews\u002F2026\u002F2026081101.html\" rel=\"noopener\">Joint release\u003C\u002Fa>, \u003Ca href=\"https:\u002F\u002Fpr.tsmc.com\u002Fenglish\u002Fnews\u002F3332\" rel=\"noopener\">TSMC board\u003C\u002Fa>\u003C\u002Fp>\n\u003C\u002Fsection>\n\u003Csection class=\"section section--further-reading\">\n  \u003Cimg class=\"section__mark\" src=\"\u002Fbranding\u002Fstargazer.png\" width=\"54\" height=\"25\" alt=\"\" \u002F>\n  \u003Ch2 class=\"section__heading\">Something to chew on\u003C\u002Fh2>\n  \u003Cp class=\"read\">\u003Ca class=\"read__title\" href=\"https:\u002F\u002Fwww.geoffreylitt.com\u002F2026\u002F07\u002F02\u002Funderstanding-is-the-new-bottleneck\" rel=\"noopener\">Understanding is the new bottleneck\u003C\u002Fa> — The argument, from Notion design engineer Geoffrey Litt: agents can be left to check the code, so the constraint on a software project shifts to whether the humans steering it still understand what they have — the thing that lets a team iterate on purpose rather than by prompt.\u003C\u002Fp>\n\u003C\u002Fsection>",1787330503545]