Tibo says a fix for abnormal Codex usage is coming soon, with usage to be reset
开发生态
Tibo said the team had identified three issues causing Codex usage to be consumed faster than expected, with fixes due to roll out soon. Usage for all paid subscribers was scheduled to reset at around 2:00 p.m. Pacific Time on August 23, 2026.
Tibo said his team would soon deploy fixes for the Codex usage-consumption issues, with the accompanying reset for paid subscriptions expected to take effect at around 2:00 p.m. Pacific Time on August 23, 2026, equivalent to 5:00 a.m. Beijing time on August 24, 2026. According to his explanation, the issues involved inefficiency when using images in long sessions containing multiple compactions, higher Computer History usage above the p95 range, and a conversation-title generation feature consuming more usage than expected. The reset covers all paid subscribers. When the original report was published, many Plus users had reported that their usage had been reset, while the reset had not yet taken effect for Business and Pro subscriptions. Tibo also said in a reply to a user that sub2api is not supported.
Boris Cherny responds to Opus 5 criticism: concise output setting launched as an interim measure
开发生态
Claude Code creator Boris Cherny said the team has released a temporary setting for Opus 5’s verbose output. Users can switch styles by running claude /config outputStyle=concise, while a long-term fix remains a key team priority.
Responding to community criticism of Opus 5, Boris Cherny said the team has already released a concise output setting, allowing users to change the output style by running claude /config outputStyle=concise without waiting for a long-term fix. He acknowledged that Opus 5 is not perfect and said fixing the model is an important priority for the team. According to his explanation, Opus performs well on long-running tasks and coding, but verbose output is one of its current issues. The new configuration is a rapid interim measure rather than the long-term fix itself.
zcode preview 3.9.1 added computer use and model input modality selection, while a user connecting to glm-5.3-t encountered an error resembling those from internal test models. The clues point to a possible launch of Zhipu’s glm-5.3-turbo, but there is no official confirmation.
Zhipu has not officially confirmed whether glm-5.3-turbo is about to launch, and the assessment is based on feature changes in the latest zcode preview 3.9.1 and a user test. zcode added computer use and model input modality selection in this version. According to an earlier statement by zcode staff, computer use would launch together with a multimodal model. Another user tried connecting to glm-5.3-t through GLM Coding Plan and received an error that differed from those produced by other erroneous models. According to the user, this type of error had previously appeared multiple times with Zhipu’s unreleased internal test models.
Identifiers for two new Anthropic Claude models reportedly revealed
前瞻与传闻
A social media user said two new Claude model identifiers had appeared at Anthropic and speculated that an Opus or Sonnet update was more likely next week. The proposed tier and timing are personal judgments, and Anthropic has not confirmed that the models exist.
A social media user posted that two new model identifiers, claude-marshmallow-eap and claude-melon-eap, had appeared at Anthropic, but Anthropic has not confirmed that the models exist. The user judged that the models associated with the identifiers were not in the fable tier but performed reasonably well, and also speculated that an Opus or Sonnet model update was more likely next week. The claims about model tier, performance, and a possible update next week are all the poster’s personal judgments.
Hugging Face reportedly explores a sale at a potential valuation of over $13 billion
前瞻与传闻
Business Insider reported that Hugging Face is working with a bank to gauge acquisition interest from potential buyers. The deal could value the company at $13 billion or more, but no transaction has been reached and Hugging Face has not officially confirmed the report.
According to Business Insider, Hugging Face is working with a bank to assess potential acquirers’ interest in a transaction, but no related deal has been reached. The report said a potential transaction could value Hugging Face at $13 billion or more. The information currently comes only from media reporting, and Hugging Face has not issued official confirmation; the banking arrangement, potential buyer interest, and valuation have not been verified by the company.
Alibaba plans HK$80 billion share placement, with all net proceeds to be invested in AI
前瞻与传闻
Alibaba announced on August 23, 2026, a plan to place HK$80 billion of new shares with non-U.S. persons outside the United States. It will direct 100% of the net proceeds to full-stack AI capabilities and AI infrastructure, marking its first share placement since its 2019 Hong Kong listing.
On August 23, 2026, Alibaba unveiled a plan to issue HK$80 billion, or S$12.95 billion, of new shares to non-U.S. persons outside the United States and use all net proceeds for AI. This will be the company’s first new-share placement since it listed in Hong Kong in 2019. According to its explanation, 100% of the funds will be invested in full-stack AI capabilities and AI infrastructure. Alibaba said the transaction would be the largest follow-on share offering by a Hong Kong-listed company and the largest Regulation S equity offering on record. By size, it would rank as the world’s third-largest follow-on offering by a listed company in 2026, behind Alphabet and Intel.
Gelonghui said Alibaba’s latest financial report, released on August 20, 2026, showed that ARR from AI-related products had exceeded RMB49.5 billion, or S$9.39 billion. The platform projected that ARR would reach US$10 billion, or S$12.69 billion, in the next quarter, while Alibaba Cloud’s external commercialization revenue would reach US$100 billion in 2030 with a profit margin above 20%. It also said global investors had recently taken a bullish view of Alibaba’s position as China’s AI leader and raised their target prices. Meanwhile, Alibaba’s adjusted net profit for the first quarter of its current fiscal year fell 38.2% and missed market expectations. Capital expenditure increased 75% as the company continued investing in AI infrastructure. CEO Wu Yongming said the related investments were expected to pay back within three years.
NVIDIA's GB and VR series servers reportedly set for price hikes early next year as HBM costs surge
前瞻与传闻
Bloomberg, citing people familiar with the matter, reported that some major Nvidia customers were told servers using its AI chips will cost more for shipments from early 2027, with increases exceeding 15% in many configurations, including Vera Rubin and Grace Blackwell systems.
Bloomberg reported on August 22, 2026, that Nvidia had communicated server price changes to some major customers, according to several people familiar with the discussions, with the new pricing applying to products delivered from early 2027. The changes cover systems equipped with the flagship Vera Rubin and Grace Blackwell chips, and prices will rise by more than 15% in many configurations. According to the sources, the actual increase will vary by Nvidia chip generation and memory configuration, against a backdrop of rising memory-chip costs, particularly HBM. The customer communications have not been made public, so the sources requested anonymity.
