OpenAI has officially announced that the API and credit prices for GPT-5.6 Sol will be reduced by over 20% in the next three months. The API is now available and will gradually cover ChatGPT Work and Codex credits, while usage within Pro, Plus, and Business subscriptions remains unchanged. Accompanying updates include the ability to view usage and expenditure per API key, set organization or project-level monthly quotas, and automatically stop traffic upon reaching a hard limit. For teams requiring bulk model calls, these changes impact both cost budgeting and permission governance.<\/p>\n
Sources:<\/p>\n
Anthropic has stated that Claude Security scanning is now powered by Claude Mythos 5 and is open for public beta to Claude Enterprise customers. Scan results can generate suggested patches in the Claude Code web interface, and development teams do not need separate Mythos model access. The company also mentioned plans to integrate Mythos 5 with partner security products and established a $35 million open-source security Defender Advantage Fund. This clarifies product availability and supporting plans, not independent verification of its security efficacy.<\/p>\n
Sources:<\/p>\n
@MaxForAI reports, citing Bloomberg, that Anthropic has hired Amir Salek, who founded and led Google’s TPU project for years, to join its forming custom chip division. The blogger also notes that Anthropic confirmed this month it is building a Custom Silicon team for Claude and that OpenAI’s early custom chip member Clive Chan has also joined. This information in the current report comes from a single blogger’s compilation of media coverage and company statements and should be treated as a significant development lead, not as confirmation that Anthropic has completed custom chip deployment.<\/p>\n
Sources:<\/p>\n
A report by @MaxForAI states that DeepSeek will eliminate peak\/off-peak distinctions on Saturdays and Sundays starting August 23rd at 00:00, charging off-peak rates all day, with pre-effective date fees settled under old rules. If this report aligns with official notification, batch tasks, Agents, and data generation could be prioritized for weekend runs. However, current visible evidence is the blogger’s summary of a notice; the main text does not expand this into a long-term policy independently confirmed by the company.<\/p>\n
Sources:<\/p>\n
@MaxForAI summarizes @pidotdev’s development notes, arguing that after an Agent runs continuously for dozens of hours, the key is no longer just the number of tools, but whether context, state, and failure recovery can be reliably handled. Pi’s approach is to trim overly long tool results from the context, save them to disk, and retain file paths and offsets for retrieval when needed. The notes claim a 26%\u201335% reduction in context usage and a 72%\u201388% reduction in uncached prefill across 19 real sessions, while preserving information recoverability. The notes also describe intent\/result logging, Session Trees, parallel Lanes, and crash recovery, indicating long-term Agents are beginning to approach transaction system and runtime design issues.<\/p>\n
Sources:<\/p>\n
In yesterday’s samples, specific signals emerged from both engineering and creative sides pointing in the same direction: The open-source VA-Judger introduced by @Gorden_Sun no longer scores video and audio quality completely separately, but integrates text matching, audiovisual synchronization, visuals, audio, and content completeness, trained on tens of thousands of human comparison datasets. @nanyuan0412’s practical tests indicate native 2K is significantly better than post-processed super-resolution for hair, skin, clothing, and complex lighting, albeit at higher cost. @Chengzilhy believes Seedance 2.5 shows clear improvement in micro-expressions and performance. Together, they support the idea that “generation quality should consider complete audiovisual and shot expression,” not merely pursuing resolution.<\/p>\n
Sources:<\/p>\n
@nanyuan0412 discovered Zhihu now offers a CLI, preparing to integrate past high-upvote answers, featured posts, and troubleshooting records into search workflows. @luoxiaoshan_ai summarizes OpenAI Developers’ thinking on open-source Codex harnesses, emphasizing that applications should control their own interface, context, tools, and approvals. @op7418, starting from AI hardware usage, argues that transcribed audio context should be able to return to the user’s own Agent. These three items are not new model releases, but rather embedding existing model capabilities into personal information sources, toolchains, and local contexts, offering more direct practical value.<\/p>\n
Sources:<\/p>\n
@Gorden_Sun reports that Anthropic has launched Claude Academy, offering free learning and practice for real-world work and life problems. Its introduction emphasizes not memorizing specific prompts, but understanding AI’s capability boundaries, knowing which tasks are suitable for models, proactively verifying important content, and explaining AI’s involvement to colleagues or clients. As the current report is the blogger’s summary of the platform and internal training methods, it is suitable to present as a public education platform development, not extending to its course effectiveness being verified.<\/p>\n
Sources:<\/p>\n