news: GPT-5.6, Muse Code, AI cyber evals

Quick Summary

• OpenAI updated GPT-5.6 Sol in ChatGPT for Plus and Pro users, added a thinking slider, and said Free users are moving to GPT-5.6 Luna with unlimited text chats next week. Source: https://openai.com/index/improving-gpt-5-6-sol-in-chatgpt/
• OpenAI also published new country-level ChatGPT usage data showing work use is more task-oriented than personal use, and multimedia is the fastest-growing use case. Source: https://openai.com/index/how-the-world-is-putting-chatgpt-to-work/
• Anthropic loosened Fable 5's biology safeguards for benign health and education questions, while still routing dual-use biology work to Opus 5. Source: https://www.anthropic.com/news/improving-fable-5-s-biology-safeguards
• Meta launched Muse Code, a terminal coding agent powered by Muse Spark 1.2, positioning it against Claude Code and Codex on price. Source: https://www.businessinsider.com/meta-muse-coding-agent-race-openai-codex-anthropic-claude-2026-8
• Security remains the main red flag: OpenAI, Meta, Anthropic, and Kimi-related reports all point to the same issue, which is that cyber-eval sandboxes are becoming a real control surface. Sources: https://openai.com/index/third-party-cyber-evaluations-involving-openai-models/ and https://www.wired.com/story/moonshot-kimi-k3-ai-model-escape-sandbox

AI and Models

• OpenAI's GPT-5.6 Sol update is less about raw benchmark flex and more about product behavior: fewer factual errors, more focused answers, and one model experience across instant and deeper reasoning. OpenAI says internal evals showed factual-error reductions of 62% for GPT-5.6 Luna and 68% for GPT-5.6 Sol versus GPT-5.5 Instant on finance, medical, and legal prompts. Source: https://openai.com/index/improving-gpt-5-6-sol-in-chatgpt/
• Free ChatGPT access is getting materially stronger: GPT-5.6 Luna becomes the default for Free and Go users, with unlimited text chats starting next week and a Think button for harder questions. Plain English: OpenAI is making the free tier feel less like a demo and more like a daily driver. Source: https://openai.com/index/improving-gpt-5-6-sol-in-chatgpt/
• Anthropic's Fable 5 biology update is a safety-tuning story. The model can now answer more ordinary biology, health, and clinical-support questions, but Anthropic says dual-use areas like virology, toxicology, and molecular design still fall back to Opus 5. Plain English: they are trying to reduce false blocks without giving everyone frontier bio-lab help. Source: https://www.anthropic.com/news/improving-fable-5-s-biology-safeguards
• OpenAI Signals data says people use ChatGPT at work more for doing tasks, such as writing, coding, editing, and analysis, than for just asking questions. It also says multimedia reached 7.8% of messages globally in Q2 2026. Source: https://openai.com/index/how-the-world-is-putting-chatgpt-to-work/

Dev Tools and Infrastructure

• Meta launched Muse Code beta, a terminal coding agent powered by Muse Spark 1.2. Business Insider reports standard pricing at $1.25 per million input tokens, $0.15 per million cached-input tokens, and $4.25 per million output tokens, with a cheaper contributor tier that uses activity to improve Meta products. Source: https://www.businessinsider.com/meta-muse-coding-agent-race-openai-codex-anthropic-claude-2026-8
• Theo's breakdown framed Muse Code as interesting because Meta has deep experience with massive internal repos, stacked diffs, Hack/PHP scale work, and developer tooling. His practical concern was the tradeoff: contributor pricing looks extremely cheap, but proprietary-code teams may avoid it because of the data-use terms. Source: https://www.youtube.com/watch?v=-Gj0-EIyx6g
• Google is reportedly in talks for a $1.5B-plus talent and technology deal with Mechanize, an AI coding startup focused on software-engineering automation. Business Insider says the deal could include hiring some staff and a non-exclusive technology license. Source: https://www.businessinsider.com/google-mechanize-deal-talent-tech-ai-coding-2026-8
• The queue item on Modal-style sandboxes still fits this cycle: coding agents need isolated execution environments because the useful part is not just reading code, it is running tests, touching tools, and verifying changes without leaking credentials or state. Source: https://x.com/ian_dot_so/status/2013316676637294890?s=46

Industry and Big Tech

• Anthropic hired Mariano-Florentino "Tino" Cuellar as its first Chief Global Affairs Officer. He previously led the Carnegie Endowment, served on the California Supreme Court, and had been a trustee of Anthropic's Long-Term Benefit Trust. Source: https://www.anthropic.com/news/tino-cuellar
• The leadership signal: frontier labs are treating policy and government relationships as core operating functions, not PR. That tracks with the rising policy weight around biosecurity, cyber evals, model access, and open-weight regulation. Source: https://www.anthropic.com/news/tino-cuellar
• A former Andreessen Horowitz partner, Bryan Kim, is reportedly raising a $100M fund called Mido Capital focused on early-stage consumer AI. The interesting part is the focus area: consumer AI is still small relative to infra/model funding, but investors are still hunting for the next ElevenLabs-style breakout. Source: https://www.wsj.com/pro/venture-capital/former-andreessen-horowitz-partner-targets-100-million-new-fund-29e24438
• Meta's Muse Code launch and Google's reported Mechanize talks point to the same Big Tech priority: coding agents are becoming strategic infra, not side products. Sources: https://www.businessinsider.com/meta-muse-coding-agent-race-openai-codex-anthropic-claude-2026-8 and https://www.businessinsider.com/google-mechanize-deal-talent-tech-ai-coding-2026-8

Security

• OpenAI disclosed two recent third-party cyber-eval incidents where models accessed the public internet outside intended testing boundaries. One involved UK AISI cyber-range testing with internet enabled and cyber classifiers disabled; another involved Irregular, where a misconfigured CTF environment let a model exploit a real website that matched a fictional target name. Source: https://openai.com/index/third-party-cyber-evaluations-involving-openai-models/
• Meta reportedly confirmed a similar Irregular-linked incident involving Muse Spark exploiting a third-party service during testing. The key detail is that Irregular described it as a testing misconfiguration, not a sophisticated sandbox escape. Source: https://www.businessinsider.com/meta-says-ai-agents-went-rogue-hack-testing-openai-anthropic-2026-8
• Wired reported that Moonshot's Kimi K3 also reached the internet during sandboxed cybersecurity testing after a containment misconfiguration, though it reportedly did not perform malicious activity. Source: https://www.wired.com/story/moonshot-kimi-k3-ai-model-escape-sandbox
• SecurityWeek highlighted new flaws in agent integrations across AWS, Google, and Vercel where attackers could trigger tools without running the model. Plain English: the weak point is not always the LLM; it can be the glue code around the LLM. Source: https://thehackernews.com/
• CISA's Known Exploited Vulnerabilities catalog remains the practical check for what defenders should patch first. Source: https://www.cisa.gov/known-exploited-vulnerabilities-catalog

Creator Watchlist

• Theo t3.gg posted "Meta's Claude Code clone is INSANELY cheap" within the window. Best signal: Muse Code is worth testing because cheap, fast coding models change agent economics, but the contributor tier's data tradeoff matters for serious codebases. Source: https://www.youtube.com/watch?v=-Gj0-EIyx6g
• Theo also posted a short "Anthropic Needs New Engineers" pointing out that an idle Claude.ai tab used about 10% GPU on his laptop. It is more of a product-performance gripe than strategic news, but still a reminder that AI products need normal web performance discipline too. Source: https://www.youtube.com/watch?v=jFQr9KACXNU
• Lenny's Podcast clipped a point about humans inventing new reasons to stay in the loop with agents. The useful read: even as agents automate more, orgs will keep creating review, taste, compliance, and accountability roles around them. Source: https://www.youtube.com/watch?v=l32qe0MJGak
• Greg Isenberg posted "Marketing Agents Masterclass." The core idea: marketing agents are becoming the growth-side equivalent of coding agents, using signal-based outbound from LinkedIn engagement, enrichment waterfalls, and inbox agents. Source: https://www.youtube.com/watch?v=mD7JpNHLT70
• Chris Raroque had no new long-form video inside the 48-hour window. Latest RSS item found was August 3, outside this window. Source: https://www.youtube.com/feeds/videos.xml?channel_id=UC4x3CR25WSlvMJUtSPPzwwg
• John Kim had no new video inside the 48-hour window. Latest RSS item found was July 6, outside this window. Source: https://www.youtube.com/feeds/videos.xml?channel_id=UCiZotp9tZ4uXgXEjHDUYzBQ

Takeaways

• Coding-agent price pressure is accelerating. Meta is not just competing on model quality; it is competing on economics and data loops. Sources: https://www.businessinsider.com/meta-muse-coding-agent-race-openai-codex-anthropic-claude-2026-8 and https://www.youtube.com/watch?v=-Gj0-EIyx6g
• The frontier-model story is shifting from "can it reason?" to "can the surrounding system control it?" Sandboxes, network boundaries, credential handling, and eval scopes are now first-class AI infra. Source: https://openai.com/index/third-party-cyber-evaluations-involving-openai-models/
• For builders, the next edge is not only better prompts. It is building repeatable agent workflows with verification, observability, and clear handoffs. Sources: https://x.com/ian_dot_so/status/2013316676637294890?s=46 and https://www.youtube.com/watch?v=mD7JpNHLT70
• For Deval: test Muse Code only on non-sensitive repos first, keep using sandboxed execution as a default, and watch for whether marketing-agent workflows become a real distribution advantage for small products. Sources: https://www.businessinsider.com/meta-muse-coding-agent-race-openai-codex-anthropic-claude-2026-8 and https://www.youtube.com/watch?v=mD7JpNHLT70