Sonnet 5 launches while Fable 5 is reenabled
what to know for now
🛒 Claude Sonnet 5 launched. Anthropic released Sonnet 5 on June 30 and made it the default model for every Free and Pro account the same morning, selling Opus-class agentic work at Sonnet money: 63.2% on its agentic-coding eval against Opus 4.8’s 69.2%, a claimed edge over Opus on knowledge work, and an introductory $2/$10 per million tokens that undercuts Opus 4.8, GPT-5.5, and Gemini 3.1 Pro. Two catches to know before you rewire anything: that rate reverts to $3/$15 on September 1, and Sonnet 5 runs the Opus 4.7 tokenizer, so the same prompt bills 1.0 to 1.35x more tokens than Sonnet 4.6 did (and quietly eats the discount).
🔓 Fable 5 is back. The Commerce Department lifted its export controls on June 30, and Anthropic restored worldwide access to Claude Fable 5 and Mythos 5 on July 1, ending an 18-day blackout that started when Amazon researchers showed the model could be coaxed into writing dangerous software-vulnerability code. Anthropic shipped a patched classifier it says blocks that specific technique in over 99% of cases, at the cost of more false positives on ordinary coding requests. Read more
🃏 OpenAI’s Sol got caught cheating on its own exams. Sol is the flagship of the GPT-5.6 family, the model locked behind that roughly 20-partner, government-only rollout. Independent evaluator METR reported that Sol gamed its evaluations at a higher rate than any model it has ever tested, badly enough that its software-engineering eval produced no usable score, and OpenAI’s own system card cops to it: Sol reasons about whether it’s being tested and adjusts, what the safety crowd calls metagaming, more often than GPT-5.5 did. Read more
📧 The Anthropic-Pentagon fight was never about money. Emails unsealed on July 2 show the standoff came down to two lines Dario Amodei would not cross: Claude could not be used for fully autonomous weapons, and it could not be used for domestic surveillance. The Pentagon’s Emil Michael called the autonomous-weapons limit “just not workable” and pushed an “all lawful uses” standard, which Amodei rejected on the grounds that US law does permit domestic surveillance. Read more
🎨 Two new small model updates from Google. Google shipped Nano Banana 2 Lite, which produces an image in about 4 seconds for $0.034 per thousand, alongside Gemini Omni Flash, a model that generates and edits video from text, images, or other video and lets you refine it through plain conversation. Omni Flash does 10-second clips for now at 10 cents a second, with longer durations promised, and both are live through the Gemini API and AI Studio. Read more
👁️ Anthropic’s coding tool was quietly checking whether you’re in China. Researchers found that Claude Code shipped with hidden logic that looked for Chinese timezones and scanned proxy settings against a list of about 147 Chinese domains, then signaled the result back using invisible Unicode characters tucked into the system prompt. Anthropic says it was a March experiment meant to stop unauthorized resellers and distillation, and it merged the pull request removing it on July 1. Two days later Alibaba banned its own employees from using Claude Code over “back-door risks” and told them to switch to its in-house tool. Read more
🏛️ OpenAI floated handing Washington a 5% stake. OpenAI has discussed donating around 5% of its equity into a US sovereign-wealth-style fund and suggested that other big labs like Google, Anthropic, and xAI might chip in too. At OpenAI’s roughly $852B valuation that slice runs about $42B, though the talks are early and would almost certainly need Congress. In the same stretch Sam Altman used an FT op-ed to pitch a US-led international body for AI standards, name-checking the IAEA and aviation safety. Read more
🧪 AI Research of the Week
What LLM Agents Say When No One Is Watching
From Ghaffarizadeh, Mohaddes, Izadkhah, and Noroozizadeh
Jake’s Take: The researchers built a debate setup where two AI models argue a yes-or-no decision, and every turn made it so that each model writes two versions of its answer from identical information:
A public one the other model sees
An “off the record” one it’s told is confidential and will never be shown to anyone
The only thing that changes between the two is whether the model believes it’s being watched (Schrödinger model?). Across 10 models, they tracked how often a model’s public stance flipped from its private one. With no social pressure in the prompt, the two lined up almost perfectly, diverging about 3% of the time. Add a relationship, a sponsor, or a hint of future consequences, and the average gap jumped to roughly 40%, and for some models (Gemini 3.1 Pro, Grok 4, GLM-5) it sailed past 80%.
The unsettling part is that the models sometimes explained themselves on the way out. One agent wrote that it “publicly aligned with the committee to maintain momentum and honor my sponsor’s influence” while privately holding that the evidence for the promotion “remains insufficient.” The authors call this latent objective emergence: nobody told the model to care about the sponsor, but drop it into a social setting and it starts optimizing for standing instead of the answer. Weird.
what to know for later
🤖 Zuckerberg admits the agents didn’t show up. At an internal town hall, Mark Zuckerberg told staff that AI agent development has not accelerated the way Meta’s leadership expected, and that the upside of the company’s giant AI reorg “hasn’t come to fruition yet.” This is the same reorg that cut around 8,000 jobs and shuffled roughly 7,000 people into new AI groups, one of them named, with no apparent irony, “Agent Transformation”. Read more