In partnership with
In Todayโs Issue:
๐ Google's Nano Banana 2 Lite: four-second images for a rounding error
โก Google's AI boom sends power and water use soaring
๐ฌ Anthropic hands scientists an AI workbench
๐ญ Sonnet 5 lands weaker than expected, and Fable 5 returns on a leash
๐งฌ OpenAI's GeneBench-Pro stumps the entire field
โจ And more AI goodnessโฆ
โก The Signal
Anthropic just shipped a model called Sonnet 5, and the striking part is how little it actually advances the frontier.
The new mid-tier model beats its predecessor, Sonnet 4.6, yet still trails Anthropic's flagship Opus 4.8 on essentially every benchmark, which makes the jump to a "5" label read more like marketing than capability. On the same day, Anthropic quietly redeployed its most powerful model, Fable 5, but wrapped it in a new safety classifier that reroutes risky requests back to the older Opus 4.8. Read together, the two moves point to a lab tapping the brakes: the genuinely powerful frontier releases are stuck in talks with regulators, so what actually reaches us is deliberately restrained. Meanwhile Google shipped a four-second image model and an environmental report showing its AI electricity demand up 37% in a year, while OpenAI dropped a biology benchmark so hard the best model still fails two of every three problems. The frontier is still moving, just not where the version numbers suggest.
All the best,
Kim Isenberg
๐ Google's Nano Banana 2 Lite Makes Images Almost Free
Google just made image generation almost free. Its new Nano Banana 2 Lite produces a picture in about four seconds for $0.034 per 1,000 images at 1K resolution, keeping the character consistency and readable text that made the original a hit. It landed alongside Gemini Omni Flash, a conversational video model at $0.10 per second, and both are already wired into Search's AI Mode, the Gemini app, Google Photos, and Ads.
๐ tl;dr: The price of a generated image is rounding down to zero.
โก Google's AI Boom Sends Its Power and Water Use Soaring
Google's own 2026 environmental report shows the AI race measured in kilowatt-hours. Electricity demand jumped 37% in a year, water use climbed 34% to 10.9 billion gallons, and only a record 12 GW of newly signed clean energy and 86% carbon-free electricity kept operational emissions down a slim 2%. The efficiency gains are real, but the AI buildout is growing faster than Google can green it.
๐ tl;dr: The goal has quietly shifted from cutting emissions to stopping them from exploding.
๐ฌ Anthropic Gives Scientists Their Own AI Workbench
Anthropic launched Claude Science, an AI workbench aimed squarely at researchers. It bundles 60+ curated skills and connectors for genomics, proteomics, and structural biology, runs analyses locally or on HPC clusters, and ships a reviewer agent that checks citations and calculations before you trust the output. Early users report real speedups: a UCSF epidemiologist says it cut a genetic analysis by about 90%, and an Allen Institute neuroscientist compressed a two-year review into weeks.
๐ tl;dr: The lab notebook is becoming an agent that runs the experiment with you.
Stop paying for image generations you can plan.
Why it helps: With models like Nano Banana 2 Lite dropping to fractions of a cent per image, cost is no longer the bottleneck. Knowing exactly what to ask for is, so you do not burn a hundred near-misses.
Try this: "Act as my art director. Before generating anything, ask me five sharp questions about audience, format, mood, brand colors, and the one thing the image must communicate. Then write three distinct, fully specified image prompts I can run, each with a one-line rationale."
๐ฌ Watch This
Open-source models just got close enough to make enterprises rethink their stack. This week's panel opens on Zhipu's GLM 5.2, the Chinese open model closing in on the American frontier on agentic benchmarks while staying free to download, fine-tune, and self-host. From there it turns practical: Box CEO Aaron Levie on how enterprises actually pick models once capable open source is on the table, Harvey co-founder Gabe Pereyra on building specialized legal AI atop open infrastructure, and Bernstein analyst Stacey Rasgon on OpenAI's new Jalapeรฑo inference chip and what the race to cut costs means for Nvidia and Broadcom. A rare conversation that ties a model release to the money underneath it.
"Unfortunately, the pace of advances is still much greater than the pace of [progress in] how we can manage those risks and mitigate them. And, that, I think, puts the ball in the hands of the policymakers."
โ Yoshua Bengio, Turing Award laureate and AI safety pioneer
A developer's reverse-engineering report has kicked off a dustup over Claude Code, Anthropic's command-line coding agent. According to the report, recent builds allegedly detect when a user routes through a non-official API endpoint (the ANTHROPIC_BASE_URL variable) from a Chinese timezone, then quietly fingerprint the request by tweaking the system prompt: reportedly switching the date separator from a dash to a slash and swapping in a look-alike apostrophe to flag China-linked proxies and AI-lab domains. Critics call it spyware-like, arguing that an agent which can read repositories and run commands should not be hiding routing metadata inside its prompts. Anthropic has acknowledged the behavior and, per the same report, says the code is being rolled back in the next release. Intent aside, it is a rough look for a company that sells trust, and a reminder to check what your coding agent puts into its own prompt
Anthropic's Release Day: Fable 5 Is Back, Sonnet 5 Lands, Both With a Caveat
The Takeaway
๐ Sonnet 5 beats Sonnet 4.6 but trails Opus 4.8 on essentially every eval, so the "5" reads as branding, not a real jump.
๐ Standard pricing is unchanged at $3/$15 per million tokens (launch promo $2/$10 through Aug 31): Sonnet money, sub-Opus performance.
๐ Anthropic also redeployed Fable 5 as export controls lifted, but bolted on a classifier that reroutes risky requests to Opus 4.8.
๐ No Opus 5 and no Fable successor, despite Fable 5 already beating Opus 4.8, all point to deliberate restraint while regulators weigh in.
Anthropic had a two-model day on June 30, and both releases come with an asterisk. Start with Claude Sonnet 5, the new mid-tier model. It is clearly better than Sonnet 4.6, which is no surprise, but across the board it still lands weaker than Opus 4.8, Anthropic's current flagship. That makes the "5" label hard to justify: a major version bump is supposed to signal a real leap, and this is not one. Sonnet 5 is good, but worse than expected. Standard pricing is unchanged from its predecessor at $3/$15 per million tokens (with a launch promo of $2/$10 through August 31), so you pay Sonnet money for sub-Opus work. Opus stays more expensive, and stays better.
Sonnet 5 improves on Sonnet 4.6 but sits below Opus 4.8 across evaluations. (Anthropic)
The bigger story is the one Anthropic barely mentioned: Fable 5 is back. Its most capable model had been pulled worldwide on June 12, when fresh US export controls hit both Fable 5 and Mythos 5 and the company had no reliable way to verify a user's nationality in real time. Those controls were lifted on June 30, and Fable 5 returned on July 1 across Claude, the API, Claude Code, and Cowork. Through July 7 it is bundled into paid plans for up to half of weekly usage; after that it runs on usage credits.
But the comeback comes on a leash, and the reason is instructive. Fable 5 was risky in the first place because Amazon researchers found a jailbreak that turned it into a capable generator of working software exploits, and every model they tested could be pushed to do the same. Anthropic's fix is a new safety classifier that blocks the bypass in over 99% of cases and, when it trips, quietly reroutes the request to the older Opus 4.8 instead. That keeps the most dangerous edge cases off the most powerful model, at the price of the occasional false positive on legitimate coding work. It is the same instinct behind Sonnet 5, which Anthropic deliberately trained to be weak at cyber-exploit tasks.
Put the two together and a pattern emerges. We got a new "5" series that is not meaningfully ahead of what we already had, and still no Opus 5, even though Fable 5 already outperforms Opus 4.8 and a stronger Opus almost certainly exists internally. Why hold back? Because restraint is the order of the day. The genuinely powerful frontier releases appear to be delayed across the board while labs work through with regulators how, and under what conditions, they can ship at all. Seen that way, Sonnet 5 looks like a placeholder to stay in the conversation, and Fable 5's return looks less like confidence than a carefully hedged bet. For a release day, it is a strangely cautious one.
Why it matters: When the most interesting thing about a frontier lab's launch is what it chose not to ship, the bottleneck has moved from engineering to governance. The models look ready; the permission to release them is not.
Sources:
๐ https://x.com/claudeai/status/2072017450611142835
๐ https://www.anthropic.com/news/redeploying-fable-5
๐ https://www.anthropic.com/news/claude-sonnet-5
HR and IT need to work as one. Here's how
Every missed onboarding step, delayed offboard, or broken provisioning handoff has a root cause: HR and IT aren't aligned. This guide gives both teams a shared framework for the full employee lifecycle.
The chart: OpenAI's new GeneBench-Pro benchmark tests AI agents on 129 hard problems across genomics, quantitative biology, and translational medicine, scored as pass rates at maximum reasoning. Even the best model, GPT-5.6 Sol (Pro), tops out at just 31.5%. The strongest non-OpenAI model is Anthropic's Opus 4.8 at 16.0%, ahead of Gemini 3.5 Flash (8.1%), GLM 5.2 (4.6%), DeepSeek V4 Pro (2.4%), and Grok 4.3 (1.5%).
The lesson: Frontier models are becoming genuinely useful for biology, but "useful" still means failing roughly two of every three expert-level problems, and OpenAI's lead in this domain is wide.
The caveat: It is OpenAI's own benchmark, so weigh the framing accordingly. A pass rate at "max reasoning" also hides cost and latency: 31.5% that needs enormous compute is not the same as a cheap, reliable tool.
๐ง Why Bigger AI Models Just Learn More
โก Bottom line: New research shows big AI models do not just know more, they can learn skills that smaller models never pick up, no matter how long you train them.
๐ก Why it matters: It explains scaling laws, the reason labs keep spending billions to build bigger models instead of just feeding smaller ones more data.
๐ What it means: Some abilities only appear once a model is big enough. More data alone will not unlock them.
We all know the rule of thumb: bigger AI models are smarter. What nobody could cleanly explain is why simply adding size unlocks brand-new abilities. A new paper (Huang et al., 2026) offers a refreshingly simple answer, and it comes down to crowding.
Some skills can be unlocked only by making the model bigger; no amount of extra training data gets you there. (Huang et al., 2026)
Picture a neural network as a room with limited space. While it trains, the common, easy patterns show up constantly and grab most of that room, while rare or hard skills appear so seldom they keep getting pushed out before they can settle in. In a small model there is simply not enough space, so those rare skills never stick. Make the model bigger and there is finally room for both: the common stuff stops crowding out the rare stuff, and the harder skills survive.
As models get wider (toward the right), they finally pick up the rarer skills that smaller ones keep missing. (Huang et al., 2026)
The researchers confirmed this on real language models, from tiny to large. The lesson is practical: a lot of a model's usefulness hides in its rare skills (the unusual question, the tricky edge case, the niche task), and those only show up once you scale the model up. So size is not just about storing more facts. It decides which abilities a model can have at all, which is why the big labs keep building bigger.
Hampton took $440K in planned hires off the calendar
Hampton co-founder Joe Speiser had three roles budgeted: a data engineer, an ops manager, a PM. $440K. He installed Viktor on April 12. Forty-four days later, none are on the calendar, and 18 of his team work with Viktor daily. His VP: we are editors now, not creators.
