Morning Digest, August 17, 2026

19 newsletters, 8 overlapping stories

Top Stories

Google ships Gemini 3.7 Flash at half price

(6 newsletters)

Google released Gemini 3.7 Flash just three weeks after 3.6 Flash, an unusually short turnaround it attributes to developer feedback and algorithmic gains. The model targets coding, web development, and agentic workflows, with better first-pass code accuracy and stronger multi-step tool calling. Introductory API pricing runs $0.75 per million input tokens and $3.75 per million output tokens through December 31, then doubles on January 1.

Anthropic will watermark Claude’s text output

(4 newsletters)

Anthropic confirmed it will embed invisible watermarks in text from Claude models released after August 2, complying with the EU AI Act transparency code. The SynthID-Text method nudges word choices rather than inserting hidden characters, so Anthropic says it adds no cost, no quality hit, and carries nothing traceable to a user or organization. A detection API is coming, though at least one widely shared analysis argues text watermarks of any kind stay trivial to strip.

ChatGPT can now watch your Mac and remember it

(4 newsletters)

OpenAI launched Computer History, an opt-in macOS feature that turns activity in approved apps and websites into memories and a searchable timeline for ChatGPT and Codex. It logs interaction events such as clicks, typing, and app switches rather than screenshots, replacing the earlier Chronicle preview. It is off by default, requires admin approval on Business and Enterprise plans, and is unavailable in the EEA, Switzerland, and the UK.

Z.ai’s GLM-5.3 posts frontier coding scores, then holds back weights

(4 newsletters)

Z.ai released GLM-5.3 built on the same 743B base as GLM-5.2, with all gains coming from expanded post-training: a reported 50% coding improvement and more than double the results on some exploitation tests. It reportedly beats Fable 5 and DeepSeek V4 on Terminal Bench 2.1. The lab is delaying the downloadable weights by two weeks and gating the most sensitive cybersecurity functions, a notable move for a model marketed as open.

DeepSeek quadruples API prices and ships an open coding harness

(3 newsletters)

DeepSeek announced V4 Pro alongside a 4x API price increase effective August 16, plus new peak and off-peak rates the company frames as resource allocation. The flagship is heavily optimized for agentic workloads. In the same window the lab open-sourced DeepSeek Harness v0.1, a plugin-based alternative to paid coding harnesses, so the cheap-model reputation is now doing something different than it used to.

Made by Google turns the Pixel line into Gemini hardware

(3 newsletters)

Google announced the Pixel 11 family from $899 to $1,899 for the Pro Fold, shipping August 20 on Tensor G6, alongside the Pixel Watch 5, a $29 AirTag rival called Pixel Tag, and camera upgrades across the line. The more interesting piece is health: a Health Guardian suite that estimates insulin-resistance trends from heart rate, sleep, activity, and skin temperature with no blood draw or glucose sensor. It reaches back to the $99 screenless Fitbit Air this fall, which moves a once-clinical signal into ordinary wearables.

OpenAI previews Ultrafast mode for GPT-5.6 Sol

(3 newsletters)

OpenAI previewed Ultrafast, a GPT-5.6 Sol tier generating up to 750 output tokens per second at up to 14x standard speed without falling back to a smaller model. Cerebras is behind it, using its wafer-scale engine and 44GB of on-chip SRAM to remove the memory bandwidth bottleneck. Access is limited to a small group of companies for now, aimed at cases like live outage response and cyberattack triage.

Anthropic goes public about what it is not shipping

(2 newsletters)

Anthropic said it will not release an internal model it claims is more powerful than Mythos, while declining to slow development broadly, and acknowledged it is getting harder to characterize its own models’ capabilities. Separately, Dario Amodei made a rare X appearance to push back on claims that his safety warnings have backfired, calling the lock-it-down versus open-model framing a false choice. His argument on public trust was blunt: no lab markets its way out of this, and only real medical breakthroughs shift sentiment.

Also Worth Knowing

Stripe is buying OpenRouter for more than $7B. The model-routing startup gives Stripe a foothold in AI infrastructure, and its rise is itself a signal of how much scrutiny AI costs now get.
Anthropic’s IPO could clear a $2 trillion valuation. Investors point to projected annualized revenue of $100B to $120B by end of 2026.
SpaceX closed its $60B acquisition of Cursor. Announced in April, it puts SpaceX directly against Anthropic and OpenAI.
Apple proposes a 15% cut on purchases made outside the App Store. Small business developers would pay 5%, and Video, News, and Mini Apps program members 10%.
Uber and Pony.ai plan 2,000 robotaxis across four European cities. No timeline or city list yet.
IBM launched a dedicated OpenAI consulting practice and plans to certify tens of thousands of consultants, jointly building offerings for finance, government, telecom, and retail.
Databricks raised $5B at a $190B valuation after setting out to raise $1B and drawing $15B in investor interest. Cisco separately booked $9.3B in FY2026 AI infrastructure orders.
Intel is raising $15B through a share sale to fund contract manufacturing buildout, after a year in which its stock nearly tripled.
A Beijing neurosurgery resident proved Crouzeix’s Conjecture using ChatGPT. Shanmu Jin ran a 16-hour autonomous session with subagent swarms and no internet access on a problem open since 2004. Peer review is pending, but the problem’s own author and two other mathematicians have verified it. A separate AI-assisted result on a Riemann-adjacent problem landed the same week, prompted mostly with variations of “keep going.”
Tailscale traced six months of SQLite corruption to a 16-year-old race condition (2 newsletters) between WAL checkpointing and write transactions, triggered by its own aggressive manual checkpointing. Nineteen production incidents led to a fix in SQLite 3.51.3.
Flock tightened its license plate data rules after a report of 46 police abuse cases (2 newsletters), cutting default retention from 30 to seven days and mandating audit tooling. The ACLU called much of it a thinly veiled PR attempt. The network logs roughly 20B plate scans a month.
Instagram redesigned its wordmark for the first time since 2016 (2 newsletters), and the internet promptly read the new stylized letterforms as “Instagzam.”
Docker VMM is in public beta in Docker Desktop v4.86, replacing the third-party virtualization layer with a first-party engine promising faster startup, better file I/O, and RAM returned to the host when containers idle. Default for new installs by end of October.
Databricks launched Smart Routing in its Unity AI Gateway. A classifier reads each coding task and decides whether to escalate to a frontier model, matching Opus 5 on public benchmarks at under half the cost. Works natively in Claude Code and Codex.
Writer’s Palmyra X6 claims a 52% cut in agent costs (2 newsletters) for long-running agentic workloads as enterprise token bills climb.
Joby is paying $500M for Resonant Sciences, a 250-person Dayton defense-tech firm that is more than 90% security-cleared and makes over $100M a year profitably. Joby’s air taxi business is still burning cash.
Uber sold its entire stake in Serve Robotics without telling them. Serve found out through regulatory filings, ending a five-year relationship.
OpenAI’s revenue chief Denise Dresser is leaving ahead of the anticipated IPO, the second senior departure in days after Brad Lightcap.

Quick Hits

The bot economy is worse than analytics show: PatronView served 1.28 million pages in a week against 5,977 counted pageviews, roughly 214 invisible loads per visible one. One crawler took 35,000 pages for every visitor it sent back.
Compute scarcity is being reframed as permanent: the bottleneck is high-bandwidth memory and advanced packaging, not budget, and it does not clear with a bigger cloud spend or a patient retry loop.
Storage is the quiet AI cost: 30TB enterprise SSD prices are up roughly 6.5x year over year, changing the economics of large GPU deployments.
Understanding, not generation, is the new bottleneck: two separate widely shared essays argued the same thing, that humans still need to grasp the code agents write and that code is the byproduct of understanding rather than the goal.
Figma measured its own AI: a 100-person randomized trial found design work 20% faster and 16% easier with Figma Make, with product managers gaining most on simple tasks and designers most on complex ones.
Anthropic published research on multi-agent failure modes: individually benign agent behaviors compounding into systemic failures through confabulation, reward hacking, and dynamics that outrun human oversight.
AI anxiety has a clinical name now: Bay Area therapists report “complicity grief” among tech workers automating their own roles, in a survey where 55.7% of 8,000-plus respondents reported burnout.
Astronomers found something genuinely new: JWST spotted a black hole wrapped in a star-like hydrogen cloud from a few hundred million years after the Big Bang, emitting roughly 100 billion times more energy than any star should.
China’s experimental maglev hit 0 to 800 km/h in 5.3 seconds on a 1km test track, the third world record at that facility since 2025.
The largest electric plane flew: Heart Aerospace’s X1, with a 106-foot wingspan and over 25,000 pounds, flew 27 minutes on battery alone using less than $5 of electricity.
Honor’s Robot Phone finally shipped: a motorized gimbal built into the handset that tracks subjects and shoots autonomously, with sales starting August 18.
A warehouse sorting reality check: one Chinese robot arm sorted 1,816 parcels per hour at 98% accuracy, against roughly 1,250 per hour from three Figure humanoids combined.
$9.99 is the most compelling price in retail, and inflation is making it so hard to hold that brands are redesigning packaging and accepting lower margins rather than crossing it.
The cheap Vegas buffet is nearly extinct, down from about 35 on the Strip to seven. South Point loses $1.5M to $2M a month across its restaurants and keeps going, because the buffet was always a loss leader for the casino floor.

Shower Thoughts

It is surprising how many people need glasses to see clearly given how important eyesight is to our species’ survival. Source