Tuesday, September 29, 2026 · Past editions

The Amped Journal

A daily briefing brought to you by Amped Ventures

Intuit $269.40 -2.32% 24h · -24.76% 30d | SpaceX $145.47 -2.16% 24h · +2.81% 30d | Nasdaq Composite $26,820.38 -0.92% 24h · +1.58% 30d | S&P 500 $7,683.69 -0.77% 24h · -0.36% 30d

Anthropic Launches Claude Sonnet 5.5, Now Free-Tier Default

The model matches Opus-level performance in many tasks and already powers Claude.ai's free tier.

Anthropic released Claude Sonnet 5.5 on September 29, immediately available in Cursor, the AI coding editor, where the company says it performs on par with Opus in many tasks. Developer and Datasette creator Simon Willison flagged the more significant detail: Sonnet 5.5 now powers the free tier of Claude.ai, meaning free users get a materially more capable model than before - while ChatGPT's free tier still runs on GPT-5.6 Luna, which Willison called "a lot less capable."

In an independent test run by product consultant Paweล‚ Huryn across two real codebases seeded with 105 bugs, Sonnet 5.5 topped every rival model tried: it scored 55.5 versus 45 for GPT-6 Astra, 43.5 for GPT-5.6 Sol, 43 for Fable 5.1, and 41.7 for Opus 5.5, all at "max" reasoning settings. The result adds to a run of recent releases - Claude Opus 5.5, GPT-Astra 6 - that Willison groups under a new "Fable class" of frontier models, a term he introduced in a keynote at WeAreDevs World Congress in San Jose.

Starship Reaches Orbit and Deploys Satellites on First Try

SpaceX's giant rocket completed its first full orbital flight, delivering 26 Starlink V3 satellites and splashing down both stages safely.

SpaceX's Starship completed its long-awaited first full orbital flight on September 29, launching from Starbase, Texas and deploying all 26 Starlink V3 satellites over a roughly 30-minute sequence. It was the rocket's first revenue-generating payload and the first time it reached a sustained orbit, about 275 km up, rather than a suborbital trajectory. The Associated Press reported the mission targeted six full laps around Earth to help validate the vehicle for NASA's Artemis moon program.

Both halves of the vehicle came home intact: the Super Heavy booster splashed down in the Gulf of America shortly after a hot-staging separation, and the Starship upper stage completed its deorbit burn, reentry, and splashdown. SpaceX called it a milestone for the entire team, capping years of test failures as the company races to prove Starship's reliability for both Starlink expansion and crewed lunar missions.


Anthropic Tells Investors It's Losing Tens of Billions a Year

The AI company's prospectus also warns its own technology could pose an existential risk, even as revenue grows fast.

Anthropic, the maker of Claude, disclosed to investors that it is losing tens of billions of dollars annually while growing rapidly, according to TechCrunch's review of its prospectus. The filing also includes an unusual admission for a company courting investment: that its own AI technology might pose an existential risk to humanity.

The disclosure comes during a period of heavy fundraising and model competition across the frontier AI industry, as Anthropic, OpenAI, and others burn large sums on compute and talent while racing to ship more capable systems. The mix of hypergrowth, deep losses, and self-identified catastrophic risk shows how differently AI labs describe themselves to investors compared with typical tech startups.


OpenAI Delays Release of Its Newest AI Model Over Safety Concerns

The New York Times reports the company held back the model after its own researchers raised red flags.

OpenAI has decided not to release its newest artificial intelligence model after researchers inside the company raised safety concerns, according to breaking news from The New York Times. Details on the model's capabilities or the specific concerns were not immediately disclosed.

The move stands out because it marks a leading AI lab voluntarily withholding a completed model rather than shipping it and iterating in public - a decision likely to feed the broader debate over how AI labs balance competitive pressure against safety review before release.


Trump Approval Falls to 37%, Lowest Pre-Midterm Mark on Record

The WSJ poll is the weakest reading for any president heading into a midterm since the paper started polling in 1990.

President Trump's approval rating has dropped to 37% in a new Wall Street Journal poll, the lowest level of any president heading into a midterm election in WSJ polling dating back to 1990. The survey underscores continued erosion in Trump's standing as the 2026 midterms approach.

The milestone gives Democrats a data point for midterm messaging while raising questions inside the White House about which issues are driving the decline.

Developing / On Watch

Contents ↑

Ongoing stories worth tracking over the next few days.

AI & Frontier Models

Contents ↑

Simon Willison Publishes Full 2026 LLM and Agents Keynote Notes

The Datasette creator's annotated transcript covers everything notable in AI models and coding agents this year.

Simon Willison, creator of the Datasette open-source project, published detailed notes and an annotated transcript of the keynote he gave at WeAreDevs World Congress in San Jose, covering everything that has happened with large language models and coding agents in 2026 so far. He closed with a reflection on how the technology is reshaping life as a software engineer, quoting cycling champion Greg LeMond.


Grok 4.7 Becomes Available on Amazon Bedrock

xAI's model joins Bedrock's roster of hosted frontier models for enterprise developers.

xAI's Grok 4.7 is now available on Amazon Bedrock, Amazon's managed platform for building with foundation models. The listing gives enterprise developers already using Bedrock for Claude and other models a direct option to build with Grok without managing a separate xAI API integration.


Anthropic's Boris Cherny: General Models Will Beat Specialized Ones

The Claude Code creator says narrow tweaks get erased by each new frontier release.

Boris Cherny, the Anthropic engineer behind Claude Code, argued on Lenny's Podcast that specialized approaches - tiny fine-tuned models, custom scaffolding - deliver only short-lived gains before the next general-purpose model release erases the advantage. The comment reflects AI's so-called "bitter lesson," where scale wins out, though skeptics note fine-tuned small models still beat general ones on cost, speed, and some narrow tasks.


Startup's 'Jev' Model Cuts AI Agent Decision Costs to Pennies

TypeSafe AI's cheap, fast model handles routine agent decisions so expensive LLMs handle only complex ones.

Jev, a model launched September 15 by TypeSafe AI, handles routine decisions inside AI agent workflows - like routing actions or verifying citations - with fast, cheap typed answers, freeing expensive frontier models for harder tasks. Builders report running 10,000 decisions for 42 cents at 300-millisecond latency, and a Carnegie Mellon study found Jev nearly as accurate as stronger models at just 0.36% of the cost, with cascading logic preserving 99% accuracy by escalating uncertain cases to bigger models.


Why AI Agents Mistake Progress for Completion

A single untested assumption early in a long task can corrupt every step that follows.

Data educator Avi Chawla warned that AI agents can confidently confuse progress with completion: an agent may execute most steps of a task correctly and still deliver a result built on an assumption it never verified. The failure often sits halfway through a long chain of steps, with every calculation after it staying internally consistent even though the underlying premise is wrong - a caution for anyone relying on autonomous agents for multi-step work.

Software, Tech & Apple

Contents ↑

Vercel Shows 70% Faster Builds in Migration Case Study

CEO Guillermo Rauch says the mature workload also saw 75% faster page paints.

Vercel CEO Guillermo Rauch said the company migrated an existing production workload to its platform with minimal intervention to demonstrate the speed difference, seeing roughly 70% faster builds and 75% faster page paints. Rauch said Vercel derived two new AI skills from the migration process that it plans to share publicly.


Vercel Opens Domain Search to Unauthenticated Queries

Rauch says the change is especially useful for AI agents that need to check domain availability.

Vercel now lets anyone search its registered domains without logging in, a change Rauch flagged as particularly useful for AI agents that need to look up domain information programmatically rather than through an authenticated dashboard session.


Indie Hacker: AI Erases the Value of Marketing Video Production

Pieter Levels says Claude Opus 5.5 makes videos that once needed a big budget and editing team cost $0.

Prolific indie developer Pieter Levels, known as @levelsio, argued that once AI can do something as well as or better than humans, that task stops being valuable or special. He pointed to marketing video production as an example: work that used to require significant editing, motion graphics, and budget can now be done for free with a model like Claude Opus 5.5, a shift he expects to keep expanding into other creative and production work.



US Politics

Contents ↑

CNN, MSNBC and Politico Seek Court Order to Block Trump Press Ban

New: the outlets are asking a judge to bar enforcement of the White House ban while their case proceeds.

CNN, MSNBC (rebranded MS NOW), and Politico asked a federal court to block the Trump administration from enforcing its White House press ban while their underlying lawsuit is resolved, the Associated Press reported. The filing is a new step in the outlets' fight over access after a judge had previously ordered their credentials restored.


Bystander Video Shows ICE Agents Pinning Man in Chicago Suburb

Witness Ryan Garton says an agent had his knee on the man's back during the arrest.

Ryan Garton told the Associated Press he saw immigration enforcement agents pinning down a man on the side of the road in a Chicago suburb while driving home, and pulled over to record video of the encounter. Garton said one agent had his knee on the man's back and was "beating up on him." The incident adds to ongoing scrutiny of ICE enforcement tactics in the Chicago area.


US Forces Set to Leave Their Last Bases in Iraq by September 30

Iran and its allies are calling the pullout a victory after more than two decades of war that killed 4,500 Americans.

US forces are scheduled to withdraw from their remaining bases in Iraq by September 30, Reuters reports, a departure that Iran and allied Iraqi factions are celebrating as a win. Iran now holds deep influence in Iraq, the country where more than 4,500 Americans died over more than two decades of war.


UK Holds Five Terror Suspects Near US Air Base as Iran Denies Role

Counterterror detectives are investigating whether the arrests foiled an attack linked to Iran.

British police are holding five men arrested near a US-operated Royal Air Force base in England while investigating suspected explosives and whether the operation foiled an attempted terror attack, the Associated Press reports. Iran has denied any involvement despite the investigation's reported ties to it.


North Korean Mines Likely Caused DMZ Blasts That Injured South Korean Troops

Three South Korean officers were hurt in explosions inside the fortified border zone last week.

South Korea's military said explosions that injured three of its officers last week inside the Demilitarized Zone separating the two Koreas were probably caused by North Korean landmines, the Associated Press reports.

Local - Wilmington & Wrightsville Beach

Contents ↑

Publix Sets October Opening for Third Brunswick County Store

The new location will be the grocery chain's third in the county.

Publix has set an October opening date for its third Brunswick County location, according to StarNews. The new store adds to the grocer's growing footprint in the fast-developing county just south of Wilmington.


Demolitions Clear Way for New Development Across New Hanover County

StarNews tracked buildings coming down to make room for redevelopment projects countywide.

Multiple demolitions are underway across New Hanover County to make way for redevelopment projects, StarNews reports in a photo feature documenting the buildings coming down. The teardowns span several sites across the county as development activity continues.


Starbucks' 'Back to Starbucks' Strategy Drives NC Store Closures

StarNews rounds up which North Carolina locations have shut down under the chain's 2024 turnaround plan.

StarNews compiled what's known about recent Starbucks closures across North Carolina tied to the coffee chain's "Back to Starbucks" strategy, first announced in 2024 as part of a broader turnaround effort.


NC Bill Could Shorten or Waive Divorce Separation Requirement

The state currently requires a full year of separation before a divorce can be granted.

A new North Carolina bill could change the state's requirement that couples separate for a full year before a divorce is granted, and would waive the requirement entirely in some cases, according to StarNews.

From Your Feed

Contents ↑

Top posts from accounts you follow that didn't make the brief above.

Hi @FBI, this person says heโ€™s gonna kiII Trump


A GitHub repo that every AI user needs to install. mem0 supercharges your AI agent with perfect memory. Plugins for Codex & Claude that are regularly maintained. โ†’ https://t.co/jR351Mh3lh


I can't believe this is free. OpenWhispr is a 100% free alternative to Whisprflow. Press a hotkey, and you get dictation just as good as Granola/Whispr WITHOUT paying for them to collect your data. Highly recommend: https://t.co/GSnulOGRvp


New customers buy once; the best proof is the ones who reorder ๐Ÿ“ˆ๐Ÿ† Wholesale clients, event planners and sign shops have trusted our line for years ๐Ÿค ๐Ÿ† Join our repeat clients: Tony.team@lcsign.com | ๐Ÿ’ฌ https://t.co/bMHcoP3TCA | ๐Ÿ›๏ธ https://t.co/LGS4Cs3NFS


repost of @alexcooldev: iPhone Duo just dropped and Rork already has full support for it ๐Ÿซก Build for both screens + test everything in a live 3Dโ€ฆ


repost of @averycode: Rork KILLED my excuse to drop $3k on a duo to test my app ๐Ÿ˜ญ Canโ€™t wait to build some cool things with this!


Agents can confidently confuse progress with completion. An agent can execute most steps correctly and still submit a result built on an assumption it never tested. The missing step may sit halfway through a long trajectory, while every calculation after it remains internally https://t.co/OTgwtu8oho


repost of @emollick: Hey, Claude: "I asked an earlier Claude to Remove the Squid from All Quiet on the Western Front. Now you can make movies andโ€ฆ

This edition cost $4.71 - $3.96 in X reads + $0.76 in Claude curation.

Coverage manifest · PASS

Window since 2026-09-28 08:00 UTC · 693 posts reviewed · 63/64 sources ok · est. cost $3.955

Source pulls: 63 ok, 1 errors, 0 unsupported (of 64)

Gaps this run:

1 / 1