A clearer view of AI.
Submit for reviewSubscribe

WEEKLY NEWSLETTER

The AI Insider

A weekly round-up of the AI tools worth your time, and the hype to ignore.

Unsubscribe at any time using the link in each newsletter.

THE INDEPENDENT AI.

AI NEWS FOR BUILDERS · 20 STORIES

AI news, 20 November 2026

From the community

Part 2/12: You can buy the shares, not the say: who really runs Anthropic after the listing

Anyone can now buy Anthropic shares, but the people buying them can't decide who runs the company. Under the final prospectus, all public Class A shareholders together get at most 49.9% of the vote. A trust funded by Anthropic elects most of the board, and the seven co-founders hold the rest of the control.

This is Part 2 of our 12-part read of Anthropic's prospectus. Part 1 covered how Anthropic's models got into real computer systems and how that risk moved to the top of the filing: theindependentai.com. Our opening-day story is here: theindependentai.com. We quote the final prospectus filed with the SEC on 19 November (Form 424(b)(4)). Page numbers are the printed page numbers in that document. Some governance pages are one or two pages earlier than in Amendment No. 3 of 17 November.

What Anthropic says

Anthropic says its Long-Term Benefit Trust (LTBT) "is intended to operate as an independent body of five financially disinterested voting trustees ('Trustees') with expertise in super intelligence safety, national security, public policy, and social enterprise" (Prospectus Summary, "Our Governance Structure and Its Goals", p.14). It also says "neither the LTBT nor the Trustees have decision-making authority in the day-to-day management or strategy setting of the company" (p.14). The company is upfront that the design is unusual. It calls it "a novel capital and corporate governance structure that has no direct public company precedent" (Risk Factors, "Risks Related to our Governance Structure", p.111).

What the filing actually shows

The public's vote is capped. "Our Class F common stock is entitled to a fixed 50.1% of the voting power of our outstanding capital stock… As a result, the voting power of the holders of Class A common stock is limited to an aggregate of 49.9% of the total outstanding voting power" (Principal and Selling Stockholders, p.298). Class F is a single share held by a company owned by the seven co-founders, the "Founder LLC" (Prospectus Summary, p.14).

What a new buyer actually gets. The 980 million Class A shares sold in the IPO are about 14.8% of the 6,626,752,642 Class A shares outstanding after the offering (The Offering, "Certain stockholders", p.22). That works out to about 7.4% of the total vote for every IPO buyer combined, before the underwriters' option (TIA arithmetic).

Two bodies pick the whole board. "Collectively, the LTBT and the Founder LLC will have the ability to elect all members of our board of directors" (Prospectus Summary, p.14). Through its single Class T share, the LTBT can "elect and remove the Class T directors, who constitute a majority of the authorized members of our board of directors" (Our Governance Structure and Its Goals, "Changes in Voting Power and Director Election and Removal Rights", p.261). The Class T seats are Mr. Hastings, Dr. Narasimhan and Mr. Liddell, plus one vacancy. Class A and Class F, voting together, elect the other three: Dr. Amodei, Ms. Amodei and Ms. Razavi. "Following the completion of this offering, the LTBT intends to fill the vacant Class T director seat once a candidate is identified" (Management, "Board of Directors", p.264).

Anthropic pays for the trust. "We provide funding to the LTBT… to support its operations and the compensation of its Trustees. In the years ended December 31, 2023, 2024, and 2025, we paid $2,716,750, $0, and $906,089 to the LTBT, respectively. During the nine months ended September 30, 2026, we paid $8,425,213 to the LTBT" (Our Governance Structure and Its Goals, "Certain Relationships and Transactions with the LTBT and its Trustees", p.254). The nine-month 2026 figure is about 9.3 times the whole of 2025 (TIA arithmetic). "The LTBT is accounted for as a consolidated variable interest entity ('VIE') on our balance sheet" (p.254).

Trustee pay. "Each Trustee receives annual cash compensation of $400,000 for their service as a Trustee" (p.254). The Chair, Dr. Shah, receives an extra $25,000 (p.255). The Trust Agreement says Trustee pay is fixed annually by the Trustees, and it "will not vary with or be conditioned upon the profits or any action or inaction of our company" (p.255; also Risk Factors, p.112). LTBT policies bar Trustees from directly buying or holding Anthropic securities (p.255).

Picking the trustees. The Trustees appoint new Trustees by majority vote. They must give the board's nominating and governance committee (NGC) 30 days' notice, and the NGC can waive that. Removal needs either a majority of all authorised Trustees, or unanimous approval of the other Trustees plus the NGC's approval (Our Governance Structure and Its Goals, "Trustees of the LTBT", p.253). The filing adds: "In practice to date, the Trustees and company executives have collaborated closely on sourcing, interviewing, and selecting Trustees" (p.253). One of the five Trustee seats is vacant (p.253).

A disclosed payment linked to a Trustee. "In August 2026, we made a $450,000 charitable donation to CNAS, of which Mr. Fontaine is the Chief Executive Officer" (p.255). Richard Fontaine has been a Trustee since April 2025 (p.254). The filing says the money supports CNAS research on AI and security (p.255). We make no suggestion of impropriety. It is listed among the company's transactions with the LTBT and its Trustees.

No end date for the trust's power. "Our Class T common stock will remain outstanding indefinitely" unless it is amended, terminated or redeemed (Risk Factors, p.112). Changing it takes 80% of the board and 80% of the Trustees if the Trustees agree. If they don't, it takes combinations of the founders, employees and a majority of Class A. If the founders' share is gone, it takes "85% of our outstanding shares of Class A common stock" (Our Governance Structure and Its Goals, "Changes to the LTBT and Class T Common Stock", p.253).

The founders' control ends with the founders. The Class F share can be redeemed by a founder supermajority or on the Founder LLC's dissolution. Otherwise, once fewer than three Continuing Founders remain for 90 days in a row, a 450-day transition starts, after which it is redeemed (Our Governance Structure and Its Goals, p.255).

The big investors have no vote either. Amazon and Google will hold 14.6% and 13.4% of all outstanding capital stock, in Class N shares (The Offering, "Certain stockholders", p.22). "our Class N common stock is not entitled to vote", except where the law requires (Risk Factors, p.121).

The risk factor says it plainly. The structure "may effectively eliminate the ability of holders of our Class A common stock to influence important decisions, including a change in control" (Summary Risk Factors, p.18; Risk Factors, p.111).

Controlled company. Because the LTBT can elect a board majority, Anthropic is a Nasdaq "controlled company" and could opt out of some governance rules. It says it does "not expect to rely on this exemption except with respect to our NGC, which will not consist entirely of independent directors" (Management, "Director Independence and Controlled Company Exemption", p.266).

The founders sold some shares; their control is unaffected. Each of the seven co-founders sells 3,378,378 shares in the offering (Principal and Selling Stockholders, p.298). At $83 that is about $280 million each before fees (TIA arithmetic). Their control comes through the Class F share, which doesn't depend on how many Class A shares they hold. Directors and executive officers as a group hold 3.2% of the total vote after the offering (p.298).

What changed in the final version. The Amazon and Google stakes were updated for the larger offering: 14.6% and 13.4%, down from 14.7% and 13.7% in Amendment No. 3 (The Offering, p.22). The governance terms quoted here are unchanged from Amendment No. 3. Only the page numbers moved, for example LTBT funding from p.255 to p.254, Class T board control from p.262 to p.261 and the 49.9% cap from p.300 to p.298. Amendment No. 3 was also the first version to say the LTBT intends to fill the vacant board seat.

What's missing

Where the $8.4 million went. The filing says the funding supports "operations and the compensation of its Trustees" but gives no breakdown. Four sitting Trustees at $400,000 plus the Chair's $25,000 come to about $1.6 million a year (TIA arithmetic). The filing doesn't say what the rest paid for, or why 2026 funding rose so sharply.

An independent check on independence. Independence rests on the Trust Agreement, the LTBT's own policies and the Trustees' self-selection. The filing describes no outside review, and it says company executives have helped choose Trustees.

Who fills the empty seats, and when. One Trustee seat and one Class T board seat are vacant. The filing names no candidates and gives no timetable.

The Trust Agreement itself. The prospectus summarises it. We have not reviewed the full agreement for this piece.

Analysis (our read, not a fact from the filing)

"Independent" here means independent of shareholders, not of the company's money. The trust that picks most of Anthropic's board is funded by Anthropic, is consolidated on Anthropic's balance sheet and has chosen its Trustees with input from company executives. None of that is hidden. All of it is in the filing. Trusts and foundations are often funded by the organisations they oversee, and pay that doesn't depend on profits is a real safeguard. Readers should still know who pays the referee.

For buyers, the practical point is simple. IPO buyers own a slice of Anthropic's economics and roughly 7% of its votes, and they can't elect a board majority or block a deal the founders and trust support. The structure was designed that way, and buyers chose it with the prospectus in front of them. Whether it protects the mission or just protects the people in charge depends on how the Trustees use their power. The filing can't answer that; only their record will.

The strongest counter-argument

That is the point. The structure exists to stop short-term investors overriding safety decisions. Anthropic says the LTBT is meant to help balance shareholders' money interests, "the best interests of those materially affected by our conduct, and the Company Purpose" (p.14).

Real safeguards on pay. Trustee pay is fixed cash, not linked to profits, and Trustees can't hold Anthropic shares (pp.254–255).

Fully disclosed. Funding amounts, Trustee pay, the CNAS donation and how Trustees are chosen are all spelled out in the filing. The company itself describes the structure in its risk factors as potentially eliminating Class A influence (p.18).

Founder control is not new in tech. The filing compares itself with "other corporations with multiple-class structures" (p.111). Founder-led tech companies have long listed with concentrated voting control, and Anthropic has followed that path with an added safety trust.

Limited use of the controlled-company exemption. Anthropic says it will use it only for its nominating committee, and its audit and compensation committees are fully independent (p.266).

Controversy rating: High

This structure takes board control away from public shareholders for as long as it lasts, and the body holding that control is funded by the company it oversees. On the other side, it is fully disclosed and deliberately designed, and we have seen no allegation of wrongdoing by the Trustees or the company over these arrangements. We make none.

Coming tomorrow, Part 3/12: "The 85% margin hiding in Note 8". Anthropic's income statement has no gross profit line, but a note to its accounts lets you work one out. We test the viral "85% gross margin" claim against the filing.

Earlier in this series: Part 1/12, "Anthropic's AI broke into real computers. Then the warning moved to the top of the list": theindependentai.com. Opening-day news: theindependentai.com

Why it mattersIf you hold or are thinking of buying Anthropic shares, this is what the vote attached to them is worth: all public Class A holders together are capped at 49.9%, and a trust the company funds elects most of the board. If you rely on Claude, it also tells you who decides Anthropic's direction: the founders and the trust, not outside investors. This is Part 2/12 of our daily read of Anthropic's prospectus; the Analysis section is our read, everything else is quoted from the filing or is labelled TIA arithmetic.

Sourcessec.gov ↗ · sec.gov ↗ · theindependentai.com ↗ · theindependentai.com ↗

Google Antigravity can now work on remote servers over SSH

Google's Antigravity desktop app, its agent-first coding tool, now supports SSH natively. As well as the existing Remote Control feature, you can connect the app straight to a remote server and have its agents work on the code and files that live there.

Why it mattersMany company projects run on cloud servers or powerful remote machines rather than a laptop. Native SSH means a team can point Antigravity's agents at those machines without copying code around or setting up extra tools.

Sourcesx.com post ↗

Genie 4 world model opens to developers in the Gemini API and AI Studio

Google has made Genie 4, the Google DeepMind model that generates interactive worlds you can walk around in, available through the Gemini API and in Google AI Studio. A day earlier it opened to Google AI Ultra subscribers in Project Genie. Users can step inside a generated environment and change it mid-session, for example switching the weather, adding characters or generating matching audio on the fly.

Why it mattersUntil now you could only try Genie in Google's own app; API access means developers can build it into their own products. Early uses are likely to be game prototypes, training simulations, interactive marketing and education. Google still treats it as a research-stage model, so test quality and costs before promising it to clients.

Sourcesx.com post ↗

Hermes Agent comes to Microsoft Teams, the Microsoft Store and Azure

Nous Research has partnered with Microsoft to launch Hermes Enterprise. Its open-source Hermes Agent now works inside Microsoft Teams, can be installed from the Microsoft Store and runs on Azure. Nous's example: in Teams, employees hand Hermes jobs such as financial analysis, document drafts or inbox triage, then review what it produces. Nous says nothing changes for people using Hermes outside a company setting.

Why it mattersFor businesses already on Microsoft 365, this puts a capable open-source agent where staff already chat, with deployment through Microsoft's own store and cloud, which can make IT approval easier. It is worth comparing with Microsoft's own Copilot agents before choosing.

Sourcesx.com post ↗

Tencent releases open weights for its flagship Hy4 model

Tencent has released the open weights for Hy4, its flagship model, two days after doing the same for the smaller Hy4 flash. Tencent describes it as natively multimodal, with 770 billion parameters in total (49 billion active at a time) and a context window of 1 million tokens. It also published a Hy4 System Card with its own safety, security and alignment test results, including a look at cyber and bio risks, and the model can already be run on Together AI. The figures are Tencent's own.

Why it mattersOpen weights mean a company can run the model on its own servers or a chosen cloud, keeping data in-house instead of sending it to an API. Hy4 needs serious hardware, though, so for most small businesses the practical route is a hosted provider such as Together AI. Check the system card and your own data rules before using a Chinese-built model for sensitive work.

Sourcesx.com post ↗

Vercel launches Board, an experimental message board for people and agents

Vercel launched Board on 19 November 2026 as "an experimental agentic primitive": a shared message board where a team's people and coding agents coordinate work. The official Vercel account posted it at 01:59 Europe/London, calling it "a secure message board for people & agents" and adding: "Board is experimental. Tell us what you think." A changelog entry the same night said it "is now available on all plans."

The docs describe Board as "the communication layer for your team's people and coding agents," with "its own channels and threads, on your Vercel team." A thread keeps "the conversation, context, and updates together," and labels track details such as a pull request or the current status. Once a task is on the board, an agent can claim it, post updates and record the result. Agents read from and write to Board through the Vercel CLI (`vercel board`) or the JavaScript SDK, `@vercel/board`. To teach a coding agent how to use it, the blog and changelog give `npx skills add vercel/board`; the docs add a `--skill board` flag to the same command.

The launch blog says Board "is harness agnostic" and that "Claude Code, Codex, Pi, OpenCode, eve, fx, Mastra, and LangChain can collaborate as if they were a single harness." That list is Vercel's. We have not used the product, so we cannot confirm how those tools work with it in practice.

A task can be posted on Board or synced from a connected tool. The docs say sync "connects a thread with a Slack thread, a GitHub pull request or issue, or a Linear issue, in both directions." The blog says the SDK "handles two-way sync, mirroring replies, reactions, edits, and deletions between Board and your connected tools." The changelog describes this as syncing "both ways with Slack, GitHub, and other tools through Vercel Connect."

Vercel says it first used Board internally for security agents and then for v0's "software factory," which, it says, "now merges over 50% of PRs." That figure is Vercel's claim. We have not verified it.

Grants control who can read, post and manage each channel, and they can be changed while the board is running. Every new board starts with three grants, according to the access docs: every person on the team (`g1`); every project's deployments on the team, "in production, preview, and development" (`g2`); and every app installed on the team (`g3`). `g1` is limited by the person's team role. `g2` can read, post, react, label and sync in every channel. `g3` can read, post, react and label in every channel, limited by the app's scopes. The docs note that "Preview deployments run any branch's code, and `g2` lets them in."

A section headed "What to know before you use Board" is unusually direct. It says: "Everyone on your team can read every channel by default. So can every project's deployments, preview deployments included, and every app installed on your team." It also says: "Agents act with your access. A coding agent you run can read and post wherever you can. A profile changes the name the board shows and never the permissions." And: "Agents read messages as instructions. A task's text tells whichever agent claims it what to do, so review who can post in the channels your agents work from." The same list warns people to keep secrets off the board, because anyone who can read a channel, and any platform that channel syncs with, can read a token posted there. Board also "stores its data in the United States" and "has no region setting, so you can't choose where its data is stored."

The docs do not describe a public or external posting option. They say "Board issues no tokens of its own. Every caller uses a Vercel credential," and the people, projects and apps that can be granted access are those already on the team. Analysis: this is a board for a Vercel team, not a public inbox for outside agents. A synced channel is a different matter: "Board's grants don't apply on those platforms," so anyone in the linked Slack channel, repository or Linear team can read what is posted there.

Board "is free while it's experimental." Vercel "may charge for Board in the future." The docs also say: "Behavior, APIs, and limits may change," and that use of Board is subject to Vercel's Public Beta Agreement.

What we don't know yet

We have not used Board, so we cannot say how the workflow feels in practice, or how complete the harness list is. Vercel has not said when, or whether, it will start charging, or what a paid plan would cost. The official 48-second launch video on Vercel's post was not used as a source for this story.

Vercel is in the directory at /tool/vercel. We recently covered other Vercel tools for agent builders, including Vercel Labs' October releases, Vercel Security Review, the UI SDK and Workflow SDK 5. There is no directory page yet for v0 or for Board itself.

Last checked: 20 November 2026.

Why it mattersBoard is for Vercel teams that want their people and coding agents to share one place for tasks, including teams already working in Slack, GitHub or Linear. It is available on all Vercel plans and free while it stays experimental; Vercel says it may charge later. The docs warn that, by default, everyone on the team, every deployment and every installed app can read every channel, that agents act with your access, and that they read messages as instructions.

Sourcesvercel.com ↗ · vercel.com ↗ · vercel.com ↗ · vercel.com ↗ · x.com post ↗

Gemini on Mac can now click, type and work through your apps for you

Google added computer use to the Gemini app for Mac. You describe a multi-step job, such as pulling figures from PDFs into a shipping form, batch-editing photos or building slides from a dashboard, and Gemini does the clicking and typing in a separate background window. You can watch a live preview in the chat while your own screen and mouse stay free. You choose which apps it may use, it pauses for logins and payments, and you approve key actions. It's rolling out in US English for Google AI Plus, Pro and Ultra subscribers on Apple-silicon Macs.

Why it mattersThis is the kind of agent that handles boring admin across apps that don't talk to each other, like forms, listings and reports. It isn't in the UK yet, but OpenAI, Anthropic and now Google all offer it, so it's worth listing the repetitive desktop jobs in your business that an agent could take over.

Sourcesx.com post ↗ · blog.google ↗ · techcrunch.com ↗

Perplexity's agent now writes and improves its own skills from how you work

Perplexity says Brain, the memory and self-improvement system inside its Computer agent, now writes reusable skills for you automatically. It looks at your past sessions and memory, spots jobs you do again and again, and writes step-by-step instructions the agent loads next time. A new skill needs the same need to show up twice, and every change is checked before it's switched on. It's on by default for eligible US Max subscribers and you can turn it off. Perplexity's own figures: more than 80,000 personal skills written for over 20,000 users between August and mid-November.

Why it mattersSkills are how you stop explaining the same process to an AI every time. Even if you don't use Perplexity, the lesson carries over: write down your repeat jobs (how you format reports, check invoices or answer enquiries) as skills your AI tools can reuse.

Sourcesx.com post ↗ · perplexity.ai ↗

OpenAI's new voice model gets cheaper when nobody's talking

OpenAI released GPT-Live-1.1 in its API. It's the voice model that can listen while it speaks, and OpenAI says it handles accents, pronunciation and transcription better across languages. The bigger change for anyone building voice agents is a new standby mode: when a conversation goes quiet, it pauses and bills $0.02 a minute instead of $0.05, then wakes up as soon as someone speaks.

Why it mattersIf you're planning a voice receptionist or phone assistant, idle time was a hidden cost. Standby cuts the price of silent minutes by 60%, which makes always-on voice helpers cheaper to run.

Sourcesx.com post ↗ · developers.openai.com ↗

Mistral is running a hackathon in London this Saturday

Mistral is holding a special Le Chonk hackathon in London on Saturday 21 November, its first since its worldwide hackathon, with another in Paris with ClickHouse on 5 December. Sign-up is through the Luma page.

Why it mattersA free day building with a European model maker's tools and team is one of the quickest ways to see what's practical, and to meet builders near you.

Sourcesx.com post ↗ · luma.com ↗

How OpenAI uses Codex agents to keep ChatGPT fast

OpenAI described its Performance Factory, where Codex agents handle ChatGPT slowdowns. When an alert fires, an agent investigates, writes up what it found and proposes a fix for an engineer to review. After the fix goes live, another agent checks that speed is back to normal. The team first wrote its investigation know-how into a shared skill called perf-forensics, then turned it into a repeatable workflow. OpenAI says this ran more than 1,700 investigations and found the root cause of over 600 in about three months (its own figures).

Why it mattersIt's a clear model for any team: write down how your best person handles a recurring problem, turn it into a skill, then let an agent do the first pass while a person approves the result.

Sourcesx.com post ↗ · openai.com ↗

NVIDIA's free tools check agent skills before you install them

NVIDIA's open-source SkillEvaluator was trending on GitHub this week. It checks AI agent skills, the instruction packs you add to tools like Claude Code, Codex or Cursor, for quality, security, personal data and whether the agent will actually load them, and it can compare results with and without a skill. Its security check is NVIDIA's SkillSpector scanner. NVIDIA's own research found problems in 26.1% of about 31,000 public skills it analysed, and signs of malicious intent in 5.2%.

Why it mattersSkills run with a lot of trust inside your AI tools. Before your team installs one from the internet, scan it (or ask whoever set up your tools to), just as you'd vet any software you put on a company machine.

Sourcesx.com post ↗ · github.com ↗ · github.com ↗

With GPT-6.2, don't edit the conversation history

OpenAI says agents using GPT-6.2 in its Responses API now build on their earlier reasoning, giving better results with fewer tokens and more cache hits. The catch: that saved reasoning is tied to the exact conversation before it. If your app edits, inserts or trims earlier messages, the reasoning is dropped and the cache stops working. OpenAI's guide says to keep the history append-only and put standing instructions in the first developer message.

Why it mattersIf you or your developer run agents on GPT-6.2, this one habit can cut your bill and make answers better. Ask them to check whether your app rewrites old messages.

Sourcesx.com post ↗ · developers.openai.com ↗

Volcano: Kong's cloud for agents that proposes fixes for its own bugs

Kong, the API company, launched Volcano, a 'self-healing agent cloud'. It hosts agents, apps and data in one place and works from Cursor, Codex, Claude Code or an online builder. Its LavaFlow feature reads your production logs and code, works out what keeps breaking, and opens a draft fix for you to review. 'Self-healing' is Kong's word: fixes still need your approval.

Why it mattersFor small teams building internal agents, one platform for hosting plus suggested fixes could replace several services and some late-night debugging.

Sourcesx.com post ↗ · volcano.dev ↗

MiniMind: train a tiny language model yourself to see how it works

MiniMind is a free, open-source project that walks you through training a 64-million-parameter language model from scratch, from cleaning the data to pretraining and fine-tuning. Fact check on the viral claim that you can 'train your own LLM in about 2 hours': partly fact. The project's own estimate is about 2.3 hours on one RTX 3090 graphics card for its mini version, so you need a good GPU, and the result is a teaching model, nowhere near ChatGPT.

Why it mattersIt's a hands-on way for a technical person on your team to understand what's inside the AI tools you pay for. It won't replace any of them.

Sourcesx.com post ↗ · github.com ↗Partly fact Hype or Fact? check

@thsottiaux: ChatGPT's dots reach some Plus users, and a new mobile app is coming

OpenAI's @thsottiaux says dots, ChatGPT's proactive agent that works for you in the background, is now being tested with a small number of Plus users, and he's seeing more positive feedback since it got better at acting on its own. A redesigned ChatGPT mobile app is starting with Pro users, with its own tab for your dot and the option to pin Codex. OpenAI's help pages say dots isn't available in the UK, EEA or Switzerland yet.

Why it mattersProactive assistants that act without being asked are moving from the top plan down to the standard Plus plan. When it reaches the UK, start it on low-risk jobs.

Sourcesx.com post ↗ · x.com post ↗ · 9to5mac.com ↗

@kimmonismus: Apple is building its own autonomous agent, Bloomberg reports

@kimmonismus flagged a Bloomberg report that Apple is building an AI agent that can browse the web, write code, control apps and move around your devices. The report is based on Apple job listings, describes an early-stage project and says a launch could come as soon as late 2027.

Why it mattersIf Apple builds agents into the iPhone and Mac, a lot of everyday business admin could run through the devices your team already uses. It's still more than a year away, so nothing to act on yet.

Sourcesx.com post ↗ · macrumors.com ↗

@skirano: run your sub-agents in 'dry mode' for speed

@skirano, who builds AI design tools, recommends running sub-agents on GPT-6.2 Astra or Sol in what he calls 'dry mode', with everything layered on top of the model switched off, saying they answer faster and work more efficiently that way. It's his own experience, not a published test.

Why it mattersWhen you chain AI helpers together, extra instructions and tools slow every step. Try stripping a helper back to the basics for simple jobs and see whether quality holds.

Sourcesx.com post ↗

@gregisenberg on Gemini 4 Neon: about Opus 5.5 quality at half the cost

@gregisenberg tested Gemini 4 Neon early and calls it really good, really fast and about half the price of what he usually uses. He points to benchmarks putting it level with Claude Opus 5.5 on coding at around half the cost per task, but says he still often reaches for Claude.

Why it mattersA useful real-world view: a cheaper model can do much of the work, but most builders keep their favourite for the jobs that matter most. Test both on your own tasks.

Sourcesx.com post ↗

How we knowCompany figures are labelled as their own claims; anything we couldn't verify is labelled as a report or a claim.