The Diff

Archives
Log in
Subscribe
November 13, 2026

No. 041: Gemini 4 Hydrogen beats Argon on the index at $1.50 in / $7.50 out - and AA clocks it at $3.16 a task. Jensen Huang resigns as Nvidia CEO, effective immediately.

An evening double-header: Google shipped Gemini 4 Hydrogen - its speed tier now beats Argon on the Artificial Analysis index, at $1.50 in and $7.50 out per million tokens - plus Orchestrator, which swaps models mid-task. And Jensen Huang has resigned as Nvidia’s CEO, effective immediately. Plus the quiet ones nobody talks about.
View in browser: The Diff No. 041
The Diff  No. 041
Friday, 13 November 2026  ·  AI dev news, only what is verified
Google benchmark chart for Gemini 4 Hydrogen
An evening double-header: Google shipped Gemini 4 Hydrogen - its speed tier now beats Argon on the Artificial Analysis index, at $1.50 in and $7.50 out per million tokens - plus Orchestrator, which swaps models mid-task. And Jensen Huang has resigned as Nvidia’s CEO, effective immediately.

Hydrogen is live everywhere Gemini is sold: all AI Studio tiers, the API, the app, Search, and Antigravity, where a new Orchestrator mode hands a single task back and forth between Carbon and Hydrogen instead of guessing once up front. Artificial Analysis already has a page up: 54 at the new xhigh setting - and $3.16 per task, against Fable 5.2’s $26.54.

Nvidia’s board announced Huang’s resignation as President, CEO and director in a four-paragraph statement that names his successor (CFO Colette Kress, interim), its first-ever chair, and not one word about why. The investigation the board announced two days earlier goes unmentioned. Earnings are Tuesday.

The fast model got good, the multi-model split became a product, and the company under everyone’s GPUs changed hands - all after yesterday’s issue went out.

IN THIS ISSUE
🚀 Gemini 4 Hydrogen: Google’s speed tier beats Argon on the index at $1.50 in, $7.50 out - and Orchestrator swaps models mid-task
01
🚀 Gemini 4 Hydrogen: Google’s speed tier beats Argon on the index at $1.50 in, $7.50 out - and Orchestrator swaps models mid-task
Hydrogen scores 51 on Artificial Analysis at its default thinking level, above Argon’s 50, reaching 54 at the new xhigh tier. Orchestrator in Antigravity hands one task between Carbon and Hydrogen as the work changes.
02
🔩 Jensen Huang has resigned as Nvidia’s CEO, effective immediately - the board’s statement says everything except why
CFO Colette Kress is interim CEO and keeps the CFO job; Harvey C. Jones becomes the board’s first-ever chair. The statement never mentions the investigation the board announced two days earlier. Earnings are Tuesday.
🔥 VIRAL
The loud ones on Hacker News. Numbers are a morning snapshot.
Google benchmark chart for Gemini 4 Hydrogen
01  Google blogs, Nov 12 · HN 489 + 347 points at 06:25 IST · AA page live overnight

🚀 Gemini 4 Hydrogen: Google’s speed tier beats Argon on the index at $1.50 in, $7.50 out - and Orchestrator swaps models mid-task

Hydrogen scores 51 on Artificial Analysis at its default thinking level, above Argon’s 50, reaching 54 at the new xhigh tier. Orchestrator in Antigravity hands one task between Carbon and Hydrogen as the work changes.
SCENE

Open Google Antigravity today and the model dropdown has a new entry: Orchestrator. Pick it and your task no longer belongs to one model. Carbon starts in a read-only phase - exploring dependencies, resolving ambiguities, writing a plan with explicit re-planning triggers - then hands execution to Hydrogen to write edits, bump packages and run tests. If Hydrogen hits friction, the task goes back to Carbon, the plan gets repaired, and execution resumes. Google’s pitch is that static routers guess a task’s difficulty from your first prompt and get it wrong as the task evolves; this is the bidirectional version.

THE THING

The numbers, separating Google’s from everyone else’s. Google’s: Hydrogen hits 51 on the Artificial Analysis Intelligence Index at its default high setting (Argon: 50), 54 at xhigh; flat pricing across the whole 1M-token context at $1.50 per million input and $7.50 per million output; against 3.8 Flash it is 2.5x faster with Terminal-Bench 4.1 at 57% versus 18.2% and fabricated claims cut more than 5x; Orchestrator reaches near-Carbon task success at 24% less quota and 19% faster - though Google’s own two posts disagree on whether that was measured on Terminal-Bench 4.0 or 4.1. Independent: Artificial Analysis put its page up overnight and confirms 54 at xhigh (#16 of 78), $1.50/$7.50, a 95% cache discount, and $3.16 per Index task - against $26.54 for Fable 5.2, the report-card story from two days ago. The counter-texture: AA clocked xhigh at 13.5 minutes per task (slower than the 12.8 median) and ranks it #76 of 77 for verbosity, burning 304M output tokens on the Index against a 72.8M median. "Speed and efficiency" is Google’s framing; the verbosity tax is AA’s measurement.

✂️ STEAL IT

Steal the architecture, whatever you build on: plan and verify on the expensive model, execute on the cheap one, and make the hand-off bidirectional - the executor needs a way to send the task back up when it breaks, not just receive it. If you are wiring this yourself, the Antigravity SDK example pairs Hydrogen triage-and-code agents with a Carbon agent reserved for refactors and security checks; the same pattern is one router function away in any harness.

Try today  VERDICT

Hydrogen is on every AI Studio tier, so a free key runs it: take one real bug, run it on Hydrogen at default, and price the tokens against your current model. It fails you if you read the launch claims without the setting labels - 51 is the default high tier, 54 is xhigh, and AA’s verbosity rank means long agentic loops can eat the price advantage you came for.

Google: introducing Hydrogen  ·  Antigravity: Orchestrator  ·  Artificial Analysis: Hydrogen page

Tap one, then press send in the email that opens. Researched replies usually land within about 10 minutes.
More like thisLess like thisDig deeperAlready knew it
02  NVIDIA newsroom, Nov 12 · HN #1 at 3,145 points, 06:25 IST

🔩 Jensen Huang has resigned as Nvidia’s CEO, effective immediately - the board’s statement says everything except why

CFO Colette Kress is interim CEO and keeps the CFO job; Harvey C. Jones becomes the board’s first-ever chair. The statement never mentions the investigation the board announced two days earlier. Earnings are Tuesday.
SCENE

Inside Nvidia, a longtime engineer used the company’s internal suggestion box - before the resignation - to urge the board to weigh employee sentiment. Wired reviewed the correspondence: the post has drawn more than 1,600 comments, most of them arguing for Huang’s return. "If you believe in the mission," one anonymous commenter wrote, "bring back THE boss."

THE THING

What is board-confirmed, verbatim from the Nov 12 statement: "Jensen Huang has resigned as President, Chief Executive Officer, and a member of the Board, effective immediately." Kress - CFO since 2013, nearly 25 years in finance roles at Cisco and Microsoft before that - serves as interim CEO while the board runs a search open to internal and external candidates. Jones, the new chair, says the board expects "operations, customer commitments and product roadmap to continue without interruption"; Kress’s own sentence ends with "customers, partners, developers and shareholders." What is not in the statement: any reason, any mention of the independent directors’ investigation announced Nov 10 (Wachtell, Lipton and former acting SDNY U.S. Attorney Joon H. Kim retained), and any comment on the allegations outlets have reported this week. The split to keep: resignation, interim CEO and chair are board facts; the why remains reporting. The first public checkpoint is Tuesday Nov 17, when Nvidia reports Q3 FY27 results at 1:20pm PT with Kress on the call.

✂️ STEAL IT

Steal the reading habit: treat a transition statement like a changelog. What shipped - roles, dates, a named interim, an earnings date. What is absent - a cause, a timeline, a mention of the probe. Absence in an official document is a fact you can use: it tells you the board has chosen not to connect the two stories, which itself is the story to watch.

Watch  VERDICT

Nothing to do today unless you hold the stock. For builders, the load-bearing claim is Jones’s "roadmap continues without interruption" - CUDA, supply and the GTC cadence now sit with a caretaker and a search committee. The Nov 17 earnings call is the first test; listen for whether any analyst asks about the investigation on the record.

NVIDIA: the statement  ·  WIRED: the employee petition  ·  BBC: who is Colette Kress

Tap one, then press send in the email that opens. Researched replies usually land within about 10 minutes.
More like this Less like thisDig deeperAlready knew it
📊 THE ATTENTION GAP
HN points at 06:25 IST. Attention, not quality.
Jensen Huang resigns
3145 HN
Anthropic IPO
855 HN
Gemini 4 Hydrogen
489 HN
ALSO ON THE RADAR

Claude Code 2.1.323 shipped a new execution mode; the settings.local.json fix is still absent on day 7. The headline addition is Code Mode, an experimental opt-in where Claude writes JavaScript to call its tools. Buried lower: a start-up warning for Bash allow rules with a * in a variable’s value - the closest nod yet to the permissions surface behind the exposed-settings-files story. Secret scanning or a settings.local.json change: not in the changelog, a week after the HN comments attributed to Claude Code’s head promised prototypes. Claude Code changelog →

The first independent scorecard on Gemini Decision 1: 66, ninth of 18. Artificial Analysis’s Decision Index v2.1 ranks Google’s decision model just above the field’s median of 64 - first on the knowledge and reasoning components (15% of the score each), weakest on the 30% component, per OrcaRouter’s analysis. Useful API edges the launch coverage missed: noul probabilities are bounded 0.01-0.99 (a threshold above 0.99 never fires), and choice keys are sorted alphabetically in UTF-8 byte order before scoring - name your catch-all other, not none_of_the_above. OrcaRouter on the AA Decision Index →

Anthropic’s IPO portal is live, with five SEC free-writing prospectuses filed yesterday. ipo.anthropic.com carries the founders’ letter and an 11-chapter roadshow video. Yahoo Finance reports pricing is set for Wednesday Nov 18 with trading Thursday Nov 19 under ANT, attributed to a company announcement - as of this morning no matching post exists on anthropic.com/news, so treat the dates as reported, not confirmed at the primary. Anthropic IPO portal →

🧪 ONE THING TO TRY TONIGHT
Ten minutes: one bug, two models, one hand-off.
Open Antigravity (or AI Studio, where Hydrogen is on every tier) and give Orchestrator one real bug from your tracker - not a toy prompt. Watch for the Carbon-to-Hydrogen hand-off and, more importantly, the hand-back: does the fast model know when it is stuck? Then check the token math against what that bug cost you last time. If the answer is "it never handed back and shipped something broken," you have learned the same thing a static router would have taught you, cheaper.
What this issue could not see. Google’s Hydrogen claims - index scores at each thinking level, the 2.5x speed figure, Terminal-Bench numbers, quota and latency savings, partner speedups - are Google’s, from its two launch posts; those posts disagree with each other on whether the Orchestrator figures come from Terminal-Bench 4.0 or 4.1, and we could not resolve which. Artificial Analysis’s Hydrogen numbers are AA’s own measurements of the xhigh variant; the default-high variant may measure differently. Nvidia’s transition facts are from the board’s statement; the employee petition is Wired’s reporting from internal correspondence; the allegations about Huang remain outlet reporting, unaddressed by the board - our No. 037 and No. 040 treatment of that split stands. The Anthropic IPO dates are Yahoo Finance’s reporting of a company announcement we could not find at anthropic.com/news. VIDEO OF THE DAY is omitted for the eighth day running: Anthropic’s roadshow video exists but only as an embed without a linkable URL, and no fresh, relevant and verifiable video on today’s stories could be confirmed - we omit rather than pad. Reddit remains absent by policy; X blocks anonymous access and was not consulted. Deduped against every published issue, No. 001-040, via the kit shipped-index and the live home archive: Hydrogen, Orchestrator, the Nvidia transition, Claude Code 2.1.323, the Decision Index result and the IPO portal are new; the Gemini Decisions API was No. 040 and is referenced, not re-reported.
Talk to the editor 💬

Reply to this email with anything: what to chase, what to drop, what you want more of. Every reply shapes the next issue.

More of this kind of issueLess of this
Images: credited in each report; source screenshots and announcement artwork, not our tests.
Don't miss what's next. Subscribe to The Diff:
← Newer No. 042: Anthropic went down twice in a day - and some scheduled runs never happened. Perplexity publishes how it trains Decider v1.3, its decision model. Older → No. 040: Google's Gemini Decision 1 answers 1,000 questions in one pass - no text, output free. GitLab ships the settings.local.json block Anthropic hasn't.
Powered by Buttondown, the easiest way to start and grow your newsletter.