NVIDIA announced on November 12, 2026 that Jensen Huang had resigned as CEO, President and board member, ending a 33-year run at the company he co-founded. Chief Financial Officer Colette Kress took over as interim CEO while the board searches for a permanent successor. The change landed in the same week that Anthropic released Claude Fable 5.2 and Mythos 5.2, Google opened Gemini 4 Carbon to the public and followed with the cheaper Gemini 4 Hydrogen, Adobe put conversational assistants across its Creative Cloud apps, and Anthropic began the roadshow for a proposed IPO. For anyone who builds on AI hardware or works with creative and office software, the week changed both the supplier landscape and the tools on the desk.

NVIDIA's leadership change and why the hardware market cares

The confirmed facts are concise. On November 12, a press release titled "NVIDIA Announces CEO Transition" confirmed Huang's immediate resignation from all roles. Kress, who has served as Executive Vice President and CFO since September 2013, became interim CEO, and the board named Harvey C. Jones as its first-ever Chair. An executive search for a permanent CEO is under way.

Why does this matter beyond NVIDIA's own share price, which slid to third place among the world's most valuable companies after peaking at $5 trillion? NVIDIA supplies the compute that nearly every large AI provider runs on, including Google, Microsoft, Anthropic, Meta and Amazon, with Apple the notable exception among big tech firms. Huang was also the industry's most visible spokesperson, consistently arguing for open-source AI development and against heavy regulation. His absence, whether permanent or a quiet period of several months, leaves a gap in that public advocacy. Open questions now include who will lead NVIDIA, how enterprise supply relationships will hold up during the transition, and how government policy on AI chips will evolve. Kress's interim title and the active search suggest she is a transitional leader rather than the long-term choice, though that is a reading of the announcement, not a stated plan.

Two new frontier models, and a surprising cost gap

Anthropic launched Claude Fable 5.2 and Mythos 5.2 on November 9, 2026, and the published benchmark tables show Fable 5.2 leading nearly every category, including FrontierCode 2.0, FrontierSWE v2, Terminal-Bench-Science 0.1, OSWorld 2.1 and CyScenarioBench. On the Artificial Analysis Intelligence Index, an independent composite score, Fable 5.2 reached 65, five points ahead of the next model. In hands-on use, it appears strong enough to serve as an everyday model for software engineering, design work and video editing.

Google's Gemini 4 Carbon had briefly held the top spot after its preview on October 28, 2026, and its public release this week came just days before Fable 5.2 overtook it. Carbon scores 60 on the same index, second overall and ahead of Claude Opus 5.5 at 59. The timing looks unlucky for Google, but two custom tests suggest the benchmark gap does not tell the whole story.

The first test asked each model to generate a complete browser-based 3D survival game from a single prompt, with movement, enemies, collectibles and scoring. Fable 5.2 produced a playable prototype called "Shardstorm" with crystalline purple terrain and pink spires. Carbon produced a low-poly knight running through a green meadow with textured trees, destructible rocks, yellow collectible cubes and slime enemies. Judged on visuals and fun, Carbon's game came out ahead.

Two workbenches comparing a costly purple crystal game and a cheap green meadow knight game with stopwatches

▲ Cost gap between Fable 5.2 and Gemini 4 Carbon game tests

The second test asked each model to draw a detailed vector portrait in SVG format, a text-based image format, scored by an AI judge. Here the cost difference was stark.

Model AI judge score Time Cost Tokens
Claude Fable 5.2 7.2 19 min 50 sec $4.88 122,248
Gemini 4 Carbon 7.9 11 min 20 sec $0.46 47,897
Gemini 4 Hydrogen 6.9 3 min 56 sec $0.31 42,163

Carbon's 7.9 placed it second on the all-time leaderboard for this task at roughly one-tenth of Fable's cost. These are single runs on narrow tasks, so they should be read as examples rather than general rankings, but they point to a practical rule: for 3D code and vector graphics on a budget, Carbon via API deserves a look before defaulting to Fable 5.2.

Pricing reinforces the point. Fable 5.2 is slightly cheaper than Fable 5.1 but still costs $8 per million input tokens and $40 per million output tokens at maximum reasoning effort; medium reasoning is the sensible setting for most production use. Carbon's introductory API rate is $2 per million input and $10 per million output, scheduled to double to $4 and $20 after the promotional window. In the consumer Gemini app, Carbon is limited to the $200-per-month Google AI Ultra tier, a restriction that seems hard to justify for a model this capable.

Gemini 4 Hydrogen targets the mid-tier

On November 12, Google added Gemini 4 Hydrogen, a cost-optimized model with a 1-million-token context window priced at $1.50 per million input tokens and $7.50 per million output tokens. Google positions it against Claude Sonnet 5.5 and GPT-6.1 Terra rather than against flagships. Hydrogen scores 54 on the Artificial Analysis index, next to GPT-6.1 Sol at 55 and not far from Opus 5.5 at 59. Its average cost per task at the highest reasoning setting is $3.16, compared with $10.04 for Opus 5.5. In the SVG portrait test it scored 6.9, with sub-scores of 7.4 for likeness, 6.4 for facial structure and 6.8 for prompt adherence. Hydrogen will not win frontier benchmarks, but for high-volume agent workloads it looks like a solid budget option, and it would likely have led the field had it shipped three months earlier.

Decision APIs: fast answers without the prose

Two companies released specialized "decision" endpoints this week. These are APIs, meaning programmable interfaces, that return a judgment rather than a paragraph of text.

OpenAI's Decisions API, powered by GPT-6 Luna, works in three modes: Predicate returns a probability between 0 and 1, Choice selects one option from a defined list, and Score assigns an intensity rating such as a 1-to-3 severity scale. A demonstration inspected photos of returned packages and output a probability that the box had dents, tears or cracks. OpenAI says decisions run up to ten times faster than standard responses, and pricing is $0.10 per million uncached input tokens with no charge for output tokens. Suggested uses include sorting email, rating the urgency of support tickets, classifying documents as invoices, receipts or contracts, and letting autonomous agents decide whether to retry a step or alert a human.

Google's Gemini Decisions API, built on a dedicated Gemini Decision 1 model, handles calibrated multimodal evaluations across text, images, video and PDF documents, scoring candidates and rubric evaluations across up to 1,000 choices and returning confidence-ranked results with structured justifications.

For developers, the lesson is simple: when an agent loop only needs a yes-or-no, a category or a score, a decision endpoint cuts both latency and cost compared with a full completion model.

Adobe MAX: creative work moves toward conversation

At Adobe MAX in Florida, Adobe showed conversational AI Assistants in Illustrator, Premiere Pro, InDesign and Photoshop, each aimed at removing repetitive steps.

  • Illustrator: a typed instruction such as "Place grouped objects on artboards" arranges assets across multiple canvases automatically.
  • Premiere Pro: "String out underwater footage into a sequence" assembles a rough-cut timeline of trimmed whale and diver clips.
  • InDesign: attach a CSV data sheet and ask for 50 layout variations, and the assistant mass-produces multi-page magazine spreads.
  • Photoshop: ask how to boost vibrance, and the tool creates an adjustment layer and applies the settings while explaining each step.

Designer's monitor showing dozens of magazine layout variations generated from a CSV spreadsheet

▲ InDesign assistant producing 50 layouts from a CSV file

Firefly Design generates festival posters and social media variants from uploaded brand assets, and Firefly Video lets a user click an object in a clip, such as a sneaker, and type "Make the leather blue." The demonstrations also included a 15-second beverage advertisement generated from a script. Taken together, Adobe's core workflow appears to be shifting from manual tool manipulation toward prompting and automated layout generation.

Sneaks: early research

Adobe's Sneaks session previewed two experimental projects. Project Sight Span reconstructs a 3D scene from a single 2D photo, so a creator can rotate the camera around a sleeping panda and export a new angle. Project Sound Bender offers a 2D coordinate pad where dragging a point morphs a sound effect to match on-screen action. Both are impressive, though perspective shifts from single images are already partly possible in existing image models, and much of this looks like Adobe catching up to AI-first platforms such as Runway and OpenAI rather than leaping past them.

Claude becomes a workspace, and Anthropic heads to Nasdaq

Anthropic shipped a set of workspace tools for Claude: Claude Sheets builds and edits functional spreadsheets with multiple tabs, formulas and styling; Claude Whiteboards provides a canvas for planning with live sticky notes and diagrams; Claude Data connects to proprietary databases for queries and narrative analytics; Claude Meetings joins calls, transcribes notes and updates a shared whiteboard in real time; and Interactive Connectors link Salesforce, Atlassian, NetSuite and Workday directly into Claude so data no longer has to be copied between browser tabs. In one example, a bakery used Whiteboards, Sheets and Meetings together to plan a holiday product launch timeline and sales forecast.

Separately, Anthropic PBC launched the roadshow for a proposed IPO of Class A common stock under the Nasdaq ticker "ANT." The filing covers 525,000,000 shares, of which 501,351,354 come from Anthropic and 23,648,646 from existing stockholders, at a proposed range of $70.00 to $78.00 per share, plus a 123,750,000-share over-allotment option for underwriters Goldman Sachs, Morgan Stanley and J.P. Morgan. The listing date has not been finalized.

Other releases worth noting

  • Businesses in ChatGPT: OpenAI now lets verified businesses claim profiles, edit hours, contact details and catalogs, view Business Hub analytics on mentions and click-through rates, and deploy Business Agents that answer customer follow-ups. It functions much like Google Business Profiles, and claiming a listing early keeps official control over what ChatGPT tells customers.
  • Meta Muse Ember: Meta's smallest, fastest model, priced at $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 200,000 tokens, rising to $0.30 and $1.50 above that. It accepts text, code, image, audio and video and is available in Muse Code and the Meta Model API, with reasoning scores comparable to GPT-4 Luna and Claude Haiku 3.5.
  • MiniMax H3-Image: an open-weight image model that produces 16-megapixel 4K images in about 10 seconds, supports up to 8 reference images, and is available through the MiniMax API, MiniMax Code, Hailuo AI and a Hugging Face space for quick testing.
  • Automattic Spacefast: a hosting platform from the parent of WordPress.com for "vibe-coded" apps, meaning apps written largely by AI coding agents. It provisions hosting, custom domains, SQL databases and deployment pipelines directly from agents in ChatGPT, Claude, Muse, Cursor and others via an MCP server, and runs on infrastructure serving more than 15 billion pageviews a month. It competes with Lovable and Bolt.

What to take away

The hardware story and the software story pull in different directions. NVIDIA's leadership transition introduces uncertainty for every company that depends on its chips, and the choice of a permanent CEO and the direction of government policy on AI chips will shape the market for months. At the same time, the tools available to developers and creators improved sharply in a single week.

Practical steps for readers:

  1. If you rely on NVIDIA hardware, watch the CEO search closely and plan for uncertainty in supply relationships while the transition plays out.
  2. For top-end coding and design, Claude Fable 5.2 leads the benchmarks, but run it at medium reasoning to control cost; for 3D code and vector graphics on a budget, test Gemini 4 Carbon via API before its price doubles.
  3. For high-volume agent work, benchmark Gemini 4 Hydrogen or Meta's Muse Ember against your current mid-tier model.
  4. Use a decision API from OpenAI or Google for classification, routing and scoring inside agent loops.
  5. Creative teams should try Adobe's new assistants, especially CSV-driven layouts in InDesign, and businesses should claim their ChatGPT profile and connect Claude to their enterprise systems.