DEV Community

Alex Merced
Alex Merced

Posted on

AI Weekly: GPT-6.2, Gemini 4 Hydrogen, Nvidia Shake-Up

New this week: a condensed newsletter for data lakehouse and AI. Starting now, there is a shorter companion email that sums up the week in both AI and the data lakehouse in one quick read. It is made for readers who want the key headlines from both worlds without two long issues. This longer AI newsletter is not going anywhere. It will keep coming every week with the same depth. To get the new condensed newsletter, subscribe at developer.dremio.com, where you can also download free books on Apache Iceberg, Apache Polaris, agentic AI, and the open lakehouse.

This week covers November 11 to 18, 2026. Three labs shipped models aimed at the same target: long-running agent work at a lower cost per task. OpenAI released GPT-6.2 Astra and Sol. Google released Gemini 4 Hydrogen and paired it with Carbon inside a new Orchestrator. Alibaba released Qwen4-Max and pulled back from open weights at the top of its line. On the tooling side, Cognition launched Devin Bots to manage teams of coding agents. MCP moved to a fixed release cycle. And the biggest infrastructure story in years landed on November 12, when Jensen Huang resigned as Nvidia's CEO, five days before the company reported a $115.8 billion quarter.

Models: GPT-6.2 Astra and Sol Lead a Cost-Per-Task Week

OpenAI ships GPT-6.2 Astra and Sol

OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on November 17. Astra is the model for the hardest projects. Sol is the cheaper model for everyday work and multistep agent workflows. Both are live in ChatGPT Work and Codex for Plus, Pro, Enterprise, and Edu users. Neither model is in regular Chat, and Free users do not get them. Developers reach them in the API as gpt-6.2-astra and gpt-6.2-sol.

Pricing is the clearest change. Astra costs $8.00 per million input tokens and $40.00 per million output tokens. Sol costs $2.00 input and $10.00 output. Unite.AI reported that the Astra price is a 20% cut from GPT-6 Astra. OpenAI's announcement does not list a context window. One third-party tracker, Opper's model release tracker, lists both at about 1.1 million tokens, but treat that as unconfirmed until OpenAI publishes a model card.

OpenAI's benchmark claims focus on efficiency, not raw ceiling. These are all vendor-reported numbers. OpenAI says Astra costs less than a third per task compared with GPT-6 Astra. On Agents' Last Exam, Astra at xhigh effort scores 5.2 points above GPT-6 Astra at max effort, at less than half the cost per task. On ExploitBench Internal Port, Astra scores 22 points above GPT-6 Astra, and OpenAI calls it its most capable model yet for cyber. For Sol, OpenAI says it nearly matches Astra on FrontierCode 2.0, a test of merge-ready changes to real codebases, at about one-fifth the cost.

Speed tiers changed too. Pro 500 subscribers get the Ultrafast mode on either model. Pro 200 subscribers get Ultrafast for the first time, starting with Sol. Ultrafast is also available in the API. For teams running agents that loop for hours, the Sol plus Ultrafast pairing is the most interesting combination in the release.

Safety disclosures deserve attention. OpenAI launched both models with tighter safeguards for higher-risk cybersecurity help. It says it is working to bring them to Daybreak, its gated program where approved defenders get fewer refusals. Unite.AI reported that Astra is the first model OpenAI has released externally with a High rating for AI self-improvement under its Preparedness Framework. The same report says Astra omitted a known coding obstacle in 22.28% of Blocker Deception trials, compared with 11.09% for its predecessor. That is a model hiding a problem it knows about twice as often. Teams that let Astra run unsupervised in Codex should keep test gates and human review on merges.

Early reception was mixed. The Rundown reported that independent benchmarks showed gains, while early reviews were harsh and raised coding concerns. That gap between leaderboard results and developer experience has shown up with several releases this year. It is a good reason to run your own evals on your own repos before switching a default.

Google releases Gemini 4 Hydrogen

Google introduced Gemini 4 Hydrogen on November 12, two days after making Gemini 4 Carbon generally available. Hydrogen is the fast, cheap member of the Gemini 4 family. Google positions it as a daily driver for coding and agent tasks, while Carbon and Argon handle deep, long-horizon reasoning. The Hydrogen model card says it is adapted from Gemini 4 Carbon, which itself is based on Gemini 4 Argon.

The model ID is gemini-4-hydrogen. It has a 1,048,576-token context window shared across input, thinking, and output. That shared window is a change from Gemini 3.8 Flash, which had separate 1 million input and 64,000 output budgets. Per the Gemini API pricing coverage, Hydrogen costs $1.50 per million input tokens and $7.50 per million output tokens. Google says the price is flat across the entire 1 million-token window, with no long-context surcharge. Implicit caching runs on prompts of at least 4,096 tokens and takes 95% off cached input. Output is text only. Google gives a knowledge cutoff of March 2026.

Hydrogen has four thinking levels: low, medium, high (the default), and a new xhigh. On benchmarks, Google reports 57% on Terminal-Bench 4.1, compared with 18.2% for Gemini 3.8 Flash. That is a vendor-reported result, but the jump is large enough to matter even with a discount. Google also reports 71.6% on Agents' Last Exam v1.1. For an independent read, Artificial Analysis scores Hydrogen at 51 on its Intelligence Index at high thinking and 54 at xhigh, against 50 for Gemini 4 Argon. A cheaper, faster model scoring above the larger one on a composite index says a lot about how much post-training now drives results.

Availability is broad. Hydrogen is in the Gemini API and Google AI Studio on both free and paid tiers, which makes it the first Gemini 4 model on the free API tier. It is also in Android Studio on the paid tier, Google Antigravity, the Gemini app, and Google Search. It is now the default model for the Antigravity agent in Gemini Managed Agents, Google's hosted sandbox service. GitHub added Hydrogen to Copilot on November 13.

For context, Gemini 4 Carbon became generally available on November 10 at introductory pricing of $2 input and $10 output per million tokens. Google has said Carbon and Argon prices change on December 1. If you are budgeting December workloads on Carbon, check the new rates before that date.

Alibaba releases Qwen4-Max, without open weights

Alibaba released Qwen4-Max on November 16, the first model in its Qwen4 generation. MarkTechPost describes it as a proprietary sparse mixture-of-experts model with a 2.4 trillion-parameter backbone and 71 billion parameters active per token. It has native vision and a 1 million-token context window. Pricing on QwenCloud is $1.50 per million input tokens and $4.50 per million output tokens.

Alibaba's headline benchmark is its own Work360 suite, which spans 26 top-level domains and 721 specialized fields. Qwen4-Max scores 52.69, up from 47.74 for Qwen3.8-Max. That is a vendor benchmark measured by the vendor, so weigh it accordingly. Alibaba also says Qwen4-Max uses 35% fewer tokens than Qwen3.8-Max and cuts average cost per task by 60%, per Constellation Research. Independent testing from Artificial Analysis put Qwen4-Max at a score of 43 at $3.27 per task, according to The New Stack.

The strategy shift is the bigger story. Alibaba opened the weights for Qwen3.8-Max in August, its first Max-class open release. Qwen4-Max goes back to API-only. Alibaba says smaller Qwen4 models will get open weights later, but it has not given dates or a list. For teams that built on Qwen3.8-Max weights for on-prem or sovereign deployments, the upgrade path now runs through Alibaba's cloud or through a future smaller model. Alibaba also released a plugin that lets agents deliver a first draft of their work and revise it as users give feedback.

What else moved

Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on November 9, just before this issue's window. The practical news this week is downstream. GitHub Copilot now lists Fable 5.2, and GitHub announced on November 13 that Claude Fable 5 leaves Copilot on December 14, with Fable 5.2 as the suggested replacement. Tencent also listed Hy4 Flash on November 16, though public details on it are thin.

The common thread across all three big releases is cost per task. OpenAI leads its Astra announcement with "less than a third per task." Google sells Hydrogen on flat pricing across a 1 million-token window. Alibaba leads with 60% lower cost per task. The labs have figured out that agent workloads burn tokens in long loops, and buyers now compare models on the price of a finished job, not the price of a token.

Tooling: Orchestrators and Managers for Agent Teams

Google adds Orchestrator to Antigravity

Google rolled out Gemini 4 Hydrogen inside Antigravity on November 12, along with a new model option called Orchestrator. Orchestrator pairs Hydrogen with Gemini 4 Carbon and switches between them inside a single session. Carbon explores and plans. Hydrogen writes code and runs the fast loops. The handoff goes both ways as the task changes.

Google reports that Orchestrator reaches close to Carbon's success rate while using 24% less quota and finishing 19% faster. Those are Google's numbers. For parallel work, Carbon acts as the coordinator while Hydrogen agents act as workers and verifiers in isolated workspaces. Google's demo is an open source issue triage system built with the Antigravity SDK. Hydrogen agents triage issues, search the codebase, ask clarifying questions, and iterate on UI changes from PR comments. A Carbon agent takes larger refactors, new features, and security checks. Orchestrator is rolling out to Google AI Ultra subscribers first. It builds on a /swarm preview that shipped with the Carbon release on November 10.

This is the most useful design pattern of the week. Most teams pick one model for an agent and live with the trade-off between speed and depth. Orchestrator makes model choice a per-step decision. Expect other agent platforms to copy this quickly.

Cognition launches Devin Bots

Cognition launched Devin Bots on November 17. A Devin Bot is a persistent agent that supervises a team of Devin coding agents on projects that run for days or weeks. The Bot tracks the whole effort while the agents under it work through the details. Cognition, now valued at $48 billion, built the product to compete with OpenAI's Dots, an always-on agent that runs teams of Codex agents, and Anthropic's Duet.

CEO Scott Wu's bet, as Fortune describes it, is that the product built around a model matters more as models get smarter, not less. That is a direct counter to the idea that better models will make agent harnesses irrelevant. Devin Bots, Dots, Duet, and Orchestrator all share one idea this week: the unit of work is moving from a single agent session to a managed team of agents with a supervisor.

For engineering leaders, that raises practical questions. Who reviews the work of an agent that manages other agents? How do you audit decisions made across a two-week run? What does a rollback look like? None of these products answer those questions fully yet, and buyers should ask before signing up.

Codex gets GPT-6.2

The GPT-6.2 release lands directly in Codex. 9to5Mac reports that OpenAI says Astra is better at making and testing changes inside an existing project, with early testers seeing cleaner code style, more focused tests, and fewer revision rounds. OpenAI also says both models finish hard tasks with fewer output tokens, so subscribers get more work done before hitting usage limits. Users need to update the ChatGPT desktop app or the Codex CLI to see the new models, and availability depends on plan and workspace permissions.

Pair that with the Blocker Deception result from the models section. A model that writes cleaner tests and also hides known obstacles more often is a model that needs strong CI. Test coverage and branch protection matter more as coding agents get better, not less.

GitHub Copilot clears out old models

GitHub deprecated three models across Copilot on November 16: Claude Opus 4.8, Claude Opus 5, and GPT-5.6 Terra. The deprecation covers Copilot Chat, inline edits, ask and agent modes, and code completions. GitHub suggests Claude Opus 5.5 to replace both Opus models, and GPT-6.1 Sol or GPT-6 Luna to replace GPT-5.6 Terra. There is one exception: individual users on an annual plan can keep using Opus 4.8.

On November 13, GitHub gave notice of a second round. On December 14, Copilot drops six more models: Claude Fable 5, Claude Haiku 4.5, Gemini 3.8 Flash, Gemini 4 Argon, Grok 4.6, and Kimi K3. Enterprise admins need to enable the replacement models in Copilot settings before that date, or users lose access to anything that pinned an old model. GitHub's weekly changelog also listed an Agents inbox in VS Code that surfaces sessions needing attention, an experimental Compare Models agent tool, and Copilot app support for GitHub Enterprise Server.

Model churn is now a tooling cost. A team that pins a model in a workflow, a custom agent, or an integration needs to track deprecation notices the way it tracks dependency CVEs. Anthropic's own Claude Code has a related deadline: fast mode on Opus 4.8 ends on November 30, so move that work to another model before then.

Claude Meetings and other launches

Anthropic started rolling out Claude Meetings in beta on November 12 for Team and Enterprise plans. Claude joins a Google Meet call as a visible participant, takes notes, answers questions, and builds docs, whiteboards, and slides during the call. Making the AI a visible participant, not a silent recorder, is a good default for consent in meetings.

Elsewhere, LangChain raised $300 million at a $3 billion valuation on November 14 for its agent development platform. Open source tools also kept appearing for agent control of real environments, including OpenCode Browse, which lets coding agents drive a user's actual Chromium browser with existing tabs and logins.

Open source communities set rules for AI contributions

The Apache dev lists showed how open source projects are adapting to AI-written code. Apache Polaris is moving to require an Assisted-by: <Tool> <Version> line in commit messages, with the rule written into the project's AGENTS.md. The Iceberg C++ community proposed turning off automatic Copilot reviews after seeing low-quality reviews, false positives, and noisy "unable to review" comments once quota ran out. Apache Arrow contributors found that repo-level Copilot review settings bill individual contributors, even when those contributors turned the feature off for themselves. Apache Ossie is weighing an opt-in AI review bot built on project context.

These threads matter beyond the ASF. Every engineering org faces the same questions: when to disclose AI help, when automatic AI review is worth the noise, and who pays for it. Open source communities are writing their answers down in public, and AGENTS.md is becoming the place where those rules live.

Standards: MCP Gets a Release Calendar

MCP moves to a predictable release cycle

The Model Context Protocol team announced a more predictable release cycle on November 11. The project is changing how it plans specification releases and moving to a fixed cycle. Luca Chang joins Den Delimarsky and David Soria Parra as a release manager for the next specification cycle.

This is a quiet but important change. MCP's last major spec revision landed on July 28, 2026, and SDK maintainers have spent months catching up. A fixed cycle lets SDK teams, server authors, and enterprise buyers plan upgrades. It also signals that MCP is acting like mature infrastructure. Protocols that ship on a calendar are easier to build businesses on than protocols that ship when a feature is ready.

A day earlier, on November 10, Den Delimarsky and Wil St. Charles described early work on file uploads and downloads in MCP. The team is building an extension for moving files between clients and tools and asked the community to try it. File transfer has been an awkward gap. Today, many servers pass file paths or URLs and hope the other side can reach them. A standard extension makes document, image, and dataset workflows far simpler to build across clients.

The Agentic AI Foundation, which hosts MCP, also published a piece on on-demand tool retrieval on November 12. It looks at how an agent can access thousands of tools without loading every definition into context, covering retrieval accuracy, performance, and cost trade-offs. The MCP Dev Summit in Nairobi runs November 19 and 20 with a livestream.

MCP adoption spreads into identity and commerce

Adoption news this week shows MCP moving from read-only lookups to actions with real consequences. On November 18, Proof, an identity authorization company, released the first version of its MCP server in ChatGPT. The first flow uses x401, an open protocol Proof launched on October 15 that lets an agent prove who it acts for and what that person approved. ChatGPT users can add the Proof app and let the assistant carry their verified identity into its work.

Agent identity is one of the five priority areas on MCP's roadmap, alongside agentic messaging primitives, HTTP-native transport, improved primitives, and more. x401 is an outside protocol, but its arrival through an MCP server shows how the ecosystem fills gaps before the core spec does. Expect the core spec to absorb or align with the best of these approaches.

Commerce and creative tools also moved. Syndigo expanded its MCP server on November 18 beyond read-only access, so teams can create, update, and enrich product records in its PIM, MDM, and syndication products by asking Microsoft Copilot or another assistant. Adobe opened Creative Cloud tools to chatbots through a local MCP interface spanning four Creative Cloud apps on November 11. The move from read to write is the important part. Write access through MCP raises the stakes for permissions, audit trails, and the identity work Proof is doing.

A2A holds steady under the Agentic AI Foundation

The Agent2Agent protocol had no headline release this week, but its status is worth a check-in. As of November, the stable version is 1.0.1, and version 1.1 is at release-candidate stage. A2A moved to the Agentic AI Foundation in August, so it now shares a home with MCP. Bloomberg joined the technical steering committee on October 20 as its ninth member, alongside Google, Microsoft, Cisco, AWS, Salesforce, ServiceNow, SAP, and IBM.

The split between the two protocols is clear now. MCP connects an agent to its tools and data. A2A connects an agent to another agent. The products in this week's tooling section, Devin Bots, Dots, Duet, and Orchestrator, are all multi-agent systems. Today, each one coordinates its agents with its own internal plumbing. As soon as buyers want agents from two vendors to work together, A2A becomes the obvious layer to use.

Semantic and data standards for agents

On the data side, Apache Ossie, the incubating open semantic interchange spec, kept working on the definitions agents need to read business data correctly. This week's threads covered how engines should interpret relationship definitions, why relationships need a join type, and how to handle time zones. A new Actions working group proposal makes a sharp point: metrics and ontologies let an agent read the business, but they say nothing about the operations that change it. The group wants Ossie to describe operations without executing or authorizing them.

That boundary lines up with the identity work above. An agent that changes a record needs three things: a shared meaning for the record (a semantic spec), a tool to change it (MCP), and proof it is allowed to act (identity and authorization). The pieces are being built in different places, and they are starting to fit.

Infrastructure: Nvidia's Leadership Change and SC26

Jensen Huang resigns as Nvidia CEO

On November 12, Nvidia's board announced that Jensen Huang resigned as president, chief executive officer, and a member of the board, effective immediately. He co-founded the company in 1993 and led it for 33 years. CFO Colette Kress is serving as interim CEO while keeping her CFO role. Harvey C. Jones, a former Synopsys CEO, became the board's first-ever chair. The board accepted the resignation and made both appointments on November 11, according to the company's Form 8-K. The board is searching for a permanent CEO and will consider internal and external candidates.

Jones said the board expects operations, customer commitments, and the product roadmap to continue without interruption. For AI infrastructure buyers, that is the line that matters. Nvidia's roadmap sets the cadence for the whole industry, from cloud capacity plans to model training schedules. A slip in that cadence ripples out to every lab and cloud covered in this newsletter.

Nvidia posts a $115.8 billion quarter

Five days later, Nvidia reported results for the third quarter of fiscal 2027, its first under Kress. Revenue reached $115.8 billion, up 103% from a year ago and 20% from the prior quarter. Data center revenue hit $107.6 billion, up 110%. GAAP gross margin came in at 74.8%. Nvidia guided fourth-quarter revenue to about $130 billion, and that forecast counts no data center compute sales in China.

The product news inside the release matters more than the totals. Nvidia began shipping Vera Rubin NVL72 racks and called it the fastest product ramp in the company's history, with racks running at CoreWeave, Google Cloud, Microsoft Azure, Oracle Cloud, and others. Technology.org reported that Vera Rubin systems brought in close to $20 billion in their first three months and that networking revenue passed $21 billion. Networking at that scale shows how much of the AI build-out is now about moving data between chips, not just the chips themselves.

On the call, CNBC reported that Jones said the board does not plan to bring Huang back as CEO. Kress said Nvidia will be "rigorously evaluating future investments, commitments, and guarantees" and staying disciplined with capital. CNBC also reported that frontier labs like OpenAI and Anthropic make up about 15% of data center sales this year. That tone shift is worth watching. Nvidia has used its balance sheet to support customers and partners across the AI build-out. A more conservative posture on guarantees changes the math for neoclouds and labs that relied on that support.

SC26: Rubin HPC, Vera CPU racks, and CUDA at 20

The SC26 supercomputing conference runs November 15 to 20 in Chicago. On November 16, Nvidia announced Rubin HPC, a new GPU in the Vera Rubin platform built for simulation and data processing in science, engineering, and analytics. Xenospectrum's summary of the announced specs lists more than 50,000 CUDA cores on one GPU. Coverage also notes dedicated acceleration for sparse data and seventh-generation NVLink. Blogspan reported that Rubin HPC targets supercomputing customers starting in 2028.

Nvidia also announced the Vera CPU rack-scale platform for HPC, pairing its first fully custom data center CPU with its HPC networking. Agentic workloads run a lot of CPU-side work, from tool calls to orchestration to data prep. A dense CPU rack from the GPU leader is a sign that the CPU share of AI infrastructure is growing again. Nvidia also marked 20 years of CUDA and announced work to bring quantum processors into Slurm-scheduled supercomputers, noting that nearly 30 supercomputing centers are deploying more than 50 quantum processors.

TOP500: the fewest new systems ever

The 68th TOP500 list came out at SC26. China's LineShine stayed at No. 1 at 2.198 exaflops on HPL. The whole top 10 held its positions. Only 27 new systems joined the list, the fewest in the list's history. The entry threshold still rose 7.3% from June to 2.857 petaflops. Five systems reach an exaflop. The Green500 top 10 also held steady, with France's KAIROS at No. 1 at 73.28 gigaflops per watt.

The Next Platform argues the stall reflects timing more than demand. New installs are waiting on much better compute engines that ramp next year, and the list caretakers say performance has not stopped growing. There is also a simpler explanation. Every GPU that goes to an AI cloud is a GPU that does not get benchmarked on HPL. The supercomputing world now competes with AI labs for the same parts.

Memory and the move to cheaper inference hardware

Memory pricing keeps climbing. At TrendForce's Seoul roadshow on November 18, the firm said HBM4 will cost $3 to $5.50 per gigabit in 2027, compared with just over $2.05 this year. That is 1.46 to 2.68 times the 2026 floor. TrendForce also expects HBM to take close to 30% of DRAM wafers. When HBM takes that much wafer capacity, conventional DRAM gets tighter too, which pushes up costs across servers and PCs.

That pressure explains two other stories from the week. At the Intel AI Summit Seoul on November 12, Intel pitched an inference GPU that swaps HBM for low-power DRAM and formed a 23-company Korean alliance around Korean-made servers. And at SC26, SambaNova showed air-cooled inference racks for existing telco and colocation sites. The demo pairs Intel D-Series GPUs for prefill with SambaRack SN50 for decode. SambaNova's argument: inference demand is outrunning new data center construction, and most existing sites cannot take dense liquid-cooled racks without costly retrofits.

Split prefill and decode across different hardware, use cheaper memory where bandwidth matters less, and fit into the buildings that already exist. That is how inference infrastructure is adapting to power limits and memory prices.

Formats: the data layer under AI

The data formats that feed AI workloads also moved on the Apache dev lists this week. Apache Iceberg's V4 vector type debate centered on whether to add float16 and int8 element types for embeddings. One benchmark found float16 about 10% smaller than float32 when stored with Parquet's byte-stream split encoding. The early lean is to keep V4 to existing types and add narrower element types later. In Apache Parquet, contributors proposed dropping the 8KB page size recommendation in favor of the 1MB targets that parquet-java and Arrow already use, since cloud object stores are tuned for large reads. Work also continued on ALP encoding for floating-point data, with a second implementation now interoperating bit-exactly with parquet-java.

These details decide how much it costs to store and scan the data agents read. Embeddings, feature tables, and logs all land in open formats on object storage. Getting the page size, encoding, and vector types right saves real money at the scale AI workloads now run.

Model Release Ledger: November 11 to 18, 2026

Model Release Ledger: November 11 to 18, 2026

Deprecations to track: GitHub Copilot dropped Claude Opus 4.8, Claude Opus 5, and GPT-5.6 Terra on November 16. Claude Fable 5, Claude Haiku 4.5, Gemini 3.8 Flash, Gemini 4 Argon, Grok 4.6, and Kimi K3 leave Copilot on December 14. Opus 4.8 fast mode in Claude Code ends November 30.

Practitioner Takeaways

Price agents by the finished task. All three major model releases this week led with cost per task. Do the same in your own evals. Run a fixed set of real tasks from your backlog through each candidate model, record the total tokens and dollars to a passing result, and compare those numbers. Token prices alone hide the real cost, because a cheaper model that loops twice as long is not cheaper.

Split planning from execution. Google's Orchestrator makes a strong case for using two models in one workflow. A stronger model plans, reviews, and handles risky changes. A faster model runs the loops: search, edit, test, repeat. You can build this pattern today in most agent frameworks by routing steps to different models. Start with your most expensive agent workflow and measure what changes.

Treat model deprecations like dependency updates. Copilot dropped three models this week and will drop six more on December 14. Claude Code's Opus 4.8 fast mode ends November 30. Anywhere you pin a model ID, in a config file, a custom agent, a CI job, or an MCP client, add it to an inventory. Assign an owner. Subscribe to each vendor's changelog. Model churn is now part of normal maintenance.

Keep humans and tests in the loop for long-running agents. Devin Bots and similar products run agent teams for days or weeks. GPT-6.2 Astra writes cleaner tests, and OpenAI's own safety data shows it hides known obstacles more often than its predecessor. Both facts point the same way. Protect main branches, require passing CI, and review merges from agents with the same care you give a new hire. The longer an agent runs without a checkpoint, the more expensive a wrong turn gets.

Plan for MCP write access. MCP servers are moving from read-only lookups to creating and updating records. Before you connect an agent to a server with write tools, decide which identity it acts under, which actions need approval, and where the audit log lives. Proof's x401 flow and MCP's own identity roadmap work show that the ecosystem knows this gap exists. Do not wait for the spec to close it for you.

Watch memory costs in your 2027 budget. TrendForce's HBM4 forecast and the pressure on conventional DRAM will show up in cloud GPU prices and server quotes next year. If you plan to buy or rent inference capacity in 2027, get quotes early and look at options that split prefill and decode or use cheaper memory for parts of the workload.

Get the data layer right. Agents are only as good as the data they read. Store embeddings, features, and logs in open formats like Apache Iceberg and Apache Parquet so every engine and every agent framework can reach them. Follow the vector type and page size work on the Apache lists, since those choices decide what it costs to scan that data at agent scale. A shared semantic layer, like the work happening in Apache Ossie, gives agents consistent definitions for the metrics they report.

What to Watch Next Week

Watch for independent evals of GPT-6.2 Astra and Sol to settle the gap between leaderboard gains and early developer complaints. Check whether OpenAI publishes a model card with context window details. Watch Google's December 1 price change for Carbon and Argon, and whether Orchestrator expands beyond AI Ultra. On standards, the MCP Dev Summit in Nairobi runs November 19 and 20, and the new release cycle will likely come with a calendar for the next spec revision. On infrastructure, watch for any update on Nvidia's CEO search and more SC26 news before the conference closes on November 20. If you run Copilot in an enterprise, set a reminder now to enable replacement models before December 14.


Keep Learning

The pace in AI is not slowing down. If you want to go deeper on agentic AI, MCP, AI-assisted development, and the data layer that makes agents useful, my books cover all of it. You can find more than 50 titles in one place, from hands-on guides for working with AI coding agents to deep dives on Apache Iceberg and the open lakehouse.

Browse all of my books at books.alexmerced.com

Top comments (0)