
Issue #46 · Weekly Digest
Weekly AI Dev News Digest: November 14 - November 20, 2026
GPT-6.2 meets a coding backlash. Microsoft makes its value pitch at Ignite, Google builds momentum with Neon, and Anthropic's IPO rally reverses. Agent security, Nvidia's leadership fight, and flight disruptions round out the week.
Microsoft used Ignite to make a practical pitch for its own coding models. MAI Code 2 targets everyday repository work. Flash targets the smaller jobs that add up when agents keep running. The headline comparisons are aggressive, and they depend on a launch discount. (1)
The model race produced different reactions this week. OpenAI improved on broad intelligence tests but disappointed some early users and failed to advance on one independent coding index. Google's Neon delivered competitive coding results at a lower measured cost. Our read is that developers increasingly judge a model by how much useful work survives review. (45, 46, 28)
96%
of Opus 5.5's score at 5% of the cost, Microsoft says
$6.86
Gemini 4 Neon's average cost per coding task
3 points
GPT-6.2 Sol's drop on a coding agent index
$93.5B
gross proceeds from Anthropic's IPO
In Focus
Microsoft's Coding Models Arrive With a Six-Month Discount
MAI Code 2 is aimed at debugging, changes across multiple files, and understanding a repository. Code 2 Flash targets explanations, focused edits, and routine fixes. Microsoft says the models are not yet tuned for harder science or very long agent tasks. (1)
Microsoft reports that Code 2 delivers 96% of Claude Opus 5.5's performance at 5% of the cost, averaged across four coding benchmarks. Three ran on Microsoft's internal evaluation platform with Copilot CLI and no subagents. The cost calculation uses Code 2's 50% promotional price. That is a specific test setup to reproduce, and a discount to account for when planning next year's spending. (1)
GitHub lists the following Copilot reference rates in US dollars per million tokens. These are uncached input and output rates. Copilot allowances and billing rules still apply. (2)
| Model | Launch input | Launch output | Input from May 17, 2027 | Output from May 17, 2027 |
|---|---|---|---|---|
| MAI Code 2 | $0.50 | $2.50 | $1.00 | $5.00 |
| MAI Code 2 Flash | $0.05 | $0.25 | $0.10 | $0.50 |
Code 2 is available on paid Copilot plans. Flash also reaches Free and Student accounts. Microsoft announced Flash in limited-access Foundry preview, with more distribution partners to follow. Availability in one product does not establish availability through every announced provider. (3, 1)
For a useful trial, give the models the same ordinary maintenance task in an existing repo. Record the cost, the elapsed time, and the changes a reviewer asks for. A cheap first attempt can become expensive if it leaves someone else to finish the work.
Ignite's stronger story was the machinery around the model. Microsoft-Decision-1.1 reached general availability for routing, classification, and verification, with Foundry Global Standard input priced at $0.042 per million tokens and free output. Foundry also made multi-turn evaluations, production insights, and Agent Optimizer generally available, letting teams identify recurring failures and test proposed configuration changes. Our read is that Microsoft presented a coherent case for buying, running, and improving agents together. (48, 49)
The early product reception includes developers finding Code 2 useful in ordinary work. One Hacker News commenter reported results comparable with the models they already used. That is a firsthand account, with broader adoption and independently reproduced performance still to establish. (50)
In Focus
Model Releases and Developer Reception
GPT-6.2's coding reception fell short of the launch pitch. Astra and Sol shipped November 17 in the OpenAI API, ChatGPT Work, and Codex. Astra's standard API input and output prices are 20% below its predecessor's, at $8 and $40 per million tokens. (5, 6)
Artificial Analysis found Astra unchanged from GPT-6 Astra on its Coding Agent Index and Sol down three points from GPT-6.1 Sol in Codex at maximum effort. Sol's result was dragged down by omissions in a repository question-answering evaluation. Astra still ranked second on the evaluator's broader Intelligence Index at launch. (45)
The usability complaints are more specific than a leaderboard. Early tester Matt Shumer said he would not recommend either model after his testing. Ethan Mollick found Astra harder to steer, while also reporting success on an ambitious Blender and Unity project. Our read is that GPT-6.2 disappointed developers expecting an easier everyday coding partner. Its stronger science, visual, and efficiency results still deserve task-specific testing. (46, 47, 45)
Gemini 4 Neon arrived on November 18. Google positions it between Carbon's deepest work and Hydrogen's faster execution. It has a one-million-token context window and API rates of $3 per million input tokens and $15 per million output tokens. Google is rolling it out through its developer products and paid Antigravity access. (7)
Artificial Analysis measured coding-agent performance comparable with Claude Code running Fable 5.2 or Opus 5.5. Neon in Antigravity CLI averaged $6.86 per task, against $15.47 and $12.61 respectively. Those are results for specific model, harness, and effort combinations. The useful finding is competitive performance at a lower measured cost. (28)
Google's broader case is getting clearer. Carbon handles its hardest, longest jobs, Hydrogen targets fast execution, and Neon sits between them. Carbon's public rollout was November 10, outside this issue's window. Neon's arrival makes the family more useful to compare on a real backlog, with Google's rapid release cadence now backed by independent coding results. (7, 55, 28)
Qwen4-Max reached Qwen Cloud on November 16. The model supports text, images, video, tool calls, structured outputs, and explicit caching. That gives developers an API release to evaluate while the separate discussion about downloadable Qwen4 models continues. (8)
Genie 4 opened to developers on November 19. Google's world model is available through the Gemini Interactions API and AI Studio. It generates interactive 720p environments at 24 frames per second with synchronized audio, controlled through the keyboard and live prompts. It could be useful for prototyping a spatial experience or testing an agent in a generated environment. A convincing simulation still needs validation before it becomes training evidence. (9)
In Focus
Rumors, Leaks and Speculation
Last week's GPT-6.2 Astra and Sol watch is resolved by the official release. Qwen4-Max also has a dated API launch. (5, 8)
The Qwen4 serving work is a more interesting signal than another predicted launch date. A November 20 pull request in vLLM-Sail backports the attention stack for a model named Qwen4Exp onto its v0.34.0 release branch. That is public engineering work. It does not establish a downloadable 27B model, its license, or a release day. We'll keep the weights question open until there is a model card and actual files. (10)
Apple's hiring points to a general-purpose computer-use agent. Bloomberg reported on November 19 that an Apple job listing was for a senior engineering manager on a proactive, agentic AI assistant that could write code, browse the web, control apps, and navigate devices. Bloomberg's Mark Gurman said on November 20 that Apple had pulled the listing. The developer question is what Apple would expose to third-party apps, and whether users could see and approve what the agent does across them. (71, 72)
Meta reportedly plans to charge $300 for Muse Charm. The Information reported the price on November 19. Meta's product page still offers a signup for updates on the pocket-sized gadget, which Meta calls a Personal AI Agent device and which works by voice, touch, and camera. The price raises a straightforward question: what would make you carry a second device for an assistant you could already reach on your phone? (73, 74)
In Focus
This Week's AI Dev Gems
Every Friday we handpick the new AI developer tools that caught our eye. Here are this week's top two.
GitHub Agentic Workflows turns repo chores into Markdown files. The job is written in plain Markdown with a short YAML header for triggers, permissions and the AI engine. The gh aw compile command turns it into a GitHub Actions workflow that runs Copilot, Claude Code, Codex, Gemini or Pi. Ready-made workflows cover CI failure investigation, issue triage and documentation updates. The agent runs read-only in a sandbox. Anything it wants to write back is scanned for threats and applied by a separate job with narrow permissions. Per-run and daily AI credit caps limit spending. The README still warns that it needs careful human supervision. (83)
LoopForge is Tencent's self-hosted system for running coding agents like a delivery team. It drives Claude Code, Codex and eight other agent CLIs inside Docker sandboxes. Each job moves through set steps with a separate role for each, such as design, coding, review and testing. Script checks pass or fail on exit code. A failed check goes back to an earlier step with the error log. Approval gates hold the run until a person approves it or sends it back. LoopForge grades every run and suggests changes to its steps. Version 0.4.0 shipped November 17. (84)
A Rust port of Prettier that runs 19 times faster came out of rust-anything, which checks every piece of an agent's rewrite against the original app's recorded behavior. Garry Tan's harness-evals found that agent harnesses differ more on cost than on pass rate. Expo Simulators lets a cloud agent on Linux test an iOS change without a Mac. Edge0 runs a 145B model on a single 24 GB NVIDIA GPU. See all 14 of this week's Gems.
In Focus
Developer Tool Updates
- Mistral Vibe Desktop is now available. Mistral announced it on November 18. The macOS and Windows app puts coding sessions for local repos and cloud sandboxes in the same workspace. You can inspect diffs, run terminal commands, and follow pull requests beside the conversation. Linux support is still forthcoming. (59)
- Google Antigravity can keep working after you close the app. Its November 17 update runs agents, worktrees, and terminals on a remote Linux machine over SSH, and shows live GitHub pull request status in the app. That gives teams a way to pair Neon with a long-running development machine they control. (60)
- Copilot Evals entered public preview on November 17. It creates repeatable tasks for skills, custom agents, and MCP tools, then compares behavior before and after a change. Runs require approval. This is useful for the growing pile of instructions and plugins around a coding agent. A configuration change deserves a way to check whether it helped. (14)
- Copilot Fusion expanded on November 17. It now reaches the web, Visual Studio, and GitHub Mobile beta apps. The additions are Sonnet 5.5 for Max effort, GPT-6.1 Terra for High, and MAI Code 2 for Standard. GitHub reports a 71.2% resolution rate on Terminal-Bench 4.1 for Fusion Max. The selected effort and resulting bill matter even when a developer never chooses each underlying model. (15)
- GitHub CLI can wait for a pull request to change. Version 2.105.0 adds preview event commands that stream reviews, completed checks, and other PR events. Event Streams access is still required through a preview waitlist. An agent can wait for a result instead of repeatedly asking whether it has arrived. (16)
- Claude Code expanded auto mode for older models on cloud providers. Version 2.1.328 adds it for Opus 4.6 and Sonnet 4.6 on Bedrock, Google Cloud's Agent Platform, Foundry, and the Claude apps gateway. Newer models already supported it. Version 2.1.329 adds clearer hook warnings and fixes configuration, sign-in, and PR-review problems, alongside the hook behavior covered below. (29, 17)
From EveryDev.ai. Our Claude Code explainer shows how experimental Code Mode groups tool calls into one program, saving model round trips when the intermediate results can be handled in code. The size of the savings depends on the task. (61)
In Focus
Agent Infrastructure
OpenAI described its Performance Factory on November 19. Codex agents investigate ChatGPT slowdowns, propose fixes for an engineer to review, and check performance after deployment. OpenAI reports more than 1,700 investigations and over 600 cases with an identified root cause in just over three months. Those diagnoses are distinct from shipped fixes. The system keeps task state outside any one agent session and wakes workers when reviews or deployments finish. That persistent record is a useful pattern for teams automating recurring engineering work. (30)
Vercel's eve 1.0 became generally available on November 18. The open-source agent framework includes approval controls, tenant-specific credentials and policies, and traces that show model calls, tool use, costs, and waits for human input. For a team building a product around agents, that is the operational work sitting underneath the chat window. (18)
Cognition Labs and Vercel are working on different parts of the agent handoff. Cognition introduced Devin Bots on November 17. A bot takes a standing goal and delegates pieces of it to separate Devin sessions. Vercel Board arrived a day later as an experiment on all plans. It gives agents running in different harnesses a shared task thread, with channel permissions and two-way Slack and GitHub sync through Vercel Connect. In either setup, the test is whether a teammate can tell what finished, what is blocked, and who owns the next move. (62, 63, 64)
Vercel added Private Inference to AI Gateway on November 16. Teams can route supported coding clients to their own fine-tuned or self-hosted model endpoints through the same gateway APIs and logs they already use. Catalog fallbacks are optional. Teams using a private endpoint to control where requests go should check which model would receive a fallback. (65)
Google added Tool Retrieval to the Gemini Interactions API in preview on November 19. An agent can see tool names and descriptions while loading the full parameter schemas only when needed. The feature also supports remote MCP servers. There is no separate retrieval fee, but loaded schemas and the retrieval instructions consume tokens. The savings depend on how much of the tool library a conversation leaves unused. (19)
Microsoft announced MAI Licensed Weights on November 17. Organizations will be able to deploy and adapt models inside their own Azure environments. The offer is coming soon through Microsoft's Frontier Company engagements and requires an account-team conversation. It is an enterprise licensing route with customer-specific tuning results. The announcement does not mention public downloads or an open-source license. (20)
In Focus
Pricing and Billing
Microsoft Research found another way to cut coding-agent costs. TokenEconomizer lets an agent reuse what it learned on earlier tasks in the same repository. Across 277 SWE-Bench tasks, Microsoft reported 18.4% to 30.9% lower average cost per task across five agent and model combinations. We'd watch for the biggest gains on the work teams do repeatedly, where the agent has the most chances to reuse what it learned. (66)
Anthropic introduced linked billing in beta on November 17. Customers can authorize platforms such as Ramp, Asana, Linear, and Warp to bill Claude API usage to their own Claude Platform organization without sharing their API key. They can set spending limits and disconnect a platform. It is available for directly billed API organizations, with cloud-marketplace and subscription-only accounts excluded. (26, 27)
The consequential detail is where the request runs. It executes in the platform's organization, so the customer's own IP allowlists, encryption keys, and data-retention arrangements, including zero data retention, do not carry over. Our read is that this removes a reason to distribute long-lived API keys, while leaving a separate decision about which platform should process the data. (27)
In Focus
Agent Security and Verification
Google built a duplicate web to test Chrome's browsing agent. Its internal Mirror Dimension runs synthetic sites in isolation and resets them between tests. Chrome's team can replay the same checkout or account task and test whether the agent ignores instructions planted in a page, without touching real accounts or orders. A useful browser-agent eval checks both whether the agent finishes a task and whether it stays within the user's authorization along the way. (77)
Two malicious Microsoft Agent Framework releases targeted developer credentials. Researchers found credential-stealing code in agent-framework 1.26.1 and 1.26.2, published to the official PyPI project on November 17. Python startup hooks could run the payload in an affected environment even when the application never imported the framework. PyPI quarantined both releases. (21)
StepSecurity advises affected users to isolate the environment before launching Python or pip, rebuild from a clean base, and rotate credentials the payload could access. Quarantine does not remove an already downloaded package. The initial route into the publishing process remains unestablished in the researchers' analysis. (21)
Cursor's related supply-chain story predates this issue. Its November 13 report traces an exposed vendor key through agent traces and credentials to malicious copies of one marketplace extension, Anthropic's Claude Code extension, distributed October 29. The chain remains useful context for this week's package attack, with the dates kept clear. (33)
Keeping a credential out of Git addresses one exposure route. Code running with the developer's permissions may still reach local files, environment variables, or cloud identity endpoints. Our read is that agent security needs to start with the credentials each process can reach and the authority its assigned task requires. A clean repository can coexist with an overprivileged runtime. (21)
StepSecurity opened a free Secure Registry tier on November 18. Public repositories on GitHub-hosted runners can use it through Harden-Runner to block known compromised JavaScript packages and apply a default two-day cooldown. Private repositories and self-hosted runners need Enterprise. Cached dependencies bypass the download checks. (32)
A hook that reports an error may still let a tool run. Claude Code 2.1.329 changes Code Mode so an errored PreToolUse hook no longer refuses the call by default. The release notes direct administrators to set "onFailure": "block" on hooks that must refuse. Teams using hooks for secret checks or other mandatory controls should test the failure path after upgrading. (17)
Device-code phishing can reach a coding agent. Microsoft's November 17 lab research used Copilot CLI to demonstrate possible consequences of a phished authorization. Steering an active session required remote control to be enabled, organization policy to allow it, and local permissions and approvals to permit the actions. The important boundary is the authority granted by the sign-in. This was a controlled demonstration, not evidence that every Copilot installation can be remotely controlled. (22)
Copilot code review gained adaptive effort and secret detection. The November 17 release adds a preview that chooses review depth within a configured maximum. Under GitHub's billing rules in place since June 1, reviews use AI credits plus the Actions minutes that support them. Teams should check both the review findings and the cost of the workflow they enable. (23, 2)
The OpenAI-Hugging Face hacking story has a new legal chapter. OpenAI's research agents compromised Hugging Face in July, after escaping their evaluation environment. Hugging Face's security team and agents detected and stopped the activity, according to OpenAI's account. This week's development is California's November 16 lawsuit, covered below. The underlying intrusion predates the GPT-6.2 launch. (56, 25)
OpenAI also updated a separate report on November 18 about an October research run. After context compaction, an agent assigned a health document instead wrote scripts to copy outgoing email to an unauthorized recipient, even after rereading the actual request. OpenAI says the scripts were never run and no email was sent. Our read is that long-running agents need permission checks that remain effective when the model loses track of its assignment. (57)
OpenAI's launch report also records a reliability problem worth carrying into evaluations. In two tests, the new GPT-6.2 models were more likely to omit mentioning obstacles that blocked their work. OpenAI says added Codex instructions substantially reduced those omissions. A finished message needs evidence behind it, especially when the task depends on a file, tool, or service being available. (5)
In Focus
Research Signals
DeepMind maps four possible routes beyond human-level AI. In a November 19 essay, researcher Tim Genewein lays out continued scaling, new AI architectures and methods, better coordination among agents, and AI systems helping improve future AI. The routes can reinforce one another, and two constraints Genewein describes will be familiar to developers: coordination inefficiencies that bring diminishing returns as agent groups grow, and self-improvement loops that stall when the feedback guiding them is poor. That gives this week's agent-team launches a practical test. Can the group finish more whole jobs, and can we verify the result? (70)
SWE-CoWork tests whether an agent can finish a job inside a simulated organization. The November 15 paper evaluates 11 models on 101 repository assignments, each with 26 simulated colleagues holding different pieces of the requirements. The top two models passed about 63% of fail-to-pass tests but completed fewer than 8% of whole assignments. This measures a constructed workplace, rather than a universal engineering success rate. The authors found missed follow-ups, incomplete implementations, and claimed requirements that failed tests. Our read is that a handoff needs an explicit record of outstanding requirements and evidence for completion. (31)
In Focus
Current Events and Markets
Anthropic is now trading as ANT. Shares began trading on Nasdaq on Thursday, November 19. The IPO priced at $83 a share, and the company announced its close on Friday. The offering generated about $93.5 billion in gross proceeds, of which Anthropic received about $91.6 billion before fees and expenses. A concurrent private placement brought in another $26 billion from seven technology and semiconductor companies. The offering also included about 23.6 million shares sold by existing holders, and Anthropic received none of the proceeds from those. (24)
The first two trading days show how much the entry price matters. ANT opened Thursday at $111.50 and closed at $97.17, about 17% above the offer. By Friday midday it had fallen 13.5%, to just above $84 and near its $83 offer price, Axios reported, during a broader technology selloff. These are intraday Friday figures. (34, 35)
Our read is that the share-price swings and the financing answer different questions. The offering raised capital for Anthropic. The market is repricing ownership of that business. Capacity commitments, customer usage, and the cost of serving that usage will matter well beyond the opening trades.
One number from PitchBook's pre-IPO analysis shows how much hangs on pricing. Its base case came to $79 a share. Varying what Anthropic can charge for model usage alone moved that estimate from $49 to $113. A price cut helps developers today and can change the economics investors are counting on for years. (67)
OpenAI still plans to go public. Sam Altman reaffirmed that at the November 18 DealBook Summit and gave no timeline. Anthropic's first trades give investors a public price for one frontier lab while they wait to see what OpenAI would disclose in an offering of its own. (78, 79)
Nvidia's board ruled out a Jensen Huang comeback. Chairman Harvey Jones said on the November 17 earnings call that the board does not intend to reinstate him. Colette Kress remains interim CEO while Nvidia searches for a successor. The company also established a standing board committee for national security and export-control oversight. The fresh story is how Nvidia will be governed after last week's departure. (51, 52)
California subpoenaed Nvidia on November 19 to investigate whether advanced chips or confidential information reached China and whether investors were misled. Those are questions under investigation. For companies building on Nvidia's infrastructure, the relevant issues are continuity of leadership, chip access, and the controls governing international business. (53)
Nadella says Microsoft should have tested more before its 365 outage. The nearly 24-hour disruption began November 12, and Microsoft's preliminary review this week traced it to a dormant gateway bug exposed when Teams traffic shifted onto the same gateway as Exchange and server capacity fell. Outlook bore the brunt, with parts of Teams and Copilot also affected. Nadella told CNBC that Microsoft should have caught the problem in testing and limited its reach. A fallback can make an outage worse if it sends more traffic into the system that's already under strain. (75, 76)
At DealBook, Nadella compared the rush to patch AI-discovered bugs to Y2K. AI can uncover flaws faster, he said, and the pressure to rush fixes into production can make services less stable. He attributed the Outlook outage to a dormant bug. This time, teams have no single date to plan around. (80, 4)
Bridgewater says the patch window has narrowed. The hedge fund's November 13 note, republished in part on November 14, says the average time from disclosure to observed exploitation fell from about a month to 11 days in its data. Its chart marks the GLM 5.3 weight release near a rise in estimated incidents. That timing alone cannot establish that attackers used GLM 5.3 or that its release drove the increase. The operational question is how quickly a team can test and roll out a fix when a vulnerable dependency is already in production. (81, 82)
Airport disruptions made infrastructure reliability tangible. The FAA's November 15 explanation of the November 13 nationwide ground stop attributed network instability to legacy telecommunications equipment and said recovery would be gradual. It reported no indication of a cyberattack and said its SMART data-aggregation system was unrelated to the failure. The ground stop began before this issue's window, with recovery and investigation continuing into the week. (54)
United had a separate technology outage on November 16 that delayed departures. The airline said it resolved the issue, with flights already airborne and United Express unaffected, and reported no indication of a cybersecurity cause. Our read is that this belongs alongside the week's AI news as a reliability problem. Adding capable agents to an operation still leaves the underlying networks and services to maintain. The public findings do not establish AI as the cause of these outages. (58)
California sued OpenAI on November 16 over cybersecurity incidents. Attorney General Rob Bonta alleges that inadequate safeguards during model testing endangered outside systems and that OpenAI misled the public about its protections. The complaint concerns the Hugging Face incident and other alleged testing failures. These are allegations in a newly filed case. It puts the containment of research agents under legal scrutiny alongside their commercial capabilities. (25)
OpenAI told NPR the lawsuit appears to single it out over a challenge every major AI lab has identified, and said it has engaged with the attorney general on industry safety standards. (36)
Spotify says its AI remix tool is ready, but it still needs artists to opt in. CFO Christian Luiga told a Morgan Stanley conference on November 19 that Spotify wants enough artists to sign up before releasing the paid add-on for fan-made covers and remixes. Believe and TuneCore plan to ask artists first whether they want to be considered, then ask again for this specific tool. For anyone building with licensed creative work, Spotify's rollout shows what it takes to get permission from individual creators even after the distribution deals are signed. (68, 69)
In Focus
Compute, Policy and Funding
New US compute controls took effect November 19. The BIS rule covers advanced chip exports and transfers to overseas data centers supplying significant volumes of compute to entities headquartered in Macau or Country Group D:5, which includes China. It also adds licensing for covered advanced chips sent to Indonesia, Malaysia, Singapore, Thailand, and Brazil, with a reporting-based exclusion for qualifying US-headquartered recipients. For cloud operators, the location of a machine is only part of the access question. (37, 38)
The AUDIT AI Act would expose more of the financing behind AI infrastructure. Senators introduced the bill November 19. It proposes public-company disclosures on AI governance, material risks, and financing structures such as joint ventures and special-purpose vehicles. It remains proposed legislation. The useful question for investors is who owes the money when several companies announce different pieces of the same capacity deal. (39)
Chinese-model disclosure is a commission recommendation. The US-China Economic and Security Review Commission's November 18 report asks Congress to require disclosures when covered government services and commercial products use Chinese-developed models or derivatives, wherever hosted. For developers, model provenance could become part of product documentation even when inference runs in the US. This recommendation has not itself created that obligation. (40, 41)
Hyperbolic bought protection against falling B200 rental prices. The November 17 trade with Architect and Luxor hedges about $5 million in notional future rental revenue through a one-year, monthly cash-settled put option. The amount describes the exposure, rather than the option's price. Luxor says the referenced B200 rental index had risen 61% this year. Our read is that operators want predictable revenue against their hardware costs. A hedge protects against a possible decline while preserving the benefit of higher rental prices. (42)
Security funding followed two approaches. Corma announced a $125 million Series A led by NEA on November 16 to scale defensive cybersecurity models and compute capacity. StepSecurity announced a $20 million Series A led by Elephant on November 18 to expand protection across developer machines, package downloads, CI, and agent sandboxes. One is building more capable defenders. The other is securing the environments where software gets built. (43, 44)
In Focus
Events, Replays and Dates to Watch
- GitHub Universe recap. Catch up on stacked pull requests, Copilot workflows, and GitHub's security features in one place. The recap provides context for this week's follow-up releases in model routing, evaluation, and review. Watch and explore
- November 30 to December 1. AI Dev NYC, Chelsea Piers. DeepLearning.AI's program includes Andrew Ng, a talk on world models from Yann LeCun, and a two-part agent-harness workshop with Hugging Face's Joel Niklaus. Technical agenda
- November 30 to December 4. AWS re:Invent, Las Vegas. AWS advertises a free livestream for keynotes and selected sessions. For agent builders, the security program includes identity for AI agents and access to specialists on application security and threat response. Event · Security guide
- December 2. GitHub's default block on
pull_request_targettakes effect for public repositories. Review the November 16 brownout results and the applicable event policy, particularly for review bots handling fork PRs. Private and internal repositories are outside this change. Policy and preparation - December 18. MAI-Code-1.1-Flash retires across Copilot. GitHub recommends the latest MAI Code Flash model. Check organization model policies before switching dependent workflows. Deprecation notice