|
Issue · November 18, 2026
OpenAI’s GPT-6.2 launch puts cost per finished task in focus, while Gemini expands everyday delegation and developers confront the harder work of verification and containment.
AI in general
Frontier models, research and policy
|
2 stories |
Only 2 strong recent items found.
Score · 98 / 100
GPT-6.2 Astra and Sol aim to lower the cost of finished professional work
Source: OpenAI — November 17, 2026
MODELS · PROFESSIONAL WORK · COST EFFICIENCY
OpenAI has introduced GPT-6.2 Astra for demanding projects and GPT-6.2 Sol for everyday work, emphasizing better deliverables with fewer output tokens rather than intelligence gains alone. Astra targets documents, spreadsheets, presentations, coding, and computer use; Sol is positioned as the economical default. The useful distinction is between token pricing and the total cost of completing a task: OpenAI claims improvements on both, but its benchmark results still need validation against your own workloads.
What matters
- Astra’s API pricing is $8 per million input tokens, $0.40 per million cached input tokens, and $40 per million output tokens. Input and output prices are 20% below GPT-6 Astra’s.
- OpenAI reports a 5.2-point improvement over GPT-6 Astra on Agents’ Last Exam at less than one-third of the benchmark task cost. That is a workload-specific result, not a universal discount.
- Sol retains GPT-6.1 Sol’s token pricing. In Cognition’s FrontierCode 2.0 evaluation, OpenAI says it comes within 0.5 points of Astra at about one-fifth the cost, with both at maximum settings.
- Template fidelity and fewer edits to finished artifacts are central claims. Test representative deliverables and review time—not just prompt-response quality—before changing your default.
VerdictREAD FULL** — Worth reviewing the examples and evaluation caveats if you choose models or produce substantial professional deliverables.
Score · 95 / 100
Gemini Tasks becomes free in six countries and gains deeper access to everyday workflows
Source: Google — November 17, 2026
PERSONAL AGENTS · GOOGLE WORKSPACE · PRIVACY
Google is rolling out free Gemini Tasks access in the U.S., Japan, Korea, Canada, Australia, and India, powered by Gemini 4 Hydrogen. Advanced features start in the U.S. and connect delegation to calendars, phone calls, website logins, and purchases. This moves the product beyond answering questions toward acting across accounts—but availability and permission settings matter as much as the demonstrations.
What matters
- Indonesia is not in the announced six-country free rollout. Don’t assume the new capabilities are available in Yudi’s region.
- New integrations include Google Password Manager and Google Pay; Google says purchases support confirmation or pre-approval and an upfront spending cap.
- Tasks execute in account-dedicated, hardware-isolated virtual machines, with additional network and prompt-injection defenses. Google explicitly acknowledges that tasks can still fail or require a manual handoff.
- A consequential privacy caveat: TechCrunch reports that temporary and history-off Gemini chats will be retained for around two months, up from 72 hours. Google is also separating chat-history and model-training controls; “history off” does not mean immediate deletion.
VerdictREAD FULL** — Read the permission and security sections before connecting credentials, private files, or payment methods.
Software engineering
Coding agents, developer tools and infrastructure
|
3 stories |
Score · 97 / 100
Microsoft’s MAI Code 2 models compete on the economics of routine coding
Source: Microsoft AI — November 17, 2026
CODING MODELS · GITHUB COPILOT · INFERENCE COSTS
Microsoft has introduced MAI Code 2 for debugging, multi-file implementation, repository reasoning, and UI work, alongside Code 2 Flash for explanations, targeted edits, and routine fixes. The models are also arriving in GitHub Copilot. Microsoft is explicitly optimizing for common enterprise coding rather than claiming these are the best models for every difficult or long-running task.
What matters
- Across four coding benchmarks, Microsoft reports Code 2 achieves 98% of GPT-6.1 Sol’s performance at 40% of the cost. This is a vendor-reported aggregate, not a guarantee about a particular repository.
- Flash targets high-volume work where latency and unit economics can matter more than maximum reasoning capability.
- Microsoft says these models are not yet tuned for fundamental science or very long-horizon agentic work—a useful boundary when deciding what to route elsewhere.
- The practical experiment is to send routine edits to a cheaper model, then compare accepted patches, regressions, and reviewer effort against your current default.
VerdictSKIM** — Focus on workload boundaries and benchmark methodology; a small repository-specific trial will tell you more than the headline ratios.
Score · 95 / 100
GitHub’s security pilot turns thousands of AI findings into checked reports with proposed fixes
Source: GitHub Blog — November 17, 2026
AI SECURITY · OPEN SOURCE · VERIFICATION
GitHub and Microsoft’s Codename MDASH pilot offers a more useful measure of AI security work than raw vulnerability counts. Across six open-source projects, automated scanning was followed by AI triage and manual researcher verification before maintainers received private reports with proofs of concept and suggested fixes. The result supports a clear operating principle: whoever runs the scanner should absorb the verification burden, rather than exporting it to maintainers.
What matters
- Thousands of initial findings became 99 checked private reports. As of November 15, maintainers had fixed 55, with 14 fixes already released.
- The workflow combined Microsoft’s MAI Cyber model, GitHub’s triage tooling, and human review; this was not unattended scanning-to-merge.
- Seven researchers spent about two days per project checking findings. GitHub acknowledges that this careful process does not yet scale.
- For teams deploying AI scanners, require deduplication, threat-model validation, private disclosure, and a proposed fix before a report reaches a maintainer.
VerdictREAD FULL** — A concrete account of how to make AI-assisted security useful without creating another review queue.
Score · 95 / 100
External sandboxes separate an agent’s code execution from its host
Source: Simon Willison — November 17, 2026
AGENT SANDBOXING · SECURITY · DEVELOPER TOOLS
Simon Willison and Prime Radiant have demonstrated an “external sandbox” pattern: the agent stays outside the sandbox while code it executes runs inside a dedicated isolated environment. Three server implementations expose the same localhost API, allowing tools such as Pi, Claude Code, and Codex to delegate execution. The appeal is a more explicit boundary around untrusted code, particularly when experimenting with local models.
What matters
- An optional HTTP/HTTPS allow-list proxy restricts which hosts sandboxed code can contact.
- Credential replacement supports authenticated requests without handing the underlying secret to the agent.
- Network policy can be adjusted while the sandbox is running, making the approach useful for iterative development.
- The boundary protects delegated execution; it does not automatically constrain every other tool or permission available to the agent outside the sandbox.
VerdictREAD FULL** — Especially relevant if you run coding agents against real files or want to experiment without exposing your normal development environment.
Design & creative
Creative workflows and user experience
|
1 story |
Only 1 strong recent item found.
Score · 91 / 100
Gemini Carbon showcases point toward interactive creative prototypes, not just generated images
Source: Google — November 17, 2026
CREATIVE CODING · 3D · INTERACTIVE PROTOTYPING
Google’s new builder showcase collects five Gemini 4 Carbon projects spanning animated landscapes, particle simulations, glass physics, data tools, and a reference-image-to-3D environment. For designers, the interesting shift is toward generating explorable behavior and environments rather than isolated visual assets. These are curated examples, however—not evidence that the same results are reproducible in a single prompt or ready for production.
What matters
- One project turns a reference image into a 19th-century optics laboratory with 92 instruments that users can test and tune.
- Other examples include a CUDA/C++ meadow of 52,000 flowers and grass blades, and a WebGPU simulation with 1,048,576 particles.
- The useful design application is early experimentation with motion, material behavior, spatial layouts, and interactions that static mockups cannot demonstrate.
- Google does not provide a standardized account of iteration time, cost, or manual correction across the examples. Treat the showcase as inspiration, not a delivery estimate.
VerdictSKIM** — Browse the demonstrations for interaction ideas; skip it if you need a tutorial or a measured comparison.
Open-source watch
Projects gaining meaningful traction
|
3 stories |
These are newly surfaced trending picks, not necessarily newly launched projects. Traction figures come from today’s supplied GitHub Trending inventory.
Project · 01
Orca — stablyai/orca
AGENT · DEVTOOL · PRODUCTIVITY
Orca coordinates a fleet of coding agents, including Claude Code and Codex, using your own subscriptions. Desktop, mobile, and remote-runtime access make it interesting for developers whose bottleneck is supervising several workstreams rather than starting another agent.
What matters
- The inventory reports 1,052 stars today—a discovery signal, not evidence of reliability.
- Potential fit: managing parallel coding tasks across locations and devices.
- Before connecting real repositories, inspect credential handling, runtime isolation, subscription compatibility, and how concurrent changes are kept separate.
VerdictWATCH** — Promising orchestration, but evaluate isolation and review burden before scaling up agent concurrency.
Project · 02
Cursor Plugins — cursor/plugins
DEVTOOL · AGENT · EXTENSIBILITY
Cursor’s repository contains its plugin specification and official plugins. It is worth tracking as a concrete reference for packaging reusable capabilities around coding agents, rather than repeatedly distributing loose prompts and setup instructions.
What matters
- The inventory reports 196 stars today.
- Most relevant to teams building internal agent extensions or standardizing development workflows.
- Review plugin permissions and portability before treating a vendor-specific package format as shared infrastructure.
VerdictTRY** — If your team already uses Cursor, inspect an official example and test one narrowly scoped workflow.
Project · 03
Karakeep — karakeep-app/karakeep
AI · PRODUCTIVITY · SELF-HOSTED
Karakeep is a self-hostable collection app for links, notes, and images, with AI-assisted automatic tagging and full-text search. It addresses a quieter but practical problem: making saved material retrievable after the initial bookmarking impulse has passed.
What matters
- The inventory reports 38 stars today; practical relevance matters more here than leaderboard position.
- Useful for a personal research library or a collection of design references.
- Self-hosting does not by itself ensure private AI processing. Check the configured model provider and what content leaves your server.
VerdictTRY** — A sensible small-scale experiment if your saved links and references are becoming hard to find.
Editor’s note
Fresh, non-repeated developments with concrete implications for task cost, delegated authority, verification, creative prototyping, and useful developer workflows—not announcements selected merely to fill slots.
30-second feedback
How useful was today’s digest?
Your choice opens the short form with your rating filled in. Or share a quick note.
Prefer reading on Telegram? Subscribe to The Merpati Post →
The Merpati Post · Daily AI Briefing
|