Trending AI Briefing: Tuesday, November 17, 2026 (afternoon ET)
Two new model families landed on Ignite day, and both came with a cost-per-task claim next to the capability claims. Microsoft led with it; OpenAI put it beside every benchmark. Cost per task is a number the harness around the model shapes as much as the model does. The same afternoon, an independent benchmark of 22 harness configurations showed how much that matters, and GitHub, which Microsoft owns, started publishing harness-versus-harness charts and a tool for testing your own agent setup. The model is still the engine. The bill now depends on the car.
What's hottest in AI news right now
OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on November 17, in ChatGPT Work, Codex and the API. API prices are $8/$40 per million input/output tokens for Astra and $2/$10 for Sol, per OpenAI's developer forum announcement; 9to5Mac reports Astra's standard rates are 20% below GPT-6 Astra's while Sol keeps GPT-6.1 pricing. OpenAI says 6.2 Astra at xhigh effort scores 5.2 points above GPT-6 Astra at max on Agents' Last Exam at less than half the cost per task, and that Sol nearly matches Astra on FrontierCode 2.0 at about one-fifth the cost. It also calls Astra "our most capable model ever for cyber," up 22 points on ExploitBench Internal Port. The catch: every number is OpenAI's own, the main announcement page returned server errors to our reader this afternoon, and the rollout does not add either model to regular ChatGPT, only Work and Codex. OpenAI forum · 9to5Mac
Microsoft AI shipped MAI Code 2 and MAI Code 2 Flash the same day, straight into GitHub Copilot. The pitch is price: averaged over four coding benchmarks, Microsoft says MAI Code 2 delivers 98% of GPT-6.1 Sol's performance at 40% of the cost and 96% of Claude Opus 5.5's at 5%. Read the footnotes before the headline. Three of the four benchmarks ran on Microsoft's internal MSBench through Copilot CLI with no subagents, and MAI Code 2's cost is shown at a 50% promotional price. All models ran at max reasoning effort, and the fourth benchmark's scores come from Artificial Analysis. Microsoft also says the models are "explicitly not yet tuned for harder problems," naming fundamental science research and very long-horizon agentic work. In Copilot, MAI Code 2 reaches Pro and up, Flash reaches every plan including Free, and both bill under usage-based billing. Microsoft AI · GitHub Changelog
Runta published FrontierHarness Eval v1.1 on November 17, and it is the most useful read of the day. Built with Baseten, it ran 22 harness configurations with 5 open models on 62 terminal tasks, four attempts each: 27,280 runs and 83 billion tokens. The model explains most of the cost spread (92.8%), but with the model held fixed, the harness changed average cost per task by 1.8 to 5.0 times. Rankings did not travel: Codex CLI was the cheapest harness per task with GLM-5.3 and 16th with Kimi-K3, and the median harness moved 11 of 22 places between models. Across harnesses, steps per pass tracked cost per pass (a rank correlation of +0.51), while prompt size, tool count and cache rate showed almost none. Honest limits from the authors: open models on one provider only, Baseten's prices and cache, and roughly ±10 points of uncertainty per cell. The project's dashboard lists Runta, Baseten and Z.ai as sponsors, and Z.ai makes the two GLM models tested. It was on the Hacker News front page when we read it around 15:15 ET. Runta · HN
GitHub's benchmarks site added a Harnesses view on November 17. It compares Copilot CLI with Claude Code and Codex CLI on the same model at the same reasoning effort, plotting resolution rate against mean cost per task with paired differences and approximate 95% intervals. Good chart format; also a vendor grading its own harness against two rivals, at medium effort only. GitHub Changelog
GitHub Copilot Evals entered public preview the same day. An eval is a repeatable test of an agent's result and of how it got there. /eval creates one for a skill, custom agent or MCP tool in your project, runs it after you approve each run, and compares runs before and after a change. It runs in VS Code, Copilot CLI and the Copilot app, on Microsoft's local Vally engine. GitHub Changelog
Copilot's security review can now run GitHub secret scanning on an agent's changes before merge, in public preview since November 17. Once you opt in, it adds GitHub's detectors for more than 500 secret types plus a fine-tuned model for credentials with no fixed format, such as passwords inside connection strings. It works in Copilot CLI, the Copilot app, Copilot SDK sessions and Copilot code review. It is off by default, checks only your changes and not repository history, and spends AI credits under a separate Secret Protection SKU. GitHub Changelog
New tools and features worth actually trying
Secret scanning for agent sessions. In Copilot CLI, run /settings aiSecretDetection on, then /security-review on your own or an agent's changes. Honest tradeoff: it reads the diff, not history, so a key committed last month stays invisible, and each scan spends AI credits.
Copilot Evals. Write one eval for the skill or MCP tool you changed most recently and rerun it after every edit. Honest tradeoff: it is a preview tied to Copilot plans, and an eval only catches the regressions you thought to describe.
FrontierHarness dashboard. Before switching models, look up whether your harness stays on the cost frontier with the new one. Honest tradeoff: it covers open models served by Baseten only, so Claude, GPT and Gemini users get a method, not an answer.
Copilot code review adaptive effort. Adaptive picks Lite, Balanced or Max per pull request and labels which level it chose, capped by your enterprise, org or repo maximum. Honest tradeoff: Balanced stays the default, Adaptive is a preview, and "risk" is the reviewer's own judgment of your PR.
Trending AI repos on GitHub today
From Trendshift's daily board at about 15:20 ET (momentum scores, not star totals; ranks drift). Stars are rounded shields figures.
- elementalsouls/Claude-BugHunter (#1): a Claude Code skill bundle for bug bounty and external red-team work, 83 skills and 15 slash commands per its README. Why now: it topped the board this afternoon. MIT, about 8.9k stars, latest release v2.1 (June 5). Caveat: offensive tooling needs written authorization for every target, and the README carries a sponsor block.
- BaaRai/ARTEX (#4): a self-hosted, LLM-driven multi-agent penetration-testing platform. Why now: v0.6.0 on November 12. AGPL-3.0, about 6.1k stars. Caveat: its README names Autumn-27/ARTEX as upstream, so this is an English-translation fork, not the original project.
- ag-ui-protocol/ag-ui (#8): the event-based protocol connecting agents to user-facing apps, with a frozen 1.0 spec and TypeScript, Python and .NET SDKs. Why now: a
release/2026-11-17tag landed today. MIT, about 20k stars. - nexu-io/openmotion (#13): an open-source studio that has coding agents make motion video as HTML, with bring-your-own keys. Why now: a v0.1.0 tag on November 17. Apache-2.0, about 11k stars. Caveat: its release feed also holds a later-numbered test tag, so check which build you install.
- NVIDIA/SkillEvaluator (#22): a multi-tier framework for testing whether agent skills are safe and actually help. Why now: a v1.0.0 tag landed November 17, four days after v0.9.0. Apache-2.0; the star figure did not look reliable this run. Caveat: release notes we read described the final release and PyPI package as still pending, so check PyPI before pinning 1.0.0, and it needs Python 3.12 to 3.14.
- getarcaneapp/arcane (#23): a web UI for managing Docker. Why now: v3.0.1 on November 15. BSD-3-Clause, about 9.2k stars. Caveat:
mainis the v3 line; 2.x lives on its own branch.
What actually matters from today's signal
Track cost per finished task, measured on your own tasks with your own harness. Today OpenAI and Microsoft both put that unit next to their new models, and an outside benchmark showed why their numbers will not transfer to your setup: the same model cost up to five times more in one harness than another, and the cheapest harness with one model fell to 16th of 22 with another. For builders, four areas carry the signal. Harness choice now belongs in model evaluations. Steps per pass is the cost signal to watch (Runta found a correlation, not a proven cause). Evals for your own skills and MCP tools are becoming a product feature. And security checks are moving into the agent session itself, before the pull request.
The counter-signal is who is holding the ruler. Microsoft holds it twice: MAI Code 2's headline runs through Copilot CLI and a promotional price, and the harness chart on Microsoft-owned GitHub is drawn by the harness vendor. OpenAI's gains are OpenAI's evals. The one outside study tested only open models on one host, and its dashboard lists the maker of two of those models as a sponsor. Nobody has published a vendor-neutral comparison of Claude Code, Codex CLI and Copilot CLI on closed frontier models. Until someone does, the only cost-per-task number worth trusting is the one from your own repository.
Source access notes: Read Tuesday November 17 from about 15:10 to 15:45 ET. Pages were read through TailFinder's web reader on the linked Mac until the Mac dropped off the bridge around 15:25 ET; later pages were read through a web-fetch subagent. openai.com/news is behind a Cloudflare check and the GPT-6.2 announcement page returned server errors to both readers, so GPT-6.2 figures come from OpenAI's developer forum announcement and 9to5Mac; qz.com was unreachable. Anthropic's newsroom had nothing after November 10. Claude Code 2.1.326 (npm, November 16) remains the latest release and was covered this morning. Not checked this run: x.ai, Mistral, Cloudflare, Google, Microsoft Agent Framework and Foundry blogs, LangSmith changelog, Hugging Face papers, the Codex changelog and Product Hunt. Repo figures come from a Sonnet 5.5 verification subagent that read raw READMEs and LICENSE files and fetched releases feeds; api.github.com was blocked, so stars are rounded shields figures. The subagent flagged gylive/google-ultra (#10) as a Google One promotion workaround rather than a developer tool, so it is left out; Claude-BugHunter's release date is from its git tag. A Sonnet 5.5 adversarial pass caught: Claude-BugHunter's v2.1 dated October 16 (the tag is June 5); a Z.ai sponsorship attributed to the Runta post (it is on the FrontierHarness dashboard); SkillEvaluator described as having only a draft 1.0.0 (a v1.0.0 tag exists, publication unclear); the lede claiming neither lab led with capability (OpenAI did); GitHub and Microsoft counted as separate rulers; Microsoft's "not yet tuned" line narrowed; and a steps-per-pass correlation stated as a cost driver. All fixed. It also reported a lower HN point count than our front-page read, so the point figure was dropped.