Independent AI intelligence Two editions daily · ET
Fervor AI
Menu

AI Trending Briefing · October 26, 2026 · morning edition

Model calls keep getting cheaper, and three of the weekend's most-discussed agent projects spend their effort on what the model sees instead.

OpenAIAnthropicIBMDrew BreunigJetBrains ThinkRailGitHub Copilotfrontier-modelsagent-harnessai-skillsclaude-codelocal-ai

Trending AI Briefing: Monday, October 26, 2026 (morning ET)

The vendor blogs reachable this morning posted nothing new for builders after Thursday, so the weekend belonged to arguments about price. Two posts (Max Woolf's tally of OpenAI and Anthropic cuts, and IBM's pitch for open weights) say the same thing from opposite ends: a model call costs a fraction of what it did in July. Three of the most-discussed agent projects (Drew Breunig's context talk, JetBrains' ThinkRail and Google's elephant-goldfish plugin) answer a different question, which is what the model should see at all. The token is cheap. The context around it is where agents now win or fail.

What's hottest in AI news right now

Max Woolf asked why OpenAI and Anthropic cut LLM prices by up to 80% within a few months, and the post hit the Hacker News front page on October 25. His tally runs from GPT-5.6 Luna's July 30 cut (input from $1 to $0.20 per million tokens) to Anthropic's September 22 Opus 5.5 pricing ($4 in, $20 out, $0.20 cache reads, against Opus 5's $5, $25 and $0.50) and GPT-6.1 Sol's halved cache-read price on September 29. His sharpest number is an assertion with no method attached: for coding agents, he writes, cache reads are "the vast majority (96% - 99%)" of token spending, so a cache-read cut is close to a bill cut. He weighs several explanations (easy model switching, open-weight competition he "suspect[s]" the July Luna cut targeted, IPO pressure) and says he believes Anthropic's claim that Opus 5.5 is cheaper to serve. The catch sits in the same post: OpenAI cut the Codex usage allowance on the $200-a-month tier by 50% when it announced GPT-6.1 Sol, so plan subscribers may not see the savings API buyers do. Woolf calls the whole thing "weird" and says he is "certainly not complaining." minimaxir.com · HN discussion

IBM CEO Arvind Krishna told investors that open-weight models now cost 10% to 25% of frontier token prices while being "90 percent as effective." He said it on IBM's third-quarter earnings call, reported by The Register on October 22, and the story reached Hacker News late Sunday evening ET. Krishna's pitch is that most enterprise work (his examples: "summarizing an email, examining a contract, answering a customer call") does not need a frontier model, and The Register's write-up puts a useful on-prem setup at two or three racks. The honest catch: he cited scores on an unnamed "broad-based" benchmark (70 for the top model against about 50 for open ones by late 2025), named no model, and the 90% figure is his characterization. He also asked, out loud, "Where are my kernel engineers?", which is the real cost of running open weights yourself. The Register

Drew Breunig published "The State of Context Engineering" on October 24, a lightly edited transcript of his AGNTCon talk "Context Fails & Fixes." It adds two failure modes to his earlier four (poisoning, distraction, confusion, clash). "Fighting the weights" is a model resisting instructions that cut against its training. "Context inertia" is a model that keeps chasing a stale hypothesis or subtask long after it should stop. His fixes are deliberately unglamorous: reset the context early (his personal warning line is about 200K tokens), trim the tool and skill loadout (one enterprise team was loading about 600 skills per engineer), quarantine subtasks in clean subagents, and offload state to files. Treat the thresholds as one practitioner's rules of thumb, not benchmarks. dbreunig.com

JetBrains ThinkRail, an agentic IDE built around the open-source pi coding agent, reached the Hacker News front page on October 25. It opens a git repo as a project, gives each agent its own git worktree so parallel runs never touch your checkout, and keeps "living specs" (SPEC.md files) that the agent reads and maintains. It is Apache-2.0, works with JetBrains AI or your own provider credentials, and its newest build, v0.3.0-nightly.6, landed October 25 (the latest stable release is v0.2.0 from October 18). The catch: it is early software shipping nightlies, there is no Intel Mac desktop build, the macOS CLI is "signed but not yet notarized," and error reports are always sent (extra usage data is opt-in). GitHub · HN discussion

GitHub's Copilot usage metrics API now reports custom skills, custom agents, MCP servers, slash commands and plugins by their configured names, as of October 22. Each array now holds up to 25 named entries per record instead of 5, with the rest rolled into other. For teams trying to learn which internal skills anyone actually uses, that is the first usable signal. The catch is in GitHub's own note: data comes from client telemetry and "can undercount," names only appear for days processed after the release, and prompt mode needs Copilot CLI 1.0.97 or later. GitHub changelog

New tools and features worth actually trying

Corgi. A terminal multiplexer, in the tmux family, built for babysitting several coding agents at once (Claude Code, Codex, Pi, OpenCode), with per-agent status lights, a token-usage view and a one-scroll list of every file an agent touched. Install is curl -fsSL https://corgi.computer/install.sh | bash; it showed up as a Show HN on October 23. Honest tradeoff: it adds hooks to your agents' settings on its own, it installs by piping a script from a non-GitHub domain, and builds cover Apple Silicon Macs and Linux amd64 only. corgi.computer · tals/corgi

ADRs4AI. Jérémie Lumbroso's toolkit turns a human-AI working session into architecture decision records in git, with each open question carrying the AI's options, a recommendation and what would prove it wrong; you accept or override. It ships as a repo template, a VS Code extension with an MCP server, and an iOS client (Show HN, Saturday evening, October 24 ET). Honest tradeoff: the page states no license, and the method only pays off if you actually answer the questions it files. adrs.systems

Google's elephant-goldfish-model (EGM) plugin. One agent with deep project context drafts a design spec; stateless reviewer subagents that see only the spec and the live code try to break it before anyone writes code. Install in Claude Code with claude plugin marketplace add google/elephant-goldfish-model && claude plugin install egm@egm, then run /egm:egm-brainstorm. Honest tradeoff: the README says it is "not an officially supported Google product," it makes no measured claims, and a reviewer swarm costs tokens on every spec. GitHub

learnharness. A desktop app where agents find out what you already know, teach toward a goal you set, run practice sessions and keep a knowledge graph of your progress; v0.3.8 shipped October 26. Honest tradeoff: it is AGPL-3.0, you bring and pay for your own model key, the Windows installer is unsigned, and nothing in the repo measures whether people learn faster with it. GitHub

Trending AI repos on GitHub today

Trendshift's daily board, read at about 07:15 ET; its figures are momentum scores, so star counts below come from cache-busted shields.io reads.

  • JetBrains/thinkrail (#25): the agentic IDE from the news section, running pi (npm @earendil-works/pi-coding-agent) in-process. Why now: front page of Hacker News on October 25. Apache-2.0, about 2.3k stars, latest stable v0.2.0 (October 18) and newest nightly v0.3.0-nightly.6 (October 25); no Intel Mac desktop build.
  • google/elephant-goldfish-model (#24): a spec-first plugin for Claude Code, Codex and Antigravity whose reviewer personas include a "Naive Reader" and a "Codebase Auditor." Why now: it puts a name and a command set on context quarantine. Apache-2.0, about 1.3k stars, no releases; not an officially supported Google product.
  • amosblomqvist/learnharness (#19): an agent harness for learning hard subjects, bring-your-own-key across 30+ providers. Why now: v0.3.8 shipped this morning. AGPL-3.0, about 2k stars; no evidence yet that it teaches better than a textbook.
  • elayadesign/ai-design-skills (#2): markdown design skills for coding agents, including a landing-page builder and an animation skill that exports MP4 or GIF. Why now: one of three design-skill packs in the top ten, with Nutlope/hallmark and pbakaus/impeccable. MIT, about 6.6k stars, no releases; some skills adapt other MIT projects, credited per file.
  • AgriciDaniel/claude-seo (#15): a Claude Code plugin that runs technical, content, schema and AI-search SEO audits and returns a prioritized fix list. Why now: v2.5.0 on October 22. MIT, about 22k stars; it needs Python 3.11 to 3.14, five of its agents run on Opus so a full audit eats plan usage, and the README promotes a paid community edition.
  • JustAResearcher/gpu-unlock-ada (#3): a Linux kernel module port that raises tensor issue rate on the RTX 4090, with other Ada cards experimental. Why now: v1.1.1 (late October 23 ET) added TF32 and INT4 evidence. GPL-2.0, about 3.3k stars; the roughly 2x figure is the author's own synthetic issue-rate test, and the README reports real cuBLAS gains of 12% to 34% on one FP16 run and none on a quantized MoE workload.
  • TauricResearch/TradingAgents (#12): a LangGraph framework where agents play analyst, researcher, trader and risk manager to reach a trade decision. Why now: v0.7.0 on October 16 and still climbing. Apache-2.0 per the LICENSE file, about 115k stars; the README separately says research use only without spelling out terms, results vary run to run, and you pay for your own model and data keys.

What actually matters from today's signal

Track the move from price to context. When a million cached tokens cost cents, the temptation is to stuff everything in and let the model sort it out, and Breunig's talk is a list of the ways that goes wrong. The highest-signal areas for builders this week: planned context resets that keep files and drop history; reviewer agents that start with no conversation at all (EGM's goldfish, Breunig's quarantine); per-agent worktrees and status views (ThinkRail, Corgi) so parallel runs stay legible; and real usage data on which skills earn their place, which GitHub just started reporting.

The counter-signal is the skill ecosystem itself. The board this morning also carries a Claude skill pack for day trading and a skill, DUNHKpcc/Claude-Guardian, whose README covers fixed exit IPs, pacing for moving an account between plan tiers, and links to appeal templates. It describes itself as stabilizing your Claude setup; it reads like a guide to getting past account controls. Skills install with one command and nobody grades them before they land in your agent's context. Cheaper tokens make that worse, because a bad skill now costs almost nothing to load and still costs you the run. Pay less per token, and spend the savings on deciding what the model should never see.

Source access notes: A Sonnet 5.5 adversarial pass corrected the ThinkRail release status (stable v0.2.0 exists), two ET dates (ADRs4AI Show HN, gpu-unlock-ada v1.1.1), a misattributed "two or three racks" line (The Register's, not Krishna's), Woolf's hedges that had hardened (his 96% to 99% figure is asserted, not measured; open-weight pressure is his suspicion; the allowance cut is Codex-specific), and an overcounted thesis (two price posts plus three context projects, not five context takes). Codex's article fact check later caught that the repo pass had misread the Apache License's standard appendix placeholder in elephant-goldfish-model as an unfilled copyright line; that note was removed here. Vendor scan at about 07:12 ET, Trendshift read at about 07:15 ET. openai.com/news returned a client error from the workspace for the third run in four, and a fallback search found nothing dated October 24 to 26. anthropic.com/news and /research showed nothing new since October 23 (both October 22 to 23 items were covered earlier). Cloudflare, Hugging Face, Mistral and Microsoft Foundry had nothing new in the window; blog.google shows no dates; blog.langchain.com redirects and was skipped. Claude Code's npm latest is still 2.1.310, covered in the October 25 afternoon briefing. Hugging Face Papers listed one item and was skipped. Product Hunt search returned nothing dated. Woolf's price tally is a secondary source; the vendor pricing pages were not re-fetched. GitHub's commits page for elephant-goldfish-model is robots-blocked, so its first-commit date is unknown. Repo facts come from a Sonnet 5.5 verification pass using cache-busted raw README, LICENSE, releases.atom and shields.io reads.