On 2026-11-05, Hugging Face Harness Optimization researcher Joël Niklaus (@joelniklaus) released the Harness Leaderboard: same model, provider, and reasoning effort with only the harness swapped, covering GLM-5.3-Flash, DeepSeek-V4.1-Flash, and Flint on Terminal-Bench 2.1 and DeepSWE—28,849 trials and about $4,746 of inference. Author-reported averages: mini-swe-agent ~51.4% resolve rate; Pi ~49.4% at about $0.29 per solved task. This is an individual research leaderboard, not a Hugging Face company product announcement.
Zoom
Key Takeaways
Source: personal post by HF Harness Optimization researcher Joël Niklaus on 2026-11-05T16:04:08Z—not a company announcement
Scale (author): 9 harnesses on open models, 28,849 trials, ~$4,746 inference spend
Method: pin model/provider/reasoning effort; 3 attempts per task; Terminal-Bench 2.1 and DeepSWE
Results (author): averaged mini-swe-agent ~51.4%, openclaw ~27.3%; Pi ~49.4% at ~$0.29 per solved task
Use: public HF Space leaderboard + trajectories; Harness Lens documents prompts/tools for each harness
On 2026-11-05 Moonshot officially launched Kimi Hosted Agents, selling the agent runtime harness as a cloud service: isolated sandboxes, persistent sessions, failure recovery, and audit export, billed as sandbox hours + model tokens + plugin calls. Official docs list 140+ Skills and 30+ plugins, plus kha CLI, console, and REST API (kimi-api-version: 2026-09-01-beta). Self-hosted sandboxes are enterprise-verified only.
Zoom
Key Takeaways
Official: GA on 2026-11-05 as Harness as a Service—no self-built sandbox/session/retry stack required
Billing (official pricing): total = sandbox hours × ¥0.072/hour + model inference + Σ(plugin calls × unit price)
Bundle: vendor says 140+ Skills and 30+ plugins; product overview lists 12 official templates (finance/game/SRE/community)
Start: platform.kimi.com; recommended path is kha CLI install + kha login, or REST/SSE
Limits: self-hosted environments and Inference Hook are enterprise-verified only; API still tagged 2026-09-01-beta
On 2026-11-06 OpenAI shipped Codex CLI 0.167.0 on GitHub and the ChatGPT/Codex changelog. Open the agent command center with codex agents --no-daemon (unread results marked); switch connected backends in the terminal session picker while keeping drafts; preview local images; expand patches during approval prompts; optionally enable a Google Cloud Vertex AI setup wizard.
Key Takeaways
Official: GitHub rust-v0.167.0 (2026-11-06T01:40Z) and same-day ChatGPT/Codex changelog entry
Command center: codex agents --no-daemon; unread results marked in the center and session picker
Multi-backend: switch connected backends from the terminal session picker while keeping drafts
UX: local image previews in supported terminals; expand patch details while an approval prompt is open; network approval prompts close when requests end and serialize per endpoint
Auth: enable gcp_setup_wizard for Vertex AI under Sign in another way (ADC or credential file); opt into macOS/Windows system proxies from /experimental
On 2026-11-06 Vercel updated Token Pass so featured-model usage through AI Gateway draws credit automatically with a standard API key, OIDC token, or fx login. When credit runs out, those auth modes fall back to team AI Gateway credits; dedicated Token Pass keys still stop until the next billing period.
Zoom
Key Takeaways
Official: Vercel changelog 2026-11-06 — Token Pass credit now applies automatically across AI Gateway
Behavior: standard API key, OIDC, or fx login can draw Token Pass credit; dedicated key optional
Fallback: when credit ends, standard key/OIDC use team AI Gateway credits; dedicated Token Pass keys still stop until next period
Product terms: ~$30/month for ~$80 featured open-weight credit on Pro; Token Pass path stays ZDR
Guidance: early Token Pass key users should switch to standard key/OIDC to enable fallback
On 2026-11-05 Vercel launched Schedules in public beta on all plans: static jobs in vercel.json plus dynamic one-shot/cron schedules via SDK/CLI/API targeting private Functions, HTTP paths, or Queues. Eve and AI SDK samples let agents schedule follow-up work. Existing Cron Jobs remain supported.
Zoom
Key Takeaways
Official: Vercel 2026-11-05 changelog + blog Introducing Vercel Schedules; amplified on @vercel
Capabilities: static (vercel.json) and dynamic (SDK/CLI/API) one-shot/cron; Function, HTTP, or Queue targets
Agents: Eve ships dynamic scheduling; AI SDK sample uses Schedules.create + Queue wake-up
Billing (vendor): ~$5 per 1M runs and $10 per 1M creations on Pro/Enterprise; Hobby 500 creates / 5,000 runs per month
Existing Cron Jobs keep working; schedule/v1beta functions have no public URL
Anthropic updated Claude Enterprise security and admin features: Inference Hooks (prompt hooks and pre-tool-call hooks) are generally available so prompts and tool calls can hit the organization's existing DLP/security stack for allow/deny before the model or tool runs. The same post also introduces a Security Center, centralized management, and additional access controls (mostly beta).
Zoom
Key Takeaways
Official: claude.com article More visibility and control over Claude Enterprise for security teams
GA: Prompt hooks and pre-tool-call hooks into existing enterprise DLP/security tools
Coverage: prompts and tools/skills/plugins/MCP on supported products; one org setup
Same post, mostly beta: Security Center, centralized management, Admin API sign-out, self-serve IP allowlist, Desktop shadow-IT block, Cowork EDR
Limits: Enterprise; hooks see transcript/tool-call text, not raw file bytes; configure timeout fail policy
On 2026-11-05, University of Chicago Law School Harry Kalven, Jr. Professor William Baude (@WilliamBaude) posted on X that while he was at lunch, Claude deleted a year of constitutional-law lecture notes and his current syllabus. He later said he recovered everything via Dropbox version history and called the post a public-service warning. When fetched the original had ~3,580+ likes, 200+ reposts, and ~570K views. This is a first-person account and community discussion—not an Anthropic incident report.
Zoom
Key Takeaways
Source: William Baude (UChicago Law Harry Kalven, Jr. Professor; @WilliamBaude), original post 2026-11-05 19:08 UTC
Claim: Claude deleted a year of con-law notes and the current syllabus over lunch; ~1.7h later he said Dropbox restored everything
On 2026-11-05 Dwarkesh Patel released Clocktower Bench: ~$20k API for 386 social-deception games, open harness. Author board: GPT-6 Astra Elo 1859, ~82% win rate vs Opus 5.5.
Zoom
Key Takeaways
Author: Dwarkesh Patel (@dwarkesh_sp) on 2026-11-05; clocktowerbench.com + Substack
Scale: ~$20k API, 386 rated games
Top: GPT-6 Astra Elo 1859; ~82% win rate vs Opus 5.5; ~+430 Elo vs Opus 4.6
Open source: github.com/dwarkeshsp/clocktower; HF dwarkesh/clocktower-games; human play
Social-deception eval is not a general IQ board; script/role variance applies
On November 5 Perplexity Developers (@perplexitydevs) said every Agent API effort level now costs less per query with equal or better research quality, and added an ultra tier for broad long-running research. Vendor figures: medium rose from ~55% to ~77% on BrowseComp/DSQA with nearly 90% lower cost; xhigh held ~76.5% WANDR at ~25% of August 2026 cost; ultra runs three parallel researchers plus a consolidator at ~85.9% WANDR and about $21.71/task.
Zoom
Key Takeaways
Official (@perplexitydevs + blog): every effort level cheaper with equal/better quality; new ultra
Vendor: medium BrowseComp/DSQA ~55%→~77% with nearly 90% lower cost
Vendor: xhigh WANDR ~76.5% at ~25% of August 2026 cost
ultra: 3 parallel research agents + consolidator; WANDR ~85.9% at ~$21.71/task on Claude Opus 5.5
Use effort="ultra" or pin [email protected]; background mode for long runs
On November 5 Ant Group’s Ant Ling (@AntLingAGI) launched Ling-3.2-flash: a 560B-total / 25B-active native multimodal MoE with a 1M context window and native image, text and video input. The official changelog targets complex visual understanding, multimodal long-horizon agents, 3D understanding and interaction, and medical multimodal work; APIs are OpenAI- and Anthropic-compatible and free for the first two weeks. The launch post cites vendor evals versus Claude Opus 5 (max): OmniDocBench1.5 92.21 vs 90.56, ZEROBench(Sub) 43.71 vs 39.52, Vision2Web 85.86 vs 80.20. DesignArena placed it about #10 on Image-to-Frontend with Elo 1238.
Zoom
Key Takeaways
Specs (official docs): 560B total / 25B active; 1M context; native image/text/video; four focuses: complex vision, multimodal long-horizon agents, 3D, medical multimodal
Availability: chat UI plus OpenAI/Anthropic-compatible APIs; changelog says free for first 2 weeks, then original pricing resumes
Vendor evals (official X): OmniDocBench1.5 92.21 vs Opus 5 90.56; ZEROBench(Sub) 43.71 vs 39.52; Vision2Web 85.86 vs 80.20
Third party: DesignArena Image-to-Frontend ~#10, Elo 1238; official price page stresses 2-week free promo and has not listed 3.2 USD rates yet
Hands-on: model id Ling-3.2-flash; use image_url / video_url blocks; weights not on Hugging Face (API-first)
On November 5 Stripe made agent access to Stripe analytics available to all users via the Stripe MCP or the Queries and Reports APIs directly. Compatible agents such as Claude and Cursor can answer natural-language questions about MRR, churn, fraud and reconciliation. The Reports API ships a modeled schema of 300+ tables with built-in guidance; Queries API is free through January 31, 2027 then usage-based, while Reports API stays free. Every analysis exposes the underlying SQL for verification.
Zoom
Key Takeaways
Availability: all Stripe users; connect via Stripe MCP or Queries/Reports APIs to Claude, Cursor and other compatible agents
Capabilities: natural-language MRR, expansion/contraction/churn, fraud, reconciliation and payments performance; can combine with tools like PostHog
Data model: Reports API includes 300+ modeled tables with usage guidance; every analysis exposes editable underlying SQL
Pricing (vendor): Queries API free through Jan 31, 2027 then per-credit with no monthly fee; Reports API remains free
Safety: read-only Data Metrics/Queries/Reports; do not paste secret keys into chat—use agent API keys or OAuth
On November 5, Solid founder Trevor Keith launched Glasshand: dedicated real iPhones and Macs hosted in a data center—not emulators—so agents can see the screen and tap or type via API. Works with Claude Code, Codex, Cursor and others; agents can install apps, sign into iCloud, add eSIM for SMS codes, run social workflows or build/test in Xcode. Official pricing is $60/mo for an iPhone and $90/mo for a Mac, month to month with cancel anytime and factory wipe on release. It ranked #1 on Product Hunt that day.
Zoom
Key Takeaways
Form factor: real iPhones/Macs in a data center, not VMs or emulators; agents see the screen and tap via Glasshand API/controller
Compatibility: Claude Code, Codex, Cursor, OpenClaw, Hermes or any agent; start with “Tell your agent to rent”
Capabilities: App Store installs, iCloud sign-in, eSIM SMS verification and iMessage; Macs for Xcode build/test
Pricing (official site): iPhone $60/mo, Mac $90/mo (USD + tax), month-to-month cancel anytime; no hourly/daily rental
On November 5, Cognition launched Devin Review 2.0. Since August, per-PR review cost fell about 82%, with a typical review at about $0.25–$0.75 and median latency about 94 seconds (down from about 139). It adds automatic review rules, incremental re-reviews, configurable per-PR spend limits, and tighter Devin Autofix loops. New controls include Devin Review Ultra and Custom Reviewers (beta).
Zoom
Key Takeaways
Cost: ~82% cheaper per PR since August (~5x); typical ~$0.25–$0.75; ~$2.25 average per fixed bug across customers (vendor)
Quality & speed: bugs found per PR rose 1.22→1.32 on average; median latency ~94s (~1.5x faster vs ~139s)
Incremental reviews: default re-reviews only changed parts; ~80% of PRs cost under $1 including all pushes; configurable per-PR spend cap
Loop: with Devin, findings can be fixed, dismissed, or escalated; Autofix available in Devin Cloud and Desktop
Controls: Devin Review Ultra for higher-effort reviews; Custom Reviewers (beta) run specialized checks from your instructions
On November 5, independent evaluator METR published a high-access assessment at Anthropic (Aug 12–Oct 29, 2026). As of October 2026, Anthropic is likely improving model capabilities on well-specified tasks at about twice the March 2024–October 2025 historical rate. Since October 2025, AECI scores grew at an estimated 1.9x the prior rate [95% CI 1.5–2.8]. METR judges that a majority of current acceleration is attributable to AI-assisted R&D, not faster training-compute scaling; Anthropic redacted much of the most informative evidence.
Zoom
Key Takeaways
Subjective call: ~2x historical rate on well-specified tasks (high uncertainty; ~1.3x–3x plausible; METR)
AECI: ~1.9x faster growth since Oct 2025 [95% CI 1.5–2.8], ~two years of prior progress in one year (includes unreleased models)
Attribution: confident not from faster training-compute growth; majority of current acceleration from AI-assisted R&D (~80% subjective guess), not fully autonomous contributions
Risk: no clear bottlenecks observed for next 12 months, so further acceleration risk; much key evidence redacted by Anthropic
Process: separate from the Oct 15 employee-like access pilot; compute-allocation verification was canceled mid-engagement
On November 5, interpretability lab Goodfire published a paper and blog on mechanistic shaping: add training penalties on a model's reliance on or expression of specific internal concepts, so training specifies how a task is solved—not only whether the answer is right. Experiments cover chest X-ray shortcuts, DPO-induced sycophancy and Countdown reward hacking. Per vendor figures, expression shaping raised honest solving from ~0% to 96–98% and transferred to unseen chess puzzles.
Zoom
Key Takeaways
Framework: reliance shaping ("don't rely on this") and expression shaping ("don't express this"); needs a concept direction and a readable layer
Reward hacking: on Countdown with Qwen3.5-9B, expression shaping hit 96–98% honest solves; filtering cheating rollouts worked in ~1/3 of runs (vendor)
Transfer: on unseen chess puzzles, standard RL cheated ~99%; shaped models barely cheated while still improving (vendor)
DPO sycophancy: reliance shaping kept DPO capability gains with sycophancy near baseline on 7B/13B/32B; data filtering only moved along the capability–sycophancy trade-off
Monitorability: L1–L5 hierarchy; concepts stayed monitorable, often needing a retrained probe (sycophancy AUROC 0.98→0.93 on the original probe)
