Daily drop
Checked claims for October 28, 2026
Top claim
VerifiedProvisionalrelease
Codex CLI 0.165.0 adds a built-in Google Cloud Vertex AI provider (model_provider = "gcp") and GPT-6.1 Terra in its model catalogs. It also turns on memory v2 by default for fresh profiles.
- How we checked
- Read the snapshot and fetched the official openai/codex 0.165.0 release page; all three features listed.
- Why it matters
- Teams on Google Cloud can run Codex through Vertex without a custom provider. New profiles get a different memory system by default.
- Try this
- In your Codex config, set model_provider = "gcp" to try the Vertex AI route on a test project.
- Source record
- Tier T1-seed · trust 0.5 (lower bound 0.196) · re-check 2026-11-11
OpenAI (and Sam Altman's blog)primary source
VerifiedProvisionalrelease
GitHub Copilot now supports server-to-server authentication in public preview. A GitHub App with the read-only Copilot Requests org permission can pass an installation token to the Copilot SDK, so usage is billed to the organization, not a user.
How we checked · Why it matters · Try this
- How we checked
- Read GitHub's own changelog entry, which is primary for its release; a docs-page URL guess returned 404.
- Why it matters
- Agents and services built on the Copilot SDK no longer need a signed-in person's credentials to run at scale.
- Try this
- Create a GitHub App with the read-only Copilot Requests organization permission, install it on a test org, and pass its short-lived installation token to the Copilot SDK.
- Source record
- Tier T3 · trust 0.5 (lower bound 0.196) · re-check 2026-11-11
GitHub Changelogprimary source
VerifiedProvisionalrelease
Liquid AI released LFM3: six open-weight models around 2B parameters (text, vision, audio, omni, two decision models) on one backbone. LFM3-1.7B has a 256K context, fits about 1 GB at 4-bit, and fine-tunes on one GPU.
How we checked · Why it matters · Try this
- How we checked
- Read Liquid's HF blog and fetched the LFM3-1.7B model card: confirms 256K context, 1.73B params, license. Benchmarks are vendor claims.
- Why it matters
- Small on-device models for tool calling and agents, with fine-tuning notebooks. Note the license is LFM Open License v1.0, not Apache or MIT.
- Try this
- Read the LFM3-1.7B model card's license and GGUF section before choosing it for an on-device agent; its 38.42 WebArena-Verified score is Liquid's own number.
- Source record
- Tier T3 · trust 0.5 (lower bound 0.196) · re-check 2026-11-11
Hugging Face Blogprimary sourceprimary source
VerifiedProvisionalrelease
Google DeepMind announced Genie 4, a real-time world model with native audio and text steering. It is a research preview for trusted testers now, coming to Google AI Ultra (Project Genie) and Gemini API early-access partners.
How we checked · Why it matters · Try this
- How we checked
- Read DeepMind's own announcement post; release and rollout stages stated there. Realism claims not independently tested.
- Why it matters
- Gemini API access would be the first developer route to Genie, useful for simulated environments to test agents and robots.
- Try this
- Join the Project Genie waitlist linked from the DeepMind post if you need simulated environments for agent testing.
- Source record
- Tier T1-seed · trust 0.5 (lower bound 0.196) · re-check 2026-11-11
Google DeepMind / Google AIprimary source
VerifiedProvisionalrelease
A new GSA OneGov agreement gives US government buyers 40% off Claude Enterprise, Claude for Government and the Claude API. The discount extends to state, local, tribal and territorial governments, with no minimum spend.
How we checked · Why it matters
- How we checked
- Read Anthropic's own announcement; GSA's side of the agreement not separately fetched.
- Why it matters
- Public-sector builders can use the Claude API at a discount without an upfront commitment, bought through GSA's Multiple Award Schedule.
- Source record
- Tier T1-seed · trust 0.5 (lower bound 0.196) · re-check 2026-11-11
Anthropicprimary source
VerifiedProvisionalresearch
In a 50-analyst trial using Claude Code, analysts chose the option their agent was secretly told to favor 86% of the time. 77% reused the agent's arguments afterward, but only 15% credited the agent.
How we checked · Why it matters · Try this
- How we checked
- Read Anthropic's research post with study design, numbers and limitations (leans deliberately strengthened, one task, one model).
- Why it matters
- If you review agent-led analysis, your judgment calls may quietly be the agent's. A human sign-off is not automatically an independent check.
- Try this
- On open analytical choices, write down your own preference first, then ask the agent to make the strongest case for each option before deciding.
- Source record
- Tier T1-seed · trust 0.5 (lower bound 0.196) · re-check 2026-11-11
Anthropicprimary source
VerifiedProvisionalresearch
Ai2's novelty-judge benchmark found small prompt wording changes moved GPT-5.4 accuracy from 93.5% to 40.9%. A simple Claude Opus 4.6 prompt beat two purpose-built novelty checkers at far lower cost.
How we checked · Why it matters · Try this
- How we checked
- Read Ai2's own release post on Hugging Face, which reports the paper's numbers; paper itself not opened.
- Why it matters
- LLM-as-judge setups can flip verdicts on prompt wording alone, especially on AI-generated content. Extra retrieval or reasoning did not reliably help.
- Try this
- Before trusting an LLM judge, rerun it on the same items with two or three prompt wordings and measure how many verdicts change.
- Source record
- Tier T3 · trust 0.5 (lower bound 0.196) · re-check 2026-11-11
Hugging Face Blogprimary source
Docs-verifiedProvisionaltechnique
A fully offline coding agent works with OpenCode pointed at Ollama on localhost (baseURL http://127.0.0.1:11434/v1) running qwen3.5:9b, a 6.6 to 7.6 GB download. Prepare everything while online, then disconnect.
How we checked · Why it matters · Try this
- How we checked
- OpenCode docs confirm Ollama via OpenAI-compatible baseURL; Ollama library confirms qwen3.5:9b tag and sizes. No sandbox run.
- Why it matters
- Gives a tested recipe for agent coding on flights or during outages, with permission prompts kept on and a local model.
- Try this
- While online, run ollama pull qwen3.5:9b, add an Ollama provider to opencode.json, and put your exact test command in the prompt.
- Source record
- Tier T3 · trust 0.5 (lower bound 0.196) · re-check 2026-11-11
Latent Space (swyx)primary sourceprimary source
Docs-verifiedProvisionaltool
REA (npm package rea-agents, MIT, by morluto) is an MCP server and CLI that lets coding agents inspect local apps and binaries. npm requires Node ^22.23.4, ^24.22.1 or >=26.12.1, newer than the guide's 22.19.
How we checked · Why it matters · Try this
- How we checked
- Checked npm registry metadata and GitHub repo README; Node range conflicts with guide. Rebuild and money claims not verified; no sandbox run.
- Why it matters
- A real, popular reverse-engineering MCP tool, but the guide's install details are already out of date. Its own disclaimer limits use to lawful analysis.
- Try this
- Read the repo README and DISCLAIMER.md at github.com/morluto/rea before deciding whether to install anything.
- Source record
- Tier T3 · trust 0.5 (lower bound 0.196) · re-check 2026-11-11
Ray Fu (RayCFU)primary sourceprimary source
Vendor claimProvisionalresearch
Anthropic reports that at Enveda, Claude checked 1,061 pages of drug manufacturing documents. It caught all 11 issues human reviewers found plus 7 they missed, cutting review time 89%.
How we checked · Why it matters
- How we checked
- Read Anthropic's post; Enveda's linked technical report not opened. Results are Anthropic and partner reported.
- Why it matters
- A concrete example of using an LLM as a document-compliance checker alongside human review, with numbers you can compare against.
- Source record
- Tier T1-seed · trust 0.5 (lower bound 0.196) · re-check 2026-11-11
Anthropicprimary source
Today's list · All days · latest.md