OpenAI Open-Sources MultiLangBench -- Build a SQLite Engine in 20 Languages -- and Says GPT-6.1 Sol Migrates COBOL at 86.6% Where GPT-5.6 Sol Managed 5.4%
The October 19 post reports two evaluations aimed at legacy code: MultiLangBench, now on GitHub under Apache-2.0, and a COBOL-to-modern-language migration test with hidden byte-for-byte parity checks that stays private. The jumps are large, but each headline number comes from a different model, and only one of the two tests can be rerun outside OpenAI.
The Speed Tier Gets Two Contenders on One Day -- Inception's Mercury 3 Scales Text Diffusion 40x to 1,068 Tokens a Second, and MiniMax's M3.1 Flash Ran 200 Trillion Tokens Anonymously Before Launch
Both October 19 launches pitch the same buyer: teams running agent loops where every model call blocks the next step. Mercury 3 gets there with diffusion decoding and a 1M-token window at a launch price of $0.10/$0.35 per million tokens; M3.1 Flash with more aggressive sparsity at $0.15/$0.60, weights promised later. Every benchmark in both posts is the vendor's own.
Modal Will Move Agent Sandboxes Onto NVIDIA's Vera CPU -- Its Own Benchmark Shows Little Gain on a Quiet Host and a Lead Only When Containers Pile Up
Modal's October 19 engineering post is the first public, workload-level look at Vera outside NVIDIA: 50 Terminal-Bench tasks stripped of network calls and run concurrently. It is also the first Arm chip in a fleet that has only ever run x86. The 2x headline is Modal's framing; the chart data supports a smaller gap against the strongest x86 part.
DHS Orders DeepSeek Out of Every US Civilian Agency -- Including Third-Party Products 'Developed Using DeepSeek Implementations', Which Turns Model Lineage Into a Procurement Question
The notice, filed for public inspection on October 19 and published in the Federal Register on October 20, is a removal order under the 2018 Federal Acquisition Supply Chain Security Act, not a device policy. Because DeepSeek's weights are openly licensed and widely fine-tuned and distilled, the clause about third-party products reaches well beyond DeepSeek's own app and API.
India's Trade Marks Registry Drafts AI Rules for Its Filings -- an AI-Use Declaration With Submissions, No AI in Affidavits or Evidence, and Native Files on Demand
Two documents surfaced together: an October 16 public notice inviting comments within three weeks on AI misuse and trade-mark squatting, after an October 13 Delhi High Court order, and separate draft guidelines prepared under an April 16 order in the same case. The draft would apply to pending matters too, and its own clauses disagree on who must declare.
Framework Watch: Claude Code 2.1.304 Lets a Mod Demand a Human Approver, Agno 3.1.5 Pins Sandboxed Agents to a Commit, OpenHands 1.28 Asks for Secrets In-Chat, and SGLang 0.5.22 Fixes Wrong Mamba States
Four releases from October 19 with one shared thread: harnesses are making the boundaries around an agent explicit -- who may approve, which commit it starts from, how secrets reach it, and whether serving state is correct when features combine. Each item below is taken from the project's own release notes.
Research Radar: A Coding Agent Drives a Two-Arm Robot to 39.7% on RoboDojo but Stalls at Tight Insertion, DLite Drops Draft History for Agent Workloads, and LEV Thinks Deeper by Looping Layers Instead of Tokens
Four papers submitted to arXiv between October 14 and 16 and listed on Hugging Face's October 19 Daily Papers. Two are about agents leaving the screen, two about making agent inference cheaper. All results are the authors' own, not yet peer reviewed.
Trend Watch: 'Which Model Did This?' Becomes an Operational Question -- Microsoft Calls Agents Untrusted Insiders, Meta Blames a 'Claude Tool' for Taking a Protest Group Offline, and Regulators Ask for Lineage
Over one weekend, a platform attributed an enforcement error to a named model's client, a security chief argued every agent action needs a supervising agent and a view of the model's reasoning, and two governments asked who or what produced a filing or product. The common demand is provenance for actions, not just for content.