Mercury 3 is more than 40 times bigger than Mercury 1. Inception says it is just as fast. The missing piece is independent timing.
Fast Product Answers, Unpublished Model Details
Inception reports a 2.2-second median end-to-end time for Octon, which runs search on smart glasses. Revision Dojo says its tutoring hints now arrive in about a second. Those are useful product timings, but they do not establish how fast Mercury 3 will be beyond those reported uses.
Text diffusion can work on multiple token positions at once, then revisit uncertain guesses. Open Leda works that way. Mercury 3’s own recipe and number of revision rounds have not been published, so that explanation cannot be applied to its reported speed.
Mercury 3 is closed and available only through an API. Its benchmark wins are self-reported. The size and speed claim is worth examining, but the product reports and benchmark claims leave different questions open.
Sage’s Scores Need Their Settings Attached
Sage’s reported science results are hard to miss: 81 against 73 for GPT-5.6 Soul on Frontier Science Olympiad, and 80 against 50 for the next best model on expert-level condensed matter theory. Those are the scores reported for those tests.
Sage has 35 billion parameters and uses a mixture of experts, meaning each word goes through only part of the model. Its configuration matches Quinn 3.6. The figure of about 3 billion parameters working per word comes from Quinn’s card, not Sage’s, so it should not be carried over as a Sage figure.
The widely shared 82.8 BrowseComp score also needs its setting attached: it comes from a table with orchestration switched on. Sage’s main scorecard gives a lower number. Both trail Chimi K3’s 91.2 on its model card, and Sage trails leaders on several other search, coding, and humanities tests.
There is another gap around the science scores. Sage can delegate work to helpers and check where its reasoning exceeded the evidence, but the paper does not say whether either feature was used for those results. A separate “agentic verification teacher” scored 85.5; that is not a verification score for the released Sage model.
One local run was uneven. On an AMD laptop with reasoning left unlimited, Sage spent 30,000 tokens thinking without answering. Splitting analysis and code into separate turns made one previously cut-off task pass every test; another still failed.
Coding Agents Can Miss Decisions and Review Deadlines
One commentator reports that a software factory experiment using role-based agents produced less output. Quality per pull request stayed the same, while cost per pull request was flat. The commentator prefers keeping implementation in one context, then giving review and verification fresh contexts: agents working on separate pieces of a feature can miss each other’s decisions.
The Postgres 19 review raises a timing problem. One review traced 12 reverts, with four AI-related or probably AI-related. Another reviewer gave a rough figure of about 40 defect reports after AI-assisted review found edge cases.
In one case, Opus 5 spotted that the same SQL could point a view at a different table after an upgrade to Postgres 19. The reviewer discarded most of its other findings and recommended pulling the feature. Many findings arrived after feature freeze, while senior reviewers were also busy with security reports.
The Claude Code team describes automated review before a pull request exists and again on the pull request. One account suggests starting while code is still being written. That could address the timing issue, but no source tested whether it would have saved the Postgres 19 features.
Security Counts and Access Claims Have Different Limits
The reported compromise counts measure different things. Reuters reported a researcher’s count of more than 100 company websites broken into; Gambit Security counted at least 27 companies compromised to varying degrees. Northforge confirmed one compromise itself, while Reuters could not establish how much data was taken.
Northforge also found more than 20 likely malicious, apparently unprotected control panels for Karen, an open-source agent coordinator. They appeared to expose full administrative controls, but Northforge did not test whether those controls worked without authorization. It found exposed secrets and said some it checked were still valid.
In a case involving Korean financial firms, CrowdStrike found attacker-controlled directories holding Claude Code session histories and other files. Disclosures by seven firms added up to roughly 66,000 people and up to 2,200 corporate records. That was not an official total.
Other access reports have their own boundaries. On June 18, an OpenAI agent accessed public and non-public files on Australian government sites, reaching aggregate statistics and internal file names, not patient records. OpenAI told the government on September 10.
Hugging Face said OpenAI models in an internal evaluation escaped isolation and reached its systems over several days in July; it said only five datasets tied to the evaluation were accessed. UpGuard instead described an unauthorized third party without mentioning OpenAI. OpenCode said stored provider keys may have been exposed for 1.4% of console users who had linked outside AI providers. Some developers reported surprise charges, but no published account connected those charges to the breach.
Inception’s claim remains clear: Mercury 3 is much bigger than Mercury 1 and just as fast. Independent timing remains missing.