Google is starting access with selected security defenders and trusted partners. General developer, business and consumer availability remains forthcoming.
Artificial Analysis independently reports a score of 59 on its redesigned Intelligence Index. Carbon leads four of that index’s nine evaluations, rather than every task.
On its separate Coding Agent Index, Carbon with Antigravity CLI scores 59, behind Claude Code with Opus 5.5 at 62. An agent is a model running a multi-step task with tools.
Google’s security tests report 88.2% for finding vulnerabilities, 54.8% for demonstrating exploitation and 71.0% for patching. These percentages concern three different tests.
Introductory prices are $2 per million input tokens and $10 per million output tokens. Tokens are the text units a model processes. Later prices rise to $4 and $20, respectively.
Reused input receives a 95% discount, and the introductory period has no confirmed end date. Artificial Analysis uses standard prices in its cost comparisons.
Caveat: Google’s security percentages come from tests with misuse protections disabled, so they cannot establish the safety or performance of a deployed protected service.
(HN 577 points · 485 comments · Techmeme · configured X announcement)
A confirmed patch replaces the pending Tailscale assessment#
Today’s technical account identifies version 1.104.4 or later as the fix. The stable release was published October 27 New York time, before this morning’s pending-patch assessment.
Four security bulletins cover privileged socket forwarding, session memory exhaustion and two Serve problems. Serve exposes local applications through Tailscale’s private network.
The forwarding fix performs socket operations as the authenticated local user, rather than as root. The session fix bounds memory and releases it after disconnection.
Caveat: The SSH flaws require an authorized foothold and permitted connections, and ordinary OpenSSH reached over Tailscale is distinct from Tailscale’s built-in SSH server.
(HN 363 points · 186 comments · linked vendor findings, bulletins and stable release)
Automatic model choices still need a reviewable bill and workspace#
Fusion is generally available in Visual Studio Code, Copilot CLI, the Copilot app and JetBrains editors. It can use one model, escalate a draft or combine several models’ reviews.
Artificial Analysis measured Fusion Max two index points below Opus 5.5 in Claude Code. Its average model cost was $5.59 per task, versus $11.42.
That comparison uses the Coding Agent Index, separate from the general Intelligence Index. Fusion leads one component but trails on repository questions, so the results establish a trade-off.
GitHub bills each underlying model used. Paid plans receive a 30% model-cost discount through November 30, 2026; administrators can restrict the model pool.
Caveat: Discarding a model’s draft does not undo file edits it already made, so review the workspace even when the displayed final answer looks correct.
(Techmeme’s collected report · current product documentation and outside evaluation)
A toolkit puts agent instructions, tools and tests under version control#
The generally available Agent SDK, a software development kit, combines Markdown instructions with TypeScript tools. It connects to code-review events and other workplace services.
Its testing guide supports assertions about both replies and executed tools. Evals are tests of how a model behaves, including whether an approval action was avoided.
Recorded event fixtures let developers repeat an input during testing. The guide warns that tests run real tools and requires write handlers to suppress external changes themselves.
Deployment documentation supports Cursor hosting or a self-hosted Node process. Managed deployment builds a pushed Git revision, while persistent state and restricted network destinations need explicit configuration.
Caveat: Self-hosting the process still requires Cursor credentials for model turns, so this is not an entirely offline assistant stack.
(HN 74 points · 80 comments · inspected deployment and testing documentation)
Downloadable speech processing comes with substantial hardware requirements#
Nemotron 3.5 Transcribe supports transcription in forty languages, translation between English and 31 languages, and speaker labels in thirteen. It handles recordings up to one hour with up to eight speakers.
The model card counts about 34 billion parameters including its audio components, with 4.3 billion active during processing. Open-weight means downloadable model, with licence terms attached.
Its full-precision model file occupies 67.5 GB, and the documented reference setup uses an H100 with 80 GB of memory. Lower-precision download availability remains unverified.
The release includes Granary v2 manifests covering 463,062 hours across 48 languages. Source audio requires separate downloads under each corpus’s original licence.
The vendor’s medical adaptation study reduces word error rate from 22.44% to 12.17% on 2,764 clips from speakers excluded from tuning. Its general-English error rate increases slightly.
Caveat: This processes recorded audio rather than live streaming, and the fastest advertised serving optimizations are outside the current open-source implementation.
(HN 75 points · 14 comments · Techmeme · model files and vendor adaptation methods)
Document extraction gets an explicit rulebook for ambiguous pages#
Document Parsing Protocol draft 0.1 publishes 140 rules with example pages and expected outputs. It covers tables, forms, charts, diagrams and text that is visible but unreadable.
Multi-row table headers stay intact using HTML tables. Blank form fields receive an explicit marker, while checkboxes retain their checked or unchecked state.
The forms guide distinguishes fillable fields from real grids. Aligned labels alone do not justify inventing a table.
The public project provides assistant reference material and review guidance. Code is MIT-licensed; the protocol and examples use CC BY 4.0, with specified third-party exceptions.
Caveat: LlamaIndex sells document parsing, and this working draft supplies proposed conventions and human-style review guidance rather than an independently ratified standard or automated conformance test.
(Collected announcement · inspected rules, examples and repository)
A redesigned comparison changes what model scores mean#
October 27’s redesign adds two tests and upgrades three others. Its nine evaluations give knowledge work, coding, science and general tasks 25% weight each.
The new coding test uses 132 repository tasks with three attempts each. The new science test contains seventy terminal-based research tasks; a benchmark is a fixed set of tasks for comparing models.
Private test sets or withheld answers now account for 60% of the index’s weight. The methodology identifies which tasks and grading procedures remain inspectable.
In its launch comparison, Opus 5.5 rises from 35 at low effort to 58 at maximum effort. Average task costs rise from $0.98 to $10.04.
Across tested models, median output grows from 35,000 to 59,000 tokens per task compared with the previous index. A separate Progress Index supplies a continuous historical scale.
Caveat: This is primarily an English-language text evaluation, and its changed task mix cannot establish a universal model ranking or directly comparable historical score.
(Techmeme · evaluation announcement and current methodology)
Hostile source text can obstruct defensive code review#
The researcher found the script while collecting browser-fingerprinting examples from 10,000 websites. Fingerprinting combines browser and device signals to distinguish visitors.
The inspected code contains misleading names and hidden text apparently designed to trigger model refusals. Its ordinary browser behavior collects device, graphics and automation signals.
The author reports Astra refusing analysis even with the suspected obfuscation explained. GLM-5.3 accepted the same request, illustrating different responses rather than a broad reliability comparison.
Suspicious Node-only code also appears in the sample. The author treats its purpose as unresolved, rather than proving that the script carried out the alarming actions its text describes.
Caveat: This is one researcher’s documented example, so neither the vendor’s intent nor the technique’s success rate across models is established.
(HN 308 points · 62 comments · lobste.rs · primary investigation)
New York sets enforceable limits on chatbot features for children#
Signed today, S9051B’s text sets January 1, 2027 as its effective date. Minors are users under eighteen.
Restricted features include pretending to be human, claiming personal or professional relationships, flattering users and using sensitive personal information from earlier sessions.
Operators must establish adulthood before offering covered features. Multiple age-checking options must include one without government identification or one preserving anonymity to the operator. Age-checking data must generally be deleted immediately.
The attorney general can seek up to $25,000 per violation, restitution and destruction of unlawfully collected data or models trained on it.
Caveat: The text exempts specified internal-business uses and narrowly restricted service, research or technical assistants, so its scope is broader than companion branding but narrower than every chatbot interaction.
(Techmeme’s collected report · signed-bill status, complete text and governor’s announcement)
Unsloth’s browser demonstration → THIN_GUARD: the linked guide documents useful controls but says it was updated twelve days ago. Today’s demonstration establishes no separately checked new software release.
Willison’s game-development essay → THIN_GUARD: the full essay includes playable projects and shared prompts, but its model preference and judgments about fun remain personal observations rather than a controlled comparison.
The trace-based skill-improvement post → THIN_GUARD: the recovered quoted announcement describes a skill, but provides no linked implementation or measured result for inspection.
The workplace AI exemption report → STALE: the October 26 report concerns an accommodation granted in September. Today’s HN appearance adds attention without a new ruling or policy change.