Google’s new model needs a careful price comparison

  • Google releases Gemini 4 Neon for coding and document work — Developers get another capable hosted model to compare, but Carbon’s temporary discount changes which model is cheaper today.
    • November 18’s rollout includes the paid API and business service, plus Pro and Ultra access in Antigravity and Gemini. An agent is a model running a multi-step task with tools.
    • Neon accepts text, images, audio and video within a million-token context. Tokens are the pieces into which models divide text. Output is text.
    • Each million tokens costs $3 for fresh input, $0.15 for reused input and $15 for output. Four reasoning-effort settings run from low through xhigh.
    • A benchmark is a fixed set of tasks for comparing models. Artificial Analysis scores Neon xhigh at 61, one point above Carbon high on its Intelligence Index.
    • Its measured cost per task is $6.44 versus Carbon’s $7.76, using Carbon’s standard prices. Average completion time is about 21 versus 36 minutes.
    • Carbon currently charges $2 for input and $10 for output per million tokens, through November 30. Neon’s corresponding token prices are higher during that promotion.
    • Caveat: The independent comparison uses particular workloads and different reasoning settings, so its savings are neither a universal speedup nor today’s discounted-price comparison. (HN 866 points · 685 comments · Techmeme · release, model card and outside measurements)

Continue local coding sessions from another screen

  • Morty brings local coding conversations into a browser — You can check work from a phone while execution stays on your computer, with substantial permissions to review before pairing.
    • The Apache-2.0 repository supports Codex, Claude Code and Pi, with no documented Hermes integration. The computer must remain awake and online. History stays there.
    • The project documents encrypted session payloads between browser and computer. Its relay still sees routing metadata, and the website sends error and performance diagnostics to Sentry.
    • A sharing link works once on one device within ten minutes. Access is time-limited and revocable, but the website and repository currently disagree about its default duration.
    • Exit the native terminal session before continuing in Morty. Otherwise, Claude Code and Pi can fork the conversation. Codex allows only one writer at a time.
    • The security policy says a compromised service could deliver browser code that reads the encryption secret. Disconnecting one device also leaves that shared secret unchanged.
    • Caveat: Codex sessions start with full computer access, without approval prompts or a sandbox, unless another mode is chosen. (Lobsters · repository, setup documentation and explicit security boundaries)

Recover actual changes, then measure the cost of delegation

  • A crash-recovery study separates duplicate prevention from knowing what happened paper notes → — Record intended actions before dispatch, then check the destination after an interruption instead of interpreting silence as success.

    • Song Xu tests ten kinds of faults in one controlled runtime. Each platform runs 7,513 trials, reusing 96 captured model responses from 24 constructed cases.
    • The full setup combines authorization, a durable action ledger, destination-side duplicate suppression and checks against the destination after failure. Windows and Linux reuse the same inputs.
    • In 200 known crash-and-replay comparisons per platform, the full setup produces zero repeated effects. Each simpler baseline produces repeated effects in 117 comparisons.
    • Disabling the destination checks leaves all 100 known post-effect crash outcomes ambiguous, despite preventing duplicates. Duplicate prevention alone therefore fails to establish whether work finished.
    • Full safeguards run at roughly one-seventh of the permissive baseline’s measured operation rate. Evidence-write faults leave 75 of 100 trials unresolved on each platform.
    • Caveat: This is one team’s controlled implementation, and its code, raw records and aggregate analysis tables are unavailable publicly. (arXiv 2611.23353 · submitted November 17 · full methods, comparisons and limitations)
  • Trying a small coding model first saves compute but takes longer paper notes → — A simple fallback rule is worth testing for background work. Your correctness checks must establish when the smaller model actually succeeded.

    • AMD and NTU researchers compare 36 downloadable models on all 500 SWE-bench Verified tasks. They run each model–task pair once using the same surrounding software.
    • Their fallback simulation uses the 466 tasks with complete records for both GLM-4.7-Flash and GLM-4.7. The larger model restarts from the original repository state.
    • With a 120-minute first-stage limit, calculated GPU-hour cost falls 16.4%. Task success rises from 71.2% to 73.4%, although single runs leave that improvement uncertain.
    • Average sequential completion time rises from 41.6 to 84.5 minutes. Cost charges each task for its full allocated GPU time, without sharing discounts for concurrent work.
    • The simulated rule consults the benchmark’s acceptance tests before deciding whether to escalate. A real deployment needs an equally dependable success check.
    • Caveat: These are combined recorded runs, not a live routing deployment, and the GPU-hour accounting is distinct from actual cluster consumption or provider bills. (arXiv 2611.24324 · submitted November 18 · full experimental and simulation methods)

A small debugging method with an inspectable implementation

  • A reverse-engineering walkthrough turns debugging guesses into recorded experiments — A reusable instruction file offers a simple way to stop assistants repeating failed approaches after their conversation has been shortened. Standalone notes →
    • The author’s new walkthrough describes inspecting unfamiliar programs, saving findings and using targeted logging or probes when static inspection stalls.
    • The linked Hypothesis Debugging skill records one explanation, one test and its conclusion in HYPOTHESES.csv. A skill is a reusable package of instructions and reference material.
    • Its actual instruction file requires recording the explanation before testing and retaining inconclusive results. The process can then resume from evidence saved on disk.
    • The skill dates to July. A November 18 commit adds its MIT license. The new walkthrough shares an existing method rather than announcing a newly invented debugging technique.
    • Caveat: The author’s productivity claims come from personal experience, without a controlled comparison showing how much this procedure improves results. (HN 236 points · 60 comments · walkthrough, instruction file and commit history)

Make interactive work, then test what players experience

  • Snap lets existing coding assistants build and publish camera effects — Game developers can bring browser games and interactive camera projects into Snapchat through their existing coding workflow.

    • Easy Lens CLI, a command-line tool, launched November 18 for Mac, Windows and Linux. Snap says it is available now and free with a Snapchat account.
    • Claude Code, Codex and Cursor can build effects, render preview videos, send versions to a phone and submit them for publication.
    • Snap documents more than 100 augmented-reality building blocks and more than 50 AI asset generators. Augmented reality places digital objects or effects over a camera view.
    • Existing browser games can become Web Lenses with leaderboards. Aisha Malik’s report confirms the launch and its cross-platform support.
    • Caveat: Generated projects still follow Snapchat’s content checks, Lens review and publishing process, rather than bypassing platform approval. (Techmeme · TechCrunch · Snap’s release and installation landing page)
  • A game-building study checks working code against what players prefer paper notes → — Passing functional tests can hide awkward controls and weak presentation, so generated games need separate checks of the actual playing experience.

    • 3D GameDev Arena introduces 108 tasks across 14 genres, starting from an empty workspace with written requirements and reference images. Four model-and-software configurations build the games.
    • Evaluation separates requirement tests, model-scored recordings of play and blind human comparisons. Some recorded situations require assisted state changes, which do not show that ordinary play reaches them.
    • GPT-6.1 Sol in Codex earns the highest functional score, 64.7 out of 100. GPT-6 Astra and Kimi K3 score 59.4 and 59.8 respectively.
    • Astra’s fitted human-preference score is 69.24 versus Kimi’s 25.86. These estimate win percentages against the average competing configuration, rather than the percentage of requirements satisfied.
    • The paper’s printed arena URL returns 404. The non-www site serves a redirect, so the direct paper remains the checked reading link.
    • Caveat: Each model runs in just one software setup, so these comparisons cannot isolate model quality or prove that more live testing caused better games. (arXiv 2611.24073 · submitted November 18 · evaluation methods, results and public-link checks)
Skipped this run (14 linked items)