Subscription limits

  • OpenAI plans to combine ChatGPT and Codex usage allowances — People using both products will need to budget their work against one allowance.
    • Starting in December, the shared allowance applies to Plus, Pro and Business plans. Instant chat will not draw from it.
    • Usage will be measured in tokens, the units of information a model processes and generates. The model and task affect how much of the allowance is consumed.
    • OpenAI says Settings → Usage will show the combined remaining allowance and reset times. After reaching a limit, users can wait, upgrade or buy credits for eligible features.
    • Caveat: The help page does not state the combined token allowance or give a specific December start date. (HN 162 points · 97 comments · OpenAI help page)

Coding tools and tests

  • DeepSeek releases a prototype coding assistant for teams — The project offers a concrete design for handling repository work while limiting unsolicited interruptions.

    • The prototype listens to one GitHub repository and optionally Feishu, a workplace chat service. An agent is a model running a multi-step task with tools.
    • Requests from allowed people open separate work sessions and directories. The assistant can reply in the original conversation with a result such as a pull request.
    • The design guide says unsolicited work requires a team-written rule and starts by recording what it would have done.
    • Caveat: DeepSeek calls this research code, not a production-ready product; the published prototype differs from its internal implementation. (Techmeme · DeepSeek repository and design guide)

  • A coding workflow tests several small-model roles on programming problems — The example shows how separate planning, solving and checking steps can be assembled without claiming a universal model victory.

    • The workflow uses Qwen3.8-27B for planning, two independent proposals, solving and review, with a repair step. Files pass information between fresh sessions; a script checks sample tests.
    • Across five passes over ten difficult programming problems, the author reports an average of 9.8 solved by this workflow and 9.6 by the original authors’ rebuilt system. A separate single-call Fable 5 baseline averaged 8.0.
    • The author reports a median 40 minutes and $0.39 per run; all 50 runs cost $35.79. These tests used hosted OpenRouter, although the workflow also documents local connections.
    • Caveat: Ten problems are too few to distinguish the two Qwen workflows reliably, and the comparison uses a different single-call setup for Fable 5. (HN 267 points · 123 comments · workflow and published run records)

  • A developer compares the cost of AI-built web frameworks — The experiment makes framework claims measurable, with enough variation to discourage a rewrite based on a single ranking.

    • Nolan Lawson asked Claude Opus 5.5 to build the same Pokémon app across twelve setups and tested the output against a specification. An agent is a model running a multi-step task with tools.
    • The median estimated API cost was $3.14 per Next.js build and $2.67 for Hono with htmx. The author used a subscription, so those figures are estimates rather than his actual bill.
    • Most frameworks had five attempts. Angular had three successful attempts after two failed to meet the reverse-proxy requirement.
    • The author also measured browser performance and accessibility, and notes that the chosen page-load metrics favor some application designs. Individual runs varied substantially.
    • Caveat: One small, mostly content-oriented app and one model cannot establish the best framework for a larger interactive product. (HN 110 points · 59 comments · author’s methods and results)
Gathered items not included (5)