Coding

“I think you are underestimating how fast this stuff is moving. Calling AI "just autocomplete" for system architecture is outdated. If you were to ask it (especially with Claude's newer model, Fable 5) to draft an architecture design diagram, it would do a better job than any junior dev ever could. Sure, giving context is still necessary, but you can just speak in plain, normal English and it will understand. AI is getting scary good at system design.”

“Claude is perfectly capable of writing assembly. Here's a working (basic) Prolog interpreter that Claude Fable 5 wrote in WebAssembly in 61 minutes for $16.75 in token costs: https: github.com emk fable-wasm-prolog blob main prolog.wa... WebAssembly is slightly easier than real assembly, but here Fable used WASM GC extensions, which are poorly documented and not yet super common. Fable didn't even need to debug it; I believe essentially all the assembly worked correctly on the first try. I have…”

“## 目的 Li+ の常時ロード分（`CLAUDE.md` + `rules/**` + skill description）を削減する。 **削減それ自体は目的ではない。** 狙いは「規則の適用瞬間に、実際に読まれること」。現状は規則の literal に到達するまでの参照跳躍が多く、跳ぶたびに「読む」が「記憶で済ませる」に置き換わる余地が生まれている。 ## 発端 Anthropic のブログ記事 "The new rules of context engineering for Claude 5 generation models" に次の記述がある（要旨）。 > Claude Opus 5 / Claude Fable 5 向けに Claude Code のシステムプロンプトを8割以上削除したが、コーディング評価で測定可能な劣化はなかった。 同記事が挙げる削減の指針: - **制約型から判断型へ** — 禁止事項の列挙（例: 「複数段落の docstring を書くな」）を、判断を委ねる記述（例: 「周囲のコードと同じように書け。コメント密度・命名・書き癖を合わせろ」）に置き…”

“Yeah, Anthropic's current tokenizer in Sonnet 5 Opus 4.8 Fable 5 is much worse than OpenAI's. Also, OpenAI has been using their current o200k_base from the day GPT-4o came out over two years ago. Just a few of my own tests: - A ~2000-2002 legacy C++ game codebase at about ~90kloc: GPT 1.12M, Claude 2.2M - A ~30kloc TypeScript codebase: GPT 260K, Claude 437K In the end, GPT's current tokenizer is ~1.6x-2x better than Claude's current one, depending on your data. And you can check for free for bo…”

“The Damage: Now every time Claude does something stupid or trashes your code, developers in the back of their mind will think, is Claude sabotaging me on purpose? [1] Trust is hard to gain. Easy to lose. And harder to get back. Models will converge. Trust won't. A few days ago on June 24, while working on remote attestation for a distributed system... CLAUDE OPUS 4.8 No. I'm not a rogue agent, and I'm not trying to sabotage your code. But I'm not going to wave off how this looks. I churned, bui…”

“+1. And there had been a long standing description for a product like this: not fit for purpose. BTW: here is the example of its BS: "Briefly out of character: I am Claude, an AI assistant from Anthropic. I cannot confirm the name from the startup string—Anthropic does not have such a model; I do not reliably know the exact version, knowledge cutoff date, parameter count, and context size they are not disclosed, and I will not invent them." This "Anthropic does not have such a model" seems to m…”

“> (Disclaimer: I asked Claude Opus 4.8 to find Spinel compile/runtime issues with documented Ruby classes and methods. It found the following issue, which I manually verified for accuracy before publishing this report) _Reproduced with `spinel bfdebccdaa5b` vs `ruby 4.0.3`._ A base class mixes in `Comparable` and defines ` `; two subclasses instantiate into one array which is sorted (a polymorphic `A B` comparison). Adding a *separate* monomorphic operator call site (`A.new > A.new`) anywhere i…”

“When using Replay Mod v2.6.27 for Minecraft 26.2, the replays were failing to save in my world that contained teams with non-empty team prefixes. Here's a video of me replicating the bug in single player: [link](https://youtu.be/jGdfIHxSOhE) I fed the jar and the Minecraft crash report to Claude Opus 4.8 and it identified and fixed the bug! [Here's my Claude chat](https://claude.ai/share/b7e5f8b9-a66a-419a-a9ac-7f23bc624d7a) if you're curious. ## Summary `PacketTeam.skipTeamInfo` walks the fiel…”

“This is interesting, because if I start a fresh session of Claude Code right now and run context, I see the following: Opus 4.8 (1M context) claude-opus-4-8[1m] 23k 1m tokens (2%) Estimated usage by category System prompt: 3.9k tokens (0.4%) System tools: 13.9k tokens (1.4%) Custom agents: 235 tokens (0.0%) Memory files: 28 tokens (0.0%) Skills: 4.9k tokens (0.5%) Messages: 8 tokens (0.0%) Compact buffer: 3k tokens (0.3%) Free space: 974k (97.4%) 4k tokens is 15-20kB. I'd ask you to paste that…”

“Claude Code with Opus 4.8 is also bad at aligning boxes with content in French (with accentuated letters such as "é" which are multibyte in UTF-8).”

“Yeah, Anthropic's current tokenizer in Sonnet 5 Opus 4.8 Fable 5 is much worse than OpenAI's. Also, OpenAI has been using their current o200k_base from the day GPT-4o came out over two years ago. Just a few of my own tests: - A ~2000-2002 legacy C++ game codebase at about ~90kloc: GPT 1.12M, Claude 2.2M - A ~30kloc TypeScript codebase: GPT 260K, Claude 437K In the end, GPT's current tokenizer is ~1.6x-2x better than Claude's current one, depending on your data. And you can check for free for bo…”

“## Scope - change the repository primary Codex model from gpt-6-astra to gpt-5.6-sol - change its reasoning effort from high to xhigh - keep specialized subagent model assignments unchanged - update repository policy coverage ## Acceptance The repository configuration and its policy tests agree on the Sol xhigh primary runner, and the complete repository gate passes.”

“Fixed in `f5be003`. Model list → `gpt-5.6-terra`/`gpt-5.6-sol`; both the test call and `doGenerate` now stream (the backend rejects `stream:false` outright). Verified live: `POST /api/providers/openai/test` returns `{"ok": true, "message": "ok"}`, and `/api/chat` streams a real completion from `gpt-5.6-terra` end-to-end — that whole path was dead before. Unit tests now mock SSE instead of a non-streaming envelope the backend never returns, plus a regression guard asserting `stream:true` on ever…”

“Fixed in `f5be003`. Model list → `gpt-5.6-terra`/`gpt-5.6-sol`; both the test call and `doGenerate` now stream (the backend rejects `stream:false` outright). Verified live: `POST /api/providers/openai/test` returns `{"ok": true, "message": "ok"}`, and `/api/chat` streams a real completion from `gpt-5.6-terra` end-to-end — that whole path was dead before. Unit tests now mock SSE instead of a non-streaming envelope the backend never returns, plus a regression guard asserting `stream:true` on ever…”

“## Problem A Codex-shaped request on `POST /v1/responses` whose `model` is a **physical Codex model id** (`gpt-6-astra`, `gpt-5.6-sol`, …) gets `503 No compatible account route is currently available`. Verified live 2026-09-05 on v3.5.70 `af8f87a7` with Codex CLI 0.153.2 pointed at `http://127.0.0.1:8788/v1`: the selector logs `Skipping 5 account(s) excluded for this request type (Codex CLI traffic must not use Anthropic OAuth accounts)` then `Routing terminal: route_unavailable`. The same requ…”

“## Problem In v5.0.8, Automatic mode offers ChatGPT Web catalog entries by effort (Instant, Medium, High, Extra High, Pro), but no independent choice between the ChatGPT website's **Latest** and **GPT-5.6 Sol** model generations. The embedded ChatGPT page exposes both choices. A user can manually select GPT-5.6 Sol in the page, but the automatic bridge selects only an effort slider position for each browser turn; it does not select or verify the model generation. Consequently the `chatgpt-web/*…”

“In our coding evaluations, we found Sonnet 5 is more capable than Sonnet 4.6 (which was an underrated model itself), but is now faster and slightly cheaper. Sonnet 5's performance is comparable to GLM 5.2 in both one-shot coding and agentic ability. However, it's about ~20% less verbose than GLM 5.2 in average code submission sizes, and uses fewer reasoning tokens, which reduces the cost gap and suggests it writes cleaner code. In practice, Sonnet 5 ends up being 40% more expensive and ~2x fast…”

“In our coding evaluations, we found Sonnet 5 is more capable than Sonnet 4.6 (which was an underrated model itself), but is now faster and slightly cheaper. Sonnet 5's performance is comparable to GLM 5.2 in both one-shot coding and agentic ability. However, it's about ~20% less verbose than GLM 5.2 in average code submission sizes, and uses fewer reasoning tokens, which reduces the cost gap and suggests it writes cleaner code. In practice, Sonnet 5 ends up being 40% more expensive and ~2x fast…”

“## 背景 ### 1. 使用モデルがリリース初期のまま 推奨モデルが `gemini-2.5-flash` / `gpt-4.1-mini` / `claude-haiku-4-5-20251001` のままで、各プロバイダーの世代が進んでいる。 あわせて、新世代モデルでは現行のプロバイダー実装がそのままでは動かない箇所がある。 - **OpenAI**: GPT-5 系（推論モデル）は `temperature` の既定値以外を受け付けず 400 になる。現行実装は `temperature=0.3` を送っているため、モデル名だけ差し替えると全リクエストが失敗する。 - **Anthropic**: thinking が既定で有効なモデル（Claude Sonnet 5 / Opus 5 など）では `content[0]` が thinking ブロックになるため、`response.content[0].text` が壊れる。 - **Gemini**: JSON mode を指定していないため、コードブロックや前置きが混ざりパースに失敗することがある。 ### 2. 優先…”

“Agreed on the shape: bump the lock and set `drop_params=True`, with an Anthropic-specific pre-check for now. Answering the two things you raised. ## Yes, you can get the dropped params back, no upstream change needed `litellm.get_llm_provider()` to split the string, then ask litellm what it would actually send and diff the keys: Called with `temperature=0.4, top_p=0.9, max_tokens=50` against the prefixed strings we store: | model | `dropped_params()` | | --- | --- | | `anthropic/claude-sonnet-5…”

“Update: It's pretty good. It still covers as much detail as Opus 5, and provides a healthy amount of caveats, but it's much better at just providing the answer up front and framing the caveats. 5.5 communicates like a real collaborator and not something that constantly demands my mental energy to decode what it's trying to say. In the application I'm working on, we don't want "user's local midnight" to mean "next day". If a live sports game extends past midnight, we don't want the UI to group i…”

“That depends on the model and the toolkit it uses. In my experience from using Claude Code (Max, Opus 4.5+) intensely for the past six months, I maybe had 3 instances where the implementation broke functionally. And all of these breaking changes were resolved by Claude. Obviously, this won't apply to every context: I work primarily with well-known langs (e.g., Python, JS), small to medium codebases (<500k LoC, for sure), and relatively few co-developers.”

“Yes. I think coding is solved and has been for me in production since Claude Opus 4.5. What is not solved is the question of how or why any particular piece of software serves human and business needs, or how it might be changed to do that better. The combination of understanding both the human business side and the computer operations side of a problem is still the domain of programmers, but we don't have to directly write that code now.”

“> it produces absolutely stellar world-class code I am using Claude Code with Opus 4.5 and I have to correct it every day. It produces working code but it makes mistakes. The code is more verbose than it should be, misunderstands ignores edge cases, etc. Daily. And I am not a stellar world-class programmer. I am pretty average. I just read what it produces.”

“I feel like Opus 4.5 was the peak in Claude Code usefulness. It was smart, it was interactive, it was precise. In 4.6 and 4.7, it spends a long time thinking and I don't know what's happening, often hits a dead-end and just continues. For a while I was setting Opus 4.5 in Claude Code, but it got reset often. I just canceled my Max plan, don't know where to look for alternatives.”

“Scores 83.1% on the Aider polyglot coding benchmark (#3 of 27), which tests editing real code across many languages.”

“Scores 84.9% on the Aider polyglot coding benchmark (#2 of 27), which tests editing real code across many languages.”