#编码

Codex is sooooo good! Damn!

Building software by prompting agents without ever reading or editing the code directly isn't actually a new thing. It's just called being an engineering *manager*. We have decades of experience about how to be effective as engineering managers. It is a different skill, though.

时代变了，过去那么多的规则、约束、文档，感觉现在只会让模型的推理效率和效果变得越来越垃圾，Anthropic说的确实对。 今天决定直接大幅重构所有的文档和规则，清空重写。。。 从此思路不再是用规则约束模型指望它一次写对，而是在前期交给模型更多的思考空间更多的注意力，把更多的限制，加在测试侧，用测试来限制一切。 文档和规则，真的要退环境了。

We removed ~80% of the Claude Code system prompt for our newest models, this is what we've learned about writing system prompts, skills and Claude.MDs for them. http://x.com/i/article/2080703729385512960

Opus 5 crushed Fable 5 at 3D destruction physics for 2x cheaper! We gave four models the same task: build three self-contained HTML scenes with real physics Prompts: - A tornado that sucks in a whole field - A wrecking ball taking down an apartment block - An overloaded truck collapsing a truss bridge Outputs: - Opus 5: 55.9K tokens, $1.40 - Fable 5: 55.1K tokens, $2.82 - Kimi K3: 35.7K tokens, $0.55 - GPT 5.6: 20.1K tokens, $0.31 Opus got all three right unlike the other models. Houses fly up the funnel and out the top, the wall breaks where the ball hits and the rubble piles up, the bridge drops the truck into the river. Fable had almost nothing on the ground for the tornado to pick up, its building collapsed on its own before the ball even touched it and its bridge blew into sticks all at once. GPT is the cheapest here but its ball never reached the building at all and its bridge fell apart in a way nothing falls apart in real life. Kimi K3, the new Chinese frontier model, ended up in the same place as GPT

隆重推出 Claude Opus 5。 这是一个深思熟虑且主动的模型，以一半的价格接近 Fable 5 的前沿智能。

Grok Build new update brings a guided /tutorial tour, smarter search tool overrides, live workflow progress with one-click failed-run resume, and a wide reliability pass across workflows, voice, and tools Release Notes: v0.2.112 Breaking Changes: • CLI version policy now has separate soft update floors/ceilings and hard startup requirements Features: • New /tutorial slash command opens an opt-in nine-topic onboarding tour of Grok. • New tool_overrides option lets you set date cutoffs and domain allowlists for the agent's built-in search tools. • New toolOverrides option lets you set date cutoffs and domain allowlists for the agent's built-in search tools. • New config options let you add query parameters or environment-backed headers to custom model providers and control which variables reach shell tools. • Terminal and environment fixes are now consolidated under the /doctor command with clearer guidance. • Marketplace add now rejects non-git URLs at add time instead of failing later. • Slash commands can now show optional bracket tags (e.g. [new]) via config or remote settings. • Queued prompts now offer an [edit] mouse button alongside Send now and cancel. • Voice shortcut toggle in settings can disable the Ctrl+Space/F8 keybind without disabling voice entirely. • Image edit can now use a remotely configured model slug instead of the hardcoded default. • `grok doctor fix` can now repair common tmux clipboard and passthrough problems. • Per-provider auth helpers now work on Windows and can run from a configurable working directory. • /resume now shows only native Grok sessions by default and shows a hint when external sessions are hidden. • `grok --resume` can now resume a session by its title as well as by ID. • Workflows overlay now shows live per-agent progress and automatically follows the active phase. • Workflow runs that failed can now be resumed; scratch file limits were also increased. • Hooks can now be defined in config.toml in addition to JSON files. • Clicking the "still running" status now opens the tasks pane. Bug Fixes: • File attachments now appear correctly when resuming or replaying conversations. • Terminal output from remote clients is now recorded so read-file hints and monitors function correctly. • Background shell commands now correctly report their real exit codes instead of always showing -1. • Marketplace source refreshes no longer hang the TUI or trap you in the extensions modal. • Background task tray now correctly clears killed tasks and keeps task descriptions after reconnect. • Dashboard overlay now correctly returns after forking a dashboard-attached session. • Linux voice dictation now works on PipeWire versions before 1.6. • Fork from a rewound session now copies the correct live-branch history. • Account pane now shows name and email even after the access token expires. • Voice mode now lets you edit already-dictated text without closing the microphone. • Fixed startup hangs on Linux after concurrent launches or rapid restarts. • MCP tools now appear without restart after enrolling or updating a managed service. • Plugin subagents now see the same MCP tools as the parent session. • Copy confirmations now show shorter messages when the clipboard succeeds. • Repeated identical tool calls now end the turn silently instead of showing a stop banner. • Web search now defaults to grok-4.5. • Voice dictation text is no longer dropped when pressing Enter to send. • Bash mode (`!`) now shows yellow prefix and action label in minimal mode. • Parked turns no longer spam duplicate "Worked for" markers in the transcript.

@Xinyu2ML the full graph makes this even weirder

claude opus 5 vs fable 5 vs gpt 5.6 sol vs kimi k3 @AnthropicAI released claude opus 5 today. per the announcement, it's "a thoughtful and proactive model that comes close to the frontier intelligence of claude fable 5 at half the price." key facts from the release: • priced at $5/mtok input, $25/mtok output – unchanged from opus 4.8 • anthropic reports it as state-of-the-art on frontier-bench v0.1 (43.3%), gdpval-aa v2 (1861), arc-agi-3 (30.2%), zapier automationbench (26%) and osworld 2.0 (70.6%) • per the same release, it stays behind mythos 5 on cybersecurity, and behind fable 5 on parts of agentic coding (deepswe 68.8 vs sol's 72.7, frontiercode 53.4 vs fable's 53.5) • runs at max thinking effort by default; a fast mode is offered at ~2.5x speed for 2x the price. our runs used the standard setting our test – 3 prompts, single-file html, @threejs, fully procedural, no assets. each has a small fire button (web audio gunshot + muzzle flash + recoil + ejected brass) and a disassemble toggle that explodes the gun into labelled parts and rebuilds it: 1. 5.56 m4 carbine – collapsible stock, safe·semi·burst lower, quad-rail with r14–r28 panel numbers, aimpoint red-dot, vertical foregrip, folding bipod, a2 flash hider, breaks into 10 parts 2. glock 18c – select-fire machine pistol, ported 18c compensator slots, "glock 18c / austria 9x19" roll-marks, extended 33-round mag, field-strips into slide, ported barrel, recoil spring, frame, mag 3. steyr tmp – ribbed polymer housing, threaded barrel, integral forward vertical foregrip, canted translucent 30-round mag ran on @aimlapi results: - cost #1 gpt 5.6 sol – $1.26 #2 kimi k3 – $3.60 #3 opus 5 – $6.22 #4 fable 5 – $8.12 - tokens #1 gpt 5.6 sol – 36,554 #2 fable 5 – 120,203 #3 opus 5 – 137,093 #4 kimi k3 – 196,596 - lines of code #1 opus 5 – 2,738 #2 gpt 5.6 sol – 2,432 #3 kimi k3 – 1,907 #4 fable 5 – 1,764 - generation time #1 gpt 5.6 sol – 7.8 min #2 fable 5 – 23.5 min #3 opus 5 – 24.8 min #4 kimi k3 – 93.8 min observations: • opus 5 wrote the most code and shipped the best-looking guns – the two go together. 2,738 lines, almost no boilerplate, and it beat fable at 3/4 of fable's price. all 3 models render, all 3 disassemble reliably • it is the only model that respected the black-on-black problem. its studio env is a narrow overhead strip with a comment explaining why ("narrow + long beats big + square here"), and parts separate by roughness, not colour – slide 0.44, ribs 0.52, grip 0.86. that is why the lighting reads like a product shot • the steyr ribbing and m4 proportions are the best in the test. the ribs are real half-buried cylinders merged into one geometry, each catching a single specular line – not a bump map. thin barrel, slim quad-rail, correct 84cm scale, all computed from named stations, not eyeballed • price per 100 shipped lines – gpt sol $0.052, kimi $0.189, opus $0.227, fable $0.460. opus buys frontier-grade detail for half of fable's per-line rate opus 5's code quality: upsides: - it engineers instead of eyeballing. everything is built to real-world scale (m4 = 84cm), the rail teeth sit at the exact 1cm pitch a real picatinny rail uses, and the steyr's ribbing is actual 3d geometry catching a specular line, not a bump texture faking it - it caught a classic three.js trap the others didn't – the bevel on extruded shapes quietly inflates the silhouette, so it pre-corrected every surface constant so decals and ribs sit flush instead of floating or z-fighting downside: - one systemic bug: every left-side engraving renders mirrored. roll-marks, the safe·semi·burst selector, the mag numbers – all reversed. single cause, but it hits all three guns opus 5 is the strongest engineer in this field – it saw traps the others walked into and reasoned about physics constants, real-world scale, and sequences instead of guessing follow @thehypedotnews for 24/7 ai news, analysis and breakdowns

Anthropic launched Claude Opus 5 at half the price of its most powerful model Fable 5, while at the same time beating Feble 5 in almost all benchmarks. Costs the same as Opus 4.8, $5/$25 per mn input/ output tokens, but exactly half as much as Fable 5. Frankly I am trying to figure out at this price of Opus 5 and with this great benchmark result why would I ever need to use Fable 5. Opus 5 also has less restrictive cyber classifiers than Fable, automatic fallback options, mid-conversation tool changes and no special data-retention requirement for general access. The cheaper model is winning specifically on economically valuable activities—coding, operating computers, research and completing business processes—not merely trivia or academic question answering. Anthropic repeatedly describes Opus 5 as more willing to verify, recover and continue until a task is actually complete. Examples include: - Creating its own computer-vision pipeline to reconstruct a 3D machine component from raw pixels - Finding the root cause of an open-source bug and fixing an edge case missed by the existing community patch - Building a market-data integration and then creating its own test harness when no live validation feed was available - Completing Zapier’s full churn-prevention workflow when previous models failed These are signs of adaptive problem-solving: the model notices that a required capability or validation mechanism is missing and constructs one instead of stopping. Anthropic also describes Opus 5 as more thorough about validating its work and recovering when necessary. The launch gives examples in which it constructed its own vision pipeline, identified the underlying cause of a software bug and built a test harness when no live data feed existed. That is potentially more commercially meaningful than a modest benchmark gain. An agent that finishes 70% of jobs without intervention can be dramatically more valuable than one that produces slightly better individual responses but frequently stalls. GPT-5.6 Sol leads DeepSWE: 72.7% versus Opus 5’s 68.8%

WOW! I thought Opus 5 would just be a cheaper Fable. I was dead wrong. This model is absolutely BONKERS. I'm genuinely blown away. Opus 5 built the best Rocket League clone i've ever seen, and it did it while consuming only 27% of my 5x Max sub. Sound on. There's layers to this onion. I'm floored rn. At least for building games and 3D modeling, this has to be the best model in the world by a notable margin. Look at the car. The reflections. The details. Playable game & open source code in the comments.

隆重推出 Claude Opus 5。 这是一个深思熟虑且主动的模型，以一半的价格接近 Fable 5 的前沿智能。

隆重推出 Claude Opus 5。 这是一个深思熟虑且主动的模型，以一半的价格接近 Fable 5 的前沿智能。