AI Models & Assistants

모든 레벨

Claude is our core — but the same skills travel. This section widens the lens to the whole AI world: the major assistants, how they differ, when to use which, and the techniques that transfer across all of them.

What you'll learn

• Pick the right model for a job without re-learning the field every time
• Move between assistants — ChatGPT, Gemini, Grok, open models — without losing your technique
• Run capable models on your own machine, and know when that is worth it
• Combine Claude with local models and agent frameworks

Start here

If you read one page in this section, read this one. Everything below is a branch off it.

• How to choose a model — a durable framework that outlives any particular release.

Then follow whichever thread matches what you are doing.

The current flagships

The September 2026 releases at the top of both stacks, read side by side.

• Claude Opus 5.5: the field guide — Anthropic's new recommended default: $4/$20, cache reads on a 5% tier, four 400-errors and one silent response-shape change
• Claude Fable 5.1: what changed and how to migrate — three breaking changes, five betas, cache reads at a quarter of the price
• GPT-6 Astra: the field guide for Claude users — same list price, different bill, Critical-tier cyber classification
• Gemini 4 Argon: the field guide for Claude users — Google's new generation: the 1M is the output limit (up from 64K), not the context window; why the independent index puts it five points behind Opus 5.5 despite the "frontier lead" headlines; 62K output tokens per task against Astra's 27K; and the Fairwind terms that mean you probably cannot call it

The workhorse tier

The cheap tier is where the interesting numbers are. Same lesson every time: match the model to the shape of the work, not to its position in the price list.

• Claude Sonnet 5.5: the field guide — Sonnet 5's exact prices, up to 30% less per task: five 400-traps, thinking that moves instead of switching off, and the one benchmark row where it beats Opus 5.5
• Claude Haiku 5.5: the 100K price cliff — the only current Claude model priced by prompt length: two rate cards 5x apart, where one token past 100,000 reprices the entire request; why the cliff arrives ~23% earlier in text than you expect; the honest saving (~87% below it, ~35% above, not the 90% on the rate card); and the one lever where trimming 21% of a prompt cuts 84% of the cost
• GPT-6.1 Sol: the field guide for Claude users — Sonnet 5.5's exact rate card at half its cached-read price: the 922K input ceiling inside a 1.05M window, the cost cliff at 272K, tool calling on only one of the two APIs, and the benchmark row Claude still wins by 8.7 points
• Gemini 3.8 Flash for Claude users — Sonnet-priced, "works harder" by default, ~30% more tokens per task
• DeepSeek-V4.1-Flash — 552B MIT open weight, 1M context, 890-byte KV cache, a 1–100 effort dial, and a routing change that silently replaced deepseek-v4-pro

Coming from another assistant

You already know one tool and want your habits to carry over.

• ChatGPT for Claude users
• Gemini for Claude users
• Grok for Claude users
• Porting prompts across models — what transfers, what breaks

Comparing before you commit

• How to read an AI benchmark without getting fooled — read this before any other row in this list: what a SWE-bench score literally measures, the 4.8 points a vendor's own footnote shows come from the harness rather than the model, the contamination discount that turns 70%-plus into ~23% on unseen code, and a six-question checklist for any number
• Is the model nerfed? How to actually tell — the question that follows every launch: what a pinned snapshot does and does not guarantee, the seven things that change without a new model, why the most-cited "GPT-4 got dumber" result was an eval bug, and a paired-design protocol you can run yourself
• Claude vs GPT vs Gemini for coding
• Reasoning models compared
• What AI costs across providers
• Hard spend caps: which providers actually stop the bill — the control that matters once agents can spend while you sleep: one event, four status codes (a 429 with no retry-after on Anthropic's tier cap, a 400 on the limit you set yourself, 429 on OpenAI, 402 on OpenRouter), why no cap can be exact, and the tier where the default safety net disappears
• Ember-1: Kimi K3's answers for 40% fewer tokens — identical rate card, materially smaller bill: why turning the effort dial down is not the same as efficiency, what one benchmark run really costs, and the reasoning-token share that caps your own saving
• Prompt caching across providers — explicit vs automatic control, a 512-to-4,096-token eligibility spread, and the break-even rule that says a prefix reused only once can cost you MORE cached than uncached
• Server-side conversation state across providers — all three majors now keep your conversation for you, and shipped three incompatible contracts: Anthropic's evaporates after 5 idle minutes and requires you to code a replay path, while OpenAI and Google retain it for 30 and 55 days by default — store: true is the setting nobody on your team chose
• Diffusion LLMs: the models that don't write left to right — a different decoding algorithm, not better hardware: why the open-weight pioneer's own README says quality peaks at one step per token, the conditional-independence bug that makes aggressive parallelism wrong rather than just coarse, how DiffusionGemma converted an autoregressive Gemma 4 for under 10% of its training budget, and the masked-infill jailbreak that reaches 100% on an aligned dLLM
• Latent reasoning: agents that think without words — the first open-weight agent that does most of its thinking as hidden states instead of tokens: 170-183 latent forward passes against ~302 text tokens where a text agent spends ~596, a ~10M-parameter adapter as the entire bridge, the reason the checkpoint loops forever in llama.cpp, and the causal test (44.9-51.3% accuracy loss) that quietly undermines the paper's own "inspectable thoughts" claim
• Optical context compression: sending text as an image — 86.5% fewer counted tokens on Anthropic, 6.95x MORE on Gemini for short inputs, and the accuracy cliff the token counter cannot see
• Generative media: image, audio, video
• World models: what you can actually build today — AMD is paying ~$8.2B for World Labs, and two opposite things share the name: frame generators where the world lives in the model's context and a session caps at 60 seconds, versus 3D generators that hand you Gaussian splats you own, render for free and cannot collide with
• AI explainer videos: the code-driven pipeline — the model writes a deterministic program, not a video: why Math.random() is banned, Remotion's 3-employee licence cliff, the two Manims that share an import name, and the dwell-time formula that makes it watchable
• Gemini 3.8 Live and Live Extended Thinking — Google's GA voice models for builders: async tools by default, the 15-min/2-min session walls, and why call cost grows faster than call length

Running models yourself

Local models trade some capability for privacy, cost control, and working offline. These pages cover when that trade is worth making.

• Run models locally with Ollama — the practical starting point
• Why a 35 KB system prompt breaks on a local model — context budgets, Ollama's silent middle-of-prompt truncation, and the failure signals to watch
• DeepSeek, Qwen & the open models
• DeepSeek V4-Flash-Vision: first native-vision frontier open weight
• DeepSeek-V4.1-Flash: the 1M-context open weight that replaced it — serving stacks, sampling defaults, and the effort-dial protocol
• Mistral Large 4 (Le Chonk): the 1.05T open weight you can't download yet — why "49B active" still needs ~525 GB at 4-bit, an Intelligence Index of 38 against Opus 5.5's 58, a 50% launch promo that hides the real cost-per-task, a 1M context evaluators measured near 512K, and the staged release where vetted partners get a less-moderated build first
• Colibrì: run a 744B MoE from your SSD — expert streaming on 16–32 GB machines, and what it really costs in tok/s
• Strata: a 125B MoE at 94 tok/s on a 12 GB card — the other end of that hierarchy, at interactive speed: why only 12 of 48 layers grow a KV cache, why a 58 GB download needs just 29.6 GB of RAM, and the quality bill nobody prints — 91.3% of SWE-bench Verified next to 98.7% of LiveCodeBench, a KV default that passes needle tests while losing 8–12% perplexity, and a 3x vision regression on identical weights
• DwarfStar (ds4) and the rise of overfit inference engines — a pure-C engine that runs three model families and refuses everything else: asymmetric expert quantization, bounded SSD streaming, KV snapshots that survive a restart, and its numbers read honestly
• DeepInfer: DeepSeek's own inference engine, read honestly — 16.3k stars in a day, and four things the coverage missed: it's a vLLM fork, "SM100" excludes every H100 and every consumer card, its decode throughput counts drafted tokens whether or not they were accepted, and batch-invariant inference is mutually exclusive with its own speculative decoder
• Ternary quantization in practice: Bonsai 2 27B in 5.95 GB — why stock llama.cpp loads it and silently emits garbage, why the bigger file is faster on an H100, and what "98.2% retention" hides
• MiMo-V2.6: Xiaomi's 1T open-weight omnimodal model — MIT, 42B active, video and audio in; the Terminal-Bench version trap, why Pro wants two nodes, and what the 10x UltraSpeed tier actually sells
• Whistle: speech-to-text in 16.9 MB — on-device ASR that genuinely beats Whisper base on close-mic speech, and is the worst of three models on real meeting audio: the benchmark bars the vendor's chart declines to label, the 30-second wall that lives in the encoder, the streaming API the launch thread concluded did not exist, and the one knob that took a tester from 41% to 96%
• A private local AI stack
• Claude plus local models — using both, deliberately

Agents beyond Claude

• Local AI agents
• Open-source agent frameworks
• The third-party memory layer: how claude-mem works — 103k stars for fixing session amnesia across Claude Code, Cursor and Codex, and what it actually is underneath: a second model in a daemon reading every tool call, whose hosted observer the project's own cost dataset prices as a DeepSeek model; plus the exit-code contract that decides whether a broken memory layer blocks your session
• Claude, MCP and local tools
• A2A: the agent-to-agent protocol
• Agent payments: x402, MPP and AgentCore GA — how agents actually pay for things now
• Commerce agents blueprint — Anthropic's open-source reference for shopping and merchant agents: the split-agent shape and the staged-write guardrail
• OpenAI Agents API vs Claude Managed Agents — the two hosted agent loops side by side: permissions, vaults, budgets, network policy, sandbox cost
• OpenAI Dots: the always-on agent — a persistent agent with its own cloud computer that works while you are logged off: the five-rung permission ladder, why proactive research is read-only at the tool layer rather than the prompt layer, and what the published safety numbers really measure
• Open Code Review: Alibaba's hybrid AI reviewer — deterministic rails around any model; 4.7x Claude Code's precision on its own benchmark, and a 20% recall ceiling
• AX (Agent Executor): Google's open agent runtime — kubectl-shaped Tasks, Workspaces, Gateways and Models on top of Agent Substrate; what survives a suspend/resume, the runner contract, and why the June docs no longer match the CLI
• System One decision models: Jev, Kev and the open clones — a model that returns calibrated probabilities instead of text; what "cannot hallucinate" really guarantees, the vendor numbers next to the one independent test, and where it sits in a Claude agent loop
• Train your own decision model: Jeff, Jeeves and the DIY stack — the recipe is cross-entropy over option letters plus a separately fitted temperature; ~2 hours on one GPU, the benchmark rows where a 0.8B beats the API by 19 points and the rows where it loses by 30, and the boring embeddings baseline you have to beat first
• Async-first agent harnesses: what Unreal Agent changes — why turn count, not token count, is what you pay for; 28 turns against 44 for the same score, and the three designs Unreal Labs refused to build
• Pi 1.0: the minimal harness and the codemode trick — the most complete published answer to MCP context bloat: a QuickJS/WASM sandbox where tool results never enter the transcript, four exposure modes where the default declares nothing, a 3,000-token ceiling on inlined declarations, and a suffix-append trick that keeps your prompt cache alive when a server connects
• DeepSeek Harness: the agent loop as a plugin — 242k stars and an "everything is a plugin" claim that holds: it runs your Claude Code hooks.json, speaks anthropic-messages, and defaults run_code to codemode
• Desktop agents without pixels — Vercel's agent-desktop reads OS accessibility trees instead of screenshots, and ships the clearest published spec for how a computer-use harness should handle uncertainty: refs scoped to a snapshot so a stale identifier cannot even be expressed, resolution that fails closed instead of guessing, properties that keep Unknown distinct from Absent, and a retry rule built on delivered vs not delivered rather than success vs failure

Everything in this section

The full list, including pages added since this index was written.

AAI Models & Assistants

Claude is the core here, but the same skills travel. Start with how to choose, then go as deep as you need on any assistant, on running models locally, or on agents.

AAI 모델 지형도: Claude, ChatGPT, Gemini 및 오픈 모델 중에서 고르기

작업에 맞는 AI 모델을 고르기 위한 오래 통하는 프레임워크 — Claude, ChatGPT, Gemini, Llama, Mistral, Grok 및 오픈/로컬 모델 전반에 걸쳐 — 그리고 이 모든 모델에 공통으로 전이되는 기술.

GGemini 4 Argon: The Field Guide for Claude Users (Sep 30, 2026)

Google's Gemini 4 generation opens with Argon and one headline number: a 1M-token OUTPUT limit, up from 64K. Context stayed at 1M. What the coverage gets wrong: the independent Artificial Analysis index puts Argon at 53 — five points BEHIND Claude Opus 5.5's 58 — and its cheaper $1.99-per-task cost comes from the rate card, not efficiency, since it burns 62K output tokens per task against GPT-6 Astra's 27K. Plus: a 15% hallucination rate (lowest measured), and why you almost certainly cannot call it yet.

CClaude Opus 5.5: The Field Guide

Opus 5.5 shipped 22 September 2026 at $4/$20 per MTok — 20% under Opus 5, with cache reads at $0.20 (a 5% tier, not the usual 10%). Four breaking changes 400-error a naive migration, one more changes the response shape without failing anything and makes your UI go quiet mid-task. Thinking can no longer be disabled at all, forced tool_choice is gone, and the default effort dropped from high to medium — while each level now thinks more than it did on Opus 5.

CClaude Fable 5.1: 무엇이 바뀌었고 어떻게 마이그레이션할까

Anthropic이 2026년 9월 1일 Claude Fable 5.1과 Mythos 5.1을 출시했습니다. 가격은 동일한 $10/$50, 캐시 읽기는 $0.25로 인하, 호환성을 깨는 API 변경 세 가지(강제 tool_choice 오류, 단방향 thinking 블록, 추가 전용 히스토리)와 다섯 가지 추가 베타가 포함됩니다. claude-fable-5 또는 claude-opus-5를 고정해 쓰는 모든 분을 위한 마이그레이션 가이드입니다.

CClaude Fable 5 & Mythos 5: 플래그십 필드 가이드

Fable 5는 Anthropic의 첫 Mythos급 플래그십이었으나 현재는 레거시 모델이며 Fable 5.1로 대체되었습니다. API는 인-밴드로 거부(HTTP 200, stop_reason: refusal)하며, 적응형 thinking만 있고, 원시 thinking을 반환하지 않습니다. 최상위 모델을 고정한다면 통합을 작성하는 방식이 바뀝니다. Opus 4.8 대비 언제 사용할지, fallback 패턴, 프롬프팅 전환.

CClaude Sonnet 5.5: The Field Guide

Sonnet 5.5 shipped 28 September 2026 at exactly Sonnet 5's prices — $2/$10 per MTok — and Anthropic still claims up to 30% less per task. The saving is in the token count, not the rate card. Five breaking changes 400-error a Sonnet 5 codebase, a sixth returns 200 and silences your progress UI, thinking can no longer be disabled (only moved between tool calls), and on Terminal-Bench 4.0 the cheap model scores 70.6% against Opus 5.5's 66.4%.

CClaude Haiku 5.5: The 100K Price Cliff

Haiku 5.5 shipped 7 October 2026 at $0.10/$0.50 per MTok — a 90% cut — and it is the only current Claude model priced by prompt length. One token past 100,000 reprices the entire request 5x. The threshold is measured on the new tokenizer, so it arrives ~23% earlier in text than you expect, and the 90% headline is ~87% in practice below the cliff and ~35% above it.

CClaude Opus 5: 실전 가이드

Opus 5는 2026년 7월 24일에 출시되었으며, 두 달 만에 나온 Anthropic의 네 번째 모델입니다. Opus 4.8과 동일한 $5/$25 가격이지만 Frontier-Bench 점수는 두 배 이상, ARC-AGI-3는 다음으로 좋은 모델의 3배이며 새로운 xhigh/max effort 계층이 추가되었습니다. 중요한 벤치마크, 순진한 마이그레이션을 깨뜨리는 두 가지 400 오류, 그리고 Opus 5가 스택에서 Fable 5를 대체하는 시점을 다룹니다.

CClaude Sonnet 5: 필드 가이드

Sonnet 5는 Claude Code의 새로운 기본값이자 Sonnet 4.6의 드롭인 후계자입니다. 하지만 마이그레이션 시 400 오류를 유발하는 API 제약이 세 가지 있고, 새 토크나이저는 동일 텍스트에 대해 약 30% 더 많은 토큰을 생성합니다. 가격 계산, effort 재매핑(Sonnet 5 medium ≈ Sonnet 4.6 high), 그리고 팀들이 자주 놓치는 프롬프팅 변화를 다룹니다.

CClaude 사용자를 위한 ChatGPT

Claude에 능숙하지만 ChatGPT를 써야 하나요? 중요한 차이점, 그대로 옮겨지는 것, 그리고 각각 언제 손을 뻗어야 하는지 — Custom GPT 폐지와 그것을 대체하는 Plugins까지.

CClaude 사용자를 위한 GPT-5.6와 ChatGPT Work

OpenAI는 2026년 7월 9일 GPT-5.6 제품군(Sol, Terra, Luna)과 ChatGPT Work를 출시했다. Claude 사용자에게 실제로 무엇이 바뀌었는가 — 1.05M 컨텍스트, $1 티어, 몇 시간 동안 스스로 일하는 에이전트, 그리고 아무도 다루지 않은 부분들.

GGPT-6.1 Sol: The Field Guide for Claude Users

OpenAI shipped GPT-6.1 Sol at DevDay on September 29, 2026 with the exact rate card Claude Sonnet 5.5 has — $2/$10 — and a cached-input price half of Sonnet's. The launch line is 'near-Astra intelligence for a fifth of the price'. The details that decide whether that is true for you: a 922K input ceiling inside a 1.05M window, a hard cost cliff at 272K tokens, tool calling that only works on one of the two APIs, and the benchmark row where Sonnet 5.5 still wins by 8.7 points.

GGPT-5.6 2026년 8월 업데이트: 노력 슬라이더, 무료 Think 버튼 & 272K 가격 절벽

2026년 8월 6일 OpenAI가 GPT-5.6에 대한 사이클 중간 업데이트를 배포했고, 대부분의 리뷰는 이를 'Sol이 더 똑똑해졌다'로 축약했다. 진짜 이야기: 요청 가격 방식을 바꾸는 6단계 노력 제어, 무료 계층 Think 버튼, 전체 호출을 다시 가격 매기는 272K 토큰 가격 절벽, Sol에서만 작동하는 Fast 모드. Claude 사용자에게 각각 의미하는 바.

OOpenAI Astra: 프리뷰 현장 노트

OpenAI의 '차세대 주요 모델' Astra는 2026년 8월 1일 Lean으로 검증된 열 개 수학 증명과 거의 없는 제품 스펙으로 프리뷰됐다. 실제로 확인된 것, 아직 알려지지 않은 것, 그리고 Claude 사용자가 기다려야 하는지 — 정직하고 출처가 있는 견해.

CCodex Agent Plugins & 카탈로그 페더레이션 (v0.147.0): 실전 가이드

2026년 8월 7일 OpenAI가 Codex CLI v0.147.0을 출시했습니다 — Skills를 이식 가능한 Agent Plugins로 바꾸고 4계층 카탈로그를 도입하며, 이름과 다르게 작동하는 --approve-for-me 플래그를 추가하고, --full-auto를 조용히 제거하며 MCP 2026-07-28에 옵트인시키는 릴리스. Claude 사용자 관점에서 각 부분이 실제로 하는 일.

CChatGPT로 로그인

OpenAI의 새로운 아이덴티티 계층, 출시 3일 차. 실제로 어떤 데이터가 넘어가는지, 엔터프라이즈 관리자가 무엇을 할 수 있는지, 그리고 OpenAI에 의도치 않은 발판을 내주지 않으면서 자신의 앱에 이를 수용하는 방법.

OOllama로 AI 모델 로컬에서 실행하기

오픈 웨이트 모델(Llama, Mistral, Qwen 등)을 내 컴퓨터에서 실행하세요 — 비공개, 오프라인, 실행 비용 무료 — Ollama로. 설치, CLI, 코드에서 호출하는 방법.

코코딩을 위한 Claude vs GPT vs Gemini

코딩용 프런티어 모델을 고르기 위한 프레임워크(순위표가 아님) — 중요한 요소들, 그리고 왜 당신의 저장소에서 돌린 평가가 모든 벤치마크를 이기는지.

코코딩 에이전트 CLI 비교

Claude Code, Codex CLI, Gemini CLI 그리고 오픈소스 터미널 에이전트 — 모델뿐 아니라 하네스(harness)가 결과를 결정하는 방식, 그리고 AGENTS.md로 설정을 이식 가능하게 유지하는 법.

SSupabase Evals — 실제 백엔드 에이전트 벤치마크

Supabase는 Claude Code, Codex, OpenCode를 목(mock)이 아닌 실제 컨테이너화된 Supabase 스택에서 실행하는 벤치마크를 오픈소스로 공개했습니다. 이 벤치마크가 무엇을 측정하는지, 모델이 실제로 문서와 스킬을 어떻게 사용하는지에 관한 발견, 그리고 저녁 한나절에 로컬에서 실행하는 방법을 다룹니다.

CClaude 사용자를 위한 Apple Foundation Models 3와 온디바이스 LLM 스택

Apple의 3세대 Foundation Models(AFM 3)가 WWDC 2026에서 등장했습니다 — 3B 밀집 온디바이스 모델, iPhone에서 실행되는 20B 스파스 MoE, 그리고 이제 여러분의 LLM 프로바이더(Claude 포함)를 가져올 수 있게 하는 Swift 네이티브 프레임워크. 실제로 출시된 것, 명백하지 않은 메커니즘, 그리고 Claude 앱을 배포하는 사람에게 온디바이스 스토리를 어떻게 바꾸는지.

CClaude 사용자를 위한 Gemini

Claude에 익숙한데 구글의 Gemini가 필요하신가요? 중요한 차이점들 — Gems, Workspace 통합, 거대한 컨텍스트, 멀티모달 — 그리고 무엇이 그대로 이어지는지.

CClaude 사용자를 위한 Gemini 3.6 Flash, Flash-Lite 그리고 Flash Cyber

2026년 7월 21일 Google은 Gemini 3.5 Pro를 완전히 건너뛰고 대신 세 개의 Flash-티어 모델을 출시했습니다. 실제 가격, 실제 벤치마크, 마이그레이션 함정, 그리고 대부분의 리뷰가 놓치는 부분 — Claude Opus 4.6을 V8 버그 사냥에서 이긴 사이버보안 파인튠 포함.

GGemini 3.7 Flash for Claude Users: Coding-Agent Workhorse at Half Price (Aug 2026)

On 13 August 2026, three weeks after 3.6 Flash, Google shipped Gemini 3.7 Flash at half the introductory price with a 16-point DeepSWE jump. The gotchas most write-ups skip: an expiring intro price, a 3× throughput advantage that matters for agent loops but not chat, and where it still trails Claude Sonnet 5.

GGemini 3.8 Flash for Claude Users: The Workhorse That Now Ties Opus 5 on DeepSWE (Sep 2026)

On September 2, 2026, three weeks after 3.7 Flash, Google shipped Gemini 3.8 Flash at the same $0.75 / $3.75 intro price: DeepSWE 73.7 (+8.4), Terminal-bench 2.1 89.4, HLE-Verified 54.9, plus a Cyber variant gated behind the new Fairwind Program. What the headlines skip: it 'works harder' by emitting ~30% more tokens per task, the same model ID migration rules from 3.7 still apply, and Terminal-bench 4.0 is where Claude Opus 5 stays 33 points ahead.

CClaude 사용자를 위한 Grok

Claude에 능숙한데 xAI의 Grok이 필요한가요? 중요한 차이점들 — 실시간 X/웹 검색, OpenAI 호환 API, Grok Build 코딩 에이전트 — 그리고 그대로 옮겨지는 모든 것.

MMuse Code + Muse Spark 1.2: Meta의 터미널 코딩 에이전트

2026년 8월 5일, Meta는 터미널 코딩 에이전트인 Muse Code와 함께 co-training된 모델 Muse Spark 1.2를 출시했습니다. append-only 이벤트 로그가 실제로 무엇을 제공하는지, contributor 티어가 표준 대비 왜 12.5배 저렴한지, Terminal-Bench 2.1에서 Claude Code와 어떻게 비교되는지, 그리고 어떤 에이전트를 쓰든 훔쳐올 만한 세 가지 번들 스킬은 무엇인지.

오오픈소스 AI 에이전트 프레임워크

LLM 에이전트를 구축하기 위한 주요 오픈 프레임워크 — LangGraph, LlamaIndex, AutoGen, CrewAI 그리고 최소 루프 접근 방식 — 에 대한 공급자 중립적인 지도와 선택 방법.

네네이티브 멀티 에이전트 API: OpenAI의 Responses 멀티 에이전트 베타 vs 직접 구축하기

2026년 7월 9일, OpenAI는 Responses API 멀티 에이전트 베타와 Sol Ultra Mode를 출시했습니다. '모델이 하나의 HTTP 요청으로 N개의 서브에이전트를 스폰하고 종합하는' 것을 최상위 프리미티브로 만든 최초의 프런티어 벤더입니다. 실제로 출시된 것, API 형태, 비용 계산, Anthropic의 대안, 그리고 네이티브 프리미티브가 DIY 팬아웃을 이기는 경우를 다룹니다.

AA2A: 에이전트 간 프로토콜

MCP는 에이전트를 도구에 연결합니다. A2A는 에이전트를 다른 에이전트에 연결합니다 — 프레임워크, 클라우드, 회사를 넘나들며. Agent Card, task, 여덟 가지 라이프사이클 상태, 스트리밍 vs. 웹훅, 그리고 A2A가 MCP와 어떻게 결합되는지 배우세요.

AAgent Payments: x402, MPP and AgentCore GA

How AI agents actually pay for things now — the x402 HTTP-402 protocol, the exact/upto/batch-settlement schemes, facilitators, Amazon Bedrock AgentCore Payments' Aug 2026 GA (payment sessions, Coinbase Bazar MCP, Stripe Privy, Machine Payment Protocol), and how to add a paywall or a spending agent without inventing a wallet.

모모델 간 프롬프트 이식하기

Claude, GPT, Gemini 및 오픈 모델 사이에서 프롬프트를 옮기는 법 — 깔끔하게 이전되는 부분, 모델별로 조정할 부분, 그리고 마이그레이션 워크플로.

모모델 라우팅 패턴: 캐스케이드, 분류기, 그리고 실제로 배포되는 것

각 요청을 올바른 AI 모델로 보내기 위한 설계 패턴 — 규칙 기반, 분류기, 복잡도, 캐스케이드, 앙상블, 폴백. 각 패턴을 언제 사용해야 하는지, 무엇이 깨지는지, 그리고 2026년에 실제로 배포되는 레시피.

DDeepSeek, Qwen 그리고 오픈 웨이트의 물결

강력한 오픈 웨이트 모델(DeepSeek, Qwen, Llama, Mistral)이 격차를 크게 좁혔습니다. '오픈 웨이트'가 무엇을 의미하는지, 폐쇄형 프런티어 대비 트레이드오프, 그리고 이를 활용하는 방법.

QQwen3.8-Max: 최초의 오픈 웨이트 Max급 모델 (가중치가 실제로 공개된다면)

2026년 8월 3일: 알리바바의 Qwen3.8-Max는 2.4T 파라미터 MoE(95B 활성), 1M 컨텍스트, OpenAI/Anthropic 규격 이중 API를 갖추고 API 전용으로 출시되었으며, OSWorld에서 GPT-5.6 Sol Max를 앞서는 벤치마크 점수를 기록했다. 진정으로 새로운 점, Hugging Face 페이지가 비어 있을 때 '오픈 웨이트'가 실제로 의미하는 바, 그리고 자체 호스팅을 비싸게 만드는 희소성 수학.

DDeepSeek V4-Flash-Vision: First Native-Vision Frontier Open Weight

On Aug 31 2026 DeepSeek open-sourced V4-Flash-Vision-Exp — a 305B MoE that adds native vision to V4-Flash without hurting text agent scores, matches or beats Claude Opus 4.8 on ApexBench / Agents' Last Exam / ZeroBench at a fraction of the price, and ships MIT. The specific numbers, the hard 384-token-per-image cap most guides ignore, and the concrete way to run it.

DDeepSeek-V4.1-Flash: A 1M-Context Open Weight That Ties Opus 5 on DeepSWE (Sep 2026)

On September 10, 2026 DeepSeek shipped V4.1-Flash — a 552B MIT-licensed MoE with native vision, a 1M-token window, a KV cache a quarter the size of V4-Flash's, and a 1–100 reasoning-effort dial. It ties Claude Opus 5 on DeepSWE v1.1 (74.2 vs 74.0) at $0.30 per million input tokens, and trails badly on Terminal-Bench. The numbers, the routing change that silently moved every deepseek-v4-pro call, and where Claude still wins.

CClaude 사용자를 위한 Kimi K2

Claude에는 익숙한데 Moonshot의 Kimi K2가 궁금하신가요? Claude Code를 실제로 겨눌 수 있는 오픈 웨이트 에이전트형 모델 — Moonshot이 Anthropic 호환 엔드포인트를 제공하기 때문입니다. 진짜로 다른 점은 무엇이고, 무엇이 그대로 넘어오는지.

GGLM-5.2: 오픈웨이트 프런티어 코딩 모델

Z.ai의 753B MoE (~40B 활성), 1M 토큰 컨텍스트, MIT 라이선스 — 장기 코딩에서 Claude Opus 4.8과 1점 이내로 근접, API 비용은 약 1/6. IndexShare가 실제로 하는 일, 로컬 실행 방법, 그리고 실제 환경에서 Claude Code를 이기는 지점.

IInkling: Thinking Machines의 오픈 웨이트 모델

Mira Murati의 Thinking Machines Lab이 Inkling을 출시했다 — 연속적인 사고 노력(thinking-effort) 다이얼을 갖춘 9750억 파라미터 Apache 2.0 멀티모달 MoE. 출시 보도가 무엇을 놓쳤는지, 왜 의도적으로 벤치마크에서 지는지, 그리고 그 노력 노브가 실제로 무엇을 하는지.

KKimi K3: 세계 최대 오픈 웨이트 모델

Moonshot AI의 2.8조 파라미터 오픈 웨이트 모델이 많은 벤치마크에서 Claude Opus 4.8을 이기고 Frontend Code Arena에서 1위. 아키텍처적으로 무엇이 다른지, 캐시 히트 가격이 비용 계산을 뒤집는 때, 손 뻗을 가치가 있는지.

KKimi K3 로컬 실행: vLLM, DSpark & 실제 하드웨어 청구서

2026년 Moonshot의 2.8T 오픈 웨이트 Kimi K3를 셀프 호스팅하기 위한 결정판 실전 가이드. DSpark speculative decoding이 실제로 하는 일, 최소 하드웨어(1× DGX B300 또는 16× DGX Spark), 정확한 vLLM serve 명령, 아무도 경고하지 않는 prefill/decode 비대칭, 호스팅 추론이 자체 랙을 언더컷할 때.

MMiMo-V2.6: Xiaomi's 1T Open-Weight Omnimodal Model

A 1.02T-parameter MoE with 42B active, 1M context, text/image/video/audio input — MIT-licensed on Hugging Face. What the benchmark tables actually say once you check which version was run, why Pro needs 32 GPUs and Flash doesn't, and what the 10x UltraSpeed tier is really selling.

GGrok Build 내부 들여다보기: 오픈 에이전트 하니스 읽기

xAI가 자사 코딩 에이전트의 전체 Rust 소스를 Apache 2.0으로 공개했다 — 에이전트 루프, 툴 레이어, TUI, 확장 시스템까지. 크레이트 맵이 에이전트 구축에 대해 가르쳐 주는 것, 여기서 '오픈 소스'가 왜 소스 공개(source-available)를 의미하는지, 그리고 트리에 여전히 남아 있는 데이터 수집기 코드가 무엇을 말해 주는지.

PPoolside Laguna: 체급을 뛰어넘는 오픈 웨이트 코딩 모델

Poolside AI의 Laguna 패밀리(XS 2.1, S 2.1, M.1) — 3B–23B 활성 파라미터, 1M 토큰 컨텍스트, OpenMDW-1.1 라이선스를 가진 MoE 코딩 모델. 8B 활성의 S 2.1이 실제로 무엇을 이기는지, 왜 3B 활성의 XS 2.1이 MacBook에서 돌아가는지, 최초로 알려진 FP8 기반 RL 학습 파이프라인, 그리고 Claude, GLM-5.2, Kimi K3 대신 이 모델을 선택해야 할 때.

생생성형 미디어: 이미지, 오디오 및 비디오 AI

텍스트 채팅을 넘어서 — 이미지, 비디오, 음성, 음악을 위한 AI 지도: 주요 도구, 지속되는 기술, 그리고 중요한 권리/윤리.

OOpen Code Review: Alibaba's Hybrid AI Reviewer vs Claude Code, Codex and Any Model (2026)

Alibaba's open-code-review (ocr) is the AI code reviewer trending on GitHub: a Go CLI that puts deterministic file selection, rule matching and line anchoring around an LLM agent. Same Claude model, 4.7x the precision of a Claude Code review at 1/15 the tokens on its own benchmark — and a best-case recall of 20%. How it works, how to run it with Claude, GPT, Qwen or a local model, delegation mode, CI, and the gotchas.

MMiniMax H3 (Hailuo 3.0): 오픈 웨이트 옴니 비디오 모델

MiniMax는 33B 단일 스트림 트랜스포머를 오픈소스로 공개했습니다. 이 모델은 한 번의 패스로 네이티브 스테레오 오디오와 함께 2K 해상도의 4~15초 비디오를 생성합니다. 라이선스는 미국/EU/영국/한국에서의 셀프 호스팅을 조용히 차단하며, 오픈 베이스는 2K가 아닌 768p이고, 33B 파라미터 중 약 13B는 추론 시 건너뛸 수 있는 AdaLN 브랜치입니다. H3가 실제로 무엇이며, 비용은 얼마이고, 언제 클로즈드 비디오 API를 이길 수 있는지.

AAI의 실제 비용 (제공업체별 비교)

제공업체 간 AI 비용을 비교하기 위한 견고한 프레임워크 — 가격 책정 원형, 워크로드 추정 방법, 청구액을 줄이는 레버, 그리고 오픈/로컬이 유리한 시점.

HHard Spend Caps: Which AI Providers Actually Stop the Bill

A soft cap emails you. A hard cap blocks the request. This page maps which AI providers do which — and the trap nobody documents: the same event (you hit your cap) returns HTTP 429 with no retry-after on Anthropic's tier cap, HTTP 400 on the limit you set yourself, 429 on OpenAI, and 402 on OpenRouter. Plus why no cap can be exact, where the Custom tier loses its safety net, and how to build a pre-flight killswitch.

PPrompt Caching Across Providers

Every major provider caches your prompt prefix — but explicit vs automatic, 5 minutes vs 24 hours, and a write premium that can make caching cost you more. The portable break-even rule and the five traps when you move a workload.

AAI 에이전트가 토큰을 태우는 이유 (그리고 청구서를 억제하는 법)

에이전트 실행은 채팅보다 50~1000배 비싸다 — 모델이 멍청해져서가 아니라 컨텍스트 눈덩이 때문이다. 그 메커니즘, 놀라운 숫자들, 그리고 Claude, GPT, Gemini, 오픈 모델 전반에서 실제로 청구서를 줄이는 네 가지 레버.

SServer-Side Conversation State Across Providers

Anthropic, OpenAI and Google all now keep your conversation on their servers — and shipped three incompatible contracts. One is an ephemeral latency trick you must code a replay path for; two are durable storage that is ON by default, for 30 and 55 days. The axis that actually decides your design.

OOptical Context Compression: Sending Text as an Image

Render your context into a picture and the vision encoder becomes a compression dial. Measured across Anthropic, OpenAI and Google, an image of code counts 76–87% fewer input tokens than the same code as text — and on Gemini it can cost 6.95x MORE. The mechanism, the accuracy cliff, and how to test it on your own documents.

AAI Explainer Videos: The Code-Driven Pipeline

Opus 5.5 explainer videos hit the Hacker News front page — and almost nobody explains what is actually happening: the model writes a deterministic animation program, not a video. The mechanism, the three render engines and their licences, the two traps that break agent-written animation code, the pacing formula that makes the difference between watchable and unwatchable, and what the bill really looks like.

로로컬 AI 에이전트 구축하기

로컬 오픈 웨이트 모델로 완전히 자신의 머신에서 실행되는 자율 에이전트 — 프라이빗하고, 오프라인이며, 실행 비용이 무료입니다. 아키텍처, 트레이드오프, 그리고 시작하는 방법.

CClaude + 로컬 모델: 하이브리드 패턴

Claude와 로컬 오픈웨이트 모델을 시너지로 함께 작동시키세요 — 라우터, 초안 후 다듬기, 프라이버시 마스킹, 대량 전처리 — 프런티어의 추론 능력을 로컬의 프라이버시와 비용으로.

AAI 게이트웨이: LiteLLM, OpenRouter, Portkey, Vercel

프로덕션 AI 게이트웨이는 여러분의 앱과 모든 모델 — Claude, GPT, Gemini, Llama — 사이의 잃어버린 라우터입니다. LiteLLM, OpenRouter, Portkey, Vercel AI Gateway를 비교하고, Claude Code를 자체 프록시를 통해 배선하는 법.

로로컬 코딩 에이전트 (그리고 Claude와의 조합 방법)

당신의 머신에서 실행되며 레포지토리를 편집하는 코딩 에이전트 — Aider, Cline, Continue, OpenHands, Claude Code — 품질을 위한 Claude 또는 프라이버시를 위한 로컬 모델로 구동됩니다.

MMCP로 Claude를 로컬 도구 및 에이전트에 연결하기

Model Context Protocol을 사용하면 Claude가 여러분의 컴퓨터에서 비공개로 실행되는 도구, 데이터, 에이전트를 오케스트레이션할 수 있습니다 — Claude를 두뇌로, 로컬 기능을 손으로 삼는 시너지입니다. 로컬 MCP 서버를 추가하거나 직접 만들어 보세요.

프프라이빗 로컬 AI 스택 구축하기 (엔드투엔드)

모든 것을 하나로 연결하기: 로컬 모델(Ollama) + 모델 비종속 에이전트 + 로컬 MCP 서버를 통한 도구 — 당신의 기기에서 동작하는 프라이빗 어시스턴트, Claude는 선택적 스마트 레이어로.

로로컬 에이전트 vs Claude: 의사결정 가이드

완전 로컬 에이전트, Claude 기반, 아니면 하이브리드? 이를 결정하는 요인들 — 프라이버시, 작업 난이도, 비용, 지연 시간, 신뢰성 — 그리고 왜 하이브리드가 종종 이기는지.

로로컬 및 하이브리드 에이전트 보안

파일을 편집하고 명령을 실행할 수 있는 에이전트는 강력하면서도 위험합니다. 최소 권한, 샌드박싱, 사람의 승인, 프롬프트 인젝션 방어, 예산 상한 — 에이전트를 안전하게 실행하는 방법.

추추론 모델 비교

Claude, GPT, Gemini, DeepSeek, Qwen 전반의 사고 예산(thinking budget)과 추론 강도(reasoning effort) — 언제 사고에 토큰을 쓰고, 언제 쓰지 말아야 하는가.

전전이중 음성 AI: 음성 에이전트가 갑자기 진짜가 된 이유

GPT-Live와 새로운 음성 네이티브 모델들은 말하면서 동시에 듣습니다. 전이중이 무엇을 바꾸는지, 어떻게 동작하는지, 그리고 지금의 음성 AI 지형.

CClaude 음성과 Opus & Sonnet (2026년 7월): 추론 우선 음성 vs GPT-Live

Anthropic이 2026년 7월 23일 Claude 음성을 Opus, Sonnet, 그리고 연결 앱으로 개방했습니다 — 하지만 턴 기반 스택을 유지했습니다. 실제로 무엇이 바뀌었는지, 통화 도중 언제 모델을 전환할지, 그리고 이 설계 베팅이 OpenAI의 풀 듀플렉스 GPT-Live와 어떻게 다른지.

GGemini 3.8 Live and Live Extended Thinking: The Builder's Field Guide to Google's New Voice Models (Sep 2026)

On September 15, 2026 Google shipped gemini-3.8-live and gemini-3.8-live-extended-thinking. The launch post covers benchmarks. This page covers the rest: which model to pick, the migration traps coming from 3.1 Flash Live (async tools are now the default, thinkingLevel errors, proactive audio can't be turned off), the 15-minute and 2-minute session walls, and why a voice session's cost grows faster than its length.

CCommerce agents blueprint (Anthropic reference)

Anthropic's Sept 2026 open-source blueprint for shopping and merchant agents on Claude — the split-agent shape, the staged-write guardrail, and the patterns you can steal even if you're not building commerce.

토토큰 속도: AI 추론이 갑자기 10-15배 빨라진 이유

웨이퍼 스케일 칩과 LPU가 프런티어 모델을 초당 수백 토큰으로 서빙합니다. 추론 속도를 실제로 제한하는 것, 새로운 하드웨어 경쟁, 그리고 빠른 토큰이 에이전트·음성·추론을 바꾸는 이유.

DDiffusion LLMs: The Models That Don't Write Left to Right

Mercury 2 answers a full reasoning question in 1.7 seconds at 1,009 tok/s; Gemini Diffusion samples at 1,479 tok/s against Gemini 2.5 Flash's 272.4. Same transformers, different decoding algorithm. The practical guide: why the open-weight pioneer's own README says quality peaks when you use one step per token, the conditional-independence bug that makes parallel decoding wrong, block diffusion as the fix that restored the KV cache, how DiffusionGemma converted an autoregressive Gemma 4 for under 10% of its training budget, and the masked-infill jailbreak that hits 100% on an aligned dLLM.

컴컴퓨터 사용 에이전트 비교

Claude, GPT, Gemini 모두 이제 화면을 조작할 수 있습니다 — 하지만 각자 다른 액션 공간, 좌표 계약, 안전 게이트를 노출합니다. 실제로 무엇이 깨지고, 그것을 견디는 루프를 어떻게 만드는가.

AAI 에이전트를 위한 공유 로그인 브라우저: ego lite 패턴

새로운 부류의 브라우저 — 2026년 7월 24일 GitHub Trending #1에 오른 ego lite가 최고의 예 — 는 Claude Code, Codex, Cursor, 그리고 어떤 코딩 에이전트든 여러분의 실제 로그인을 상속받는 격리된 Chromium 세션을 몰이하게 합니다. Spaces와 semantic snapshot이 실제로 해결하는 것, 왜 에이전트 작업에서 Playwright를 앞지르는지, 그리고 자율 프로세스와 세션 쿠키를 공유하는 보안 비용.

CCloudflare Kitesurf: The First Browser Runtime Built for AI Agents (Not Humans)

On August 6, 2026 Cloudflare launched Kitesurf, a stateless, agent-first browser that runs entirely in V8 isolates on Workers. It uses 3-7x less CPU and memory than Chromium on common agent tasks — at the cost of ~1.7x wall-clock time. Written in Rust, using Firefox's CSS parser (Stylo) and the Blitz rendering engine, drop-in compatible with Puppeteer/Playwright via the Chrome DevTools Protocol. This page: what changed, when to reach for it instead of headless Chrome, how to wire it into Claude Code / Codex / your MCP client, and the four things it flat-out can't do yet.

DDesktop Agents Without Pixels

Vercel Labs shipped agent-desktop on 28 October 2026 — accessibility-tree computer use in Rust, Apache-2.0, for macOS/Windows/Linux. The headline is 'AX trees instead of pixels'. The part worth stealing is its correctness model: refs scoped to a snapshot, fail-closed resolution that refuses to guess, tri-state properties, and a retry rule built on DELIVERED vs NOT DELIVERED rather than success vs failure.

SSKILL.md: 크로스 에이전트 오픈 표준

하나의 SKILL.md 파일, 서른두 개의 에이전트. Agent Skills 오픈 표준이 실제로 무엇인지, 점진적 공개(progressive disclosure)가 어떻게 컨텍스트 비용을 낮게 유지하는지, 무엇이 이식성을 깨뜨리는지, 그리고 Claude Code, Codex, Gemini CLI, Cursor에서 수정 없이 동작하는 스킬을 작성하는 방법.

CColibrì: Run a 744B MoE From Your SSD (and What It Really Costs You)

Colibrì is a pure-C, zero-dependency engine that runs GLM-5.2 (744B), Kimi K3 (2.8T), DeepSeek V4.1 Flash and six other frontier MoE models on a 16–32 GB machine by keeping the dense layers in RAM and streaming the routed experts from NVMe. This page explains the memory hierarchy, the numbers that decide whether it is usable for you (0.05 to 6.8 tok/s), the SSD-wear and speculative-decoding gotchas, and the exact commands to get a first token.

OOpenAI Agents API vs Claude Managed Agents: The Hosted Agent Loop, Compared

On September 10, 2026 OpenAI put the Codex harness behind a single API call: the Agents API public beta — durable sessions, hosted or self-hosted sandboxes, automatic compaction, tool search, subagents. Anthropic has run the same shape since April as Managed Agents. Both use the same four primitives; they differ on permissions, secrets, budgets, network policy and what the sandbox costs. The side-by-side, the request shapes, the six gotchas, and when to pick which.

AAX (Agent Executor): Google's Open Agent Runtime, Explained for Builders (2026)

google/ax hit the Hacker News front page on September 20–21, 2026 (507 points). It is not a model, a harness or a hosted API: it is a kubectl-shaped orchestrator that runs agent Tasks in sandboxes on top of Agent Substrate, with Workspaces (git + MCP + skills), Gateways (egress allowlists) and Models (provider + secret) as declarative resources. What the four primitives do, what survives suspend/resume and what does not, the runner contract your image must honour, the GKE requirements, and why the June-era docs no longer match the CLI.

SSystem One Decision Models: Jev, Kev and the Open Clones, Explained for Builders (2026)

TypeSafe AI's Jev (out of stealth September 15, 2026) is a model that never writes text: you send it program state plus typed questions and it returns calibrated probabilities for all of them in one parallel pass, 70–500 ms, $0.042 per million input tokens, output free. Two Hacker News front-page stories on September 21 (Kev, an Apache-2.0 recreation on Qwen3.5, and the jev-leftpad joke) made it the week's topic. What a System One model actually does, what the 'cannot hallucinate' claim really guarantees, the vendor numbers next to the one independent test, where a decision model belongs in a Claude agent loop, and how to run Kev or Nimble locally.

TTrain Your Own Decision Model: Jeff, Jeeves and the DIY System One Stack (Sep 2026)

Two weeks after Jev made decision models a category, the open stack caught up: Jeff fine-tunes a 0.8B Qwen into a Jev-compatible classifier in ~2 hours on one workstation GPU, and PostHog's Jeeves shows that letting a decision model think first beats Jev on its own benchmark. What the training recipe actually is (cross-entropy over option letters, a separately fitted temperature), where tiny models beat the API and where they get destroyed, the 26-option ceiling, the 20x CPU latency cliff, and the boring embeddings baseline that scored 93.0% and that you must beat before any of this is worth doing.

TTernary Quantization in Practice: Bonsai 2 27B in 5.95 GB

PrismML compressed Qwen3.8-27B to 1.75 bits per weight and 5.95 GB. The practical guide nobody wrote: why stock llama.cpp loads the file and silently produces garbage, why the bigger 7.21 GB file is faster on an H100 but slower on a 4090, why 5.95 GB of weights needs 12 GB of VRAM, and which capabilities the 98.2% retention headline quietly loses.

AAsync-First Agent Harnesses: What Unreal Agent Changes

Unreal Labs shipped an MIT-licensed Go harness that matches Codex's pass rate on Terminal-Bench 4.0 for 39% less money — by never making the model wait for a tool. The event-log mechanism, why turn count is the real cost unit, and the three things they deliberately refused to build.

HHindsight: agent memory that learns

The MIT-licensed memory server that topped GitHub trending — retain, recall and reflect, four parallel retrievers fused by RRF, observations that get refined instead of overwritten, and the 83.6%-vs-39.0% result that says architecture beats model size.

EEmber-1: Kimi K3's Answers for 40% Fewer Tokens (Sep 2026)

Fireworks post-trained Kimi K3 to stop over-thinking and shipped the result as Ember-1 — same price per token, 35–50% fewer reasoning tokens, and it beats K3 at max effort on two of five benchmarks. Why turning effort down is not the same thing, what the '-51.9% / -23.1 USD' column really says, the implied cost per benchmark task, the modality that quietly disappears, and the Kimi K3 licence clause that lets a derivative stay closed.

OOpenAI Dots: The Always-On Agent, and the Permission Ladder Underneath It

At DevDay on September 29, 2026 OpenAI shipped dots — persistent agents that each get their own cloud computer and browser, run on GPT-6 Astra, and keep working while you are logged off. The launch coverage stops at the demo. This page covers the mechanism: the five-rung permission ladder, why proactive research is read-only at the tool layer rather than the prompt layer, what the published safety numbers actually measure, and how a dot compares to Claude Managed Agents and Cowork.

WWorld Models: What You Can Actually Build Today

AMD is paying about $8.2 billion for World Labs, and two very different things are being sold under the same name. Frame generators like Genie keep the world inside the model and cap a session at 60 seconds. 3D generators like Marble hand you Gaussian splats you own, render for free and cannot collide with. The API shapes, the export tiers, the WorldScore evidence, the open weights you can run, and the five jobs a world model will not do.

IIs the model nerfed? How to actually tell

"They nerfed it" is the most-repeated claim in AI and the least-measured. On the Claude API the weights behind a model ID cannot change — every current ID is a pinned snapshot. But seven other things can change under you without a new model, from a published system-prompt update to a harness that silently swaps models mid-session. Here is what can and cannot move, why the most-cited 'GPT-4 got dumber' result was an eval bug, and a measurement protocol you can run in an afternoon.

PPi 1.0: The Minimal Harness, and the Codemode Trick That Fixes MCP Token Bloat

Pi 1.0 shipped 1 October 2026 and went to #1 on Hacker News. The interesting part is not the TUI — it is the context engineering: a QuickJS/WASM sandbox where tool results never enter the model's context, four MCP exposure modes, a 3,000-token ceiling on inlined tool declarations, and a trick that appends a new system-prompt section to the transcript instead of mutating tool definitions so your prompt cache survives. Plus the deliberate omissions (no sub-agents, no plan mode), the $0.05 cache-warming rule, and the trust gap that loads a cloned repo's AGENTS.md without asking.

DDeepSeek Harness: the agent where the agent loop itself is a plugin

DeepSeek Harness (dsh) passed 242,000 stars and shipped desktop installers on 1 October 2026. The headline is 'everything is a plugin' — and unlike most such claims this one is load-bearing: the model adapter, the tool registry, the session log and the agent loop are all replaceable from a YAML patch file. This page covers the three facts that matter to a Claude user (it runs your existing Claude Code hooks.json, it speaks anthropic-messages, and run_code is codemode by default), the three-mode sandbox and the kernel primitives behind it, the append-only session log invariant, and the config override that silently deletes your other providers.

SStrata: a 125B MoE at 94 tok/s on a 12 GB Card — and the Exact Bill

Strata runs Qwen3.8-Flash-Next (125B, 6B active) on a 12 GB gaming GPU at interactive speed. This page explains the three architecture facts that make it possible, reads the benchmark table honestly, and prices the quality you give up: 91.3% of SWE-bench Verified against 98.7% of LiveCodeBench, a KV-cache setting that passes needle tests while losing 8-12% perplexity, a speed mode that strips refusal directions, and a measured vision regression on identical weights.

HHow to Read an AI Benchmark Without Getting Fooled

Every launch post leads with a benchmark number. Almost none of them tell you the three things that decide what it means: which scaffold produced it, how many trials it averages, and whether the model had already seen the answers. A reader's manual for SWE-bench, SWE-bench Verified, SWE-bench Pro and the arena leaderboards — with the vendor's own footnotes as evidence.

DDwarfStar (ds4) and the Rise of Overfit Inference Engines

ds4 is a 23.6k-star, pure-C engine from the creator of Redis that runs DeepSeek V4 Flash, GLM 5.x and Qwen3.8-Flash-Next locally — and refuses to run anything else. This page explains the three mechanisms a general runtime will not give you (asymmetric routed-expert quantization, bounded expert streaming from SSD, KV snapshots that survive a restart), reads its own numbers honestly — the 126 t/s headline is an eight-GPU aggregate across 16 sessions, not your Mac — and prices the real cost of the overfit bet: deliberately disposable model support, 761 open issues, and a main branch that last moved on 20 September 2026.

MMistral Large 4 (Le Chonk): the 1.05T open weight you can't download yet

Europe's trillion-parameter MoE: why 49B active params still need ~525 GB at 4-bit, the 1M context that evaluators measured at ~512K, a 50% launch promo that hides the real cost-per-task, an Intelligence Index of 38 against Opus 5.5's 58 — and the staged weight release where vetted partners get a less-moderated build first.

WWhistle: Speech-to-Text in 16.9 MB — and the Benchmark Bars Nobody Labelled

Cactus Compute's Whistle fits a multilingual ASR model in 16.9 MB and hits 11.1 ms to first token on a CPU. The size claim is real. This page reads the vendor's own benchmark chart bar by bar — including the three competitor columns it declines to label — and finds the pattern that explains every field complaint: Whistle wins on close-mic single-speaker audio and is the worst of the three on real meeting audio, 4.6 WER points behind a model 8.6x its size. Plus the 30-second architectural wall, the streaming API the launch thread missed, and the one knob that took a tester from 70/170 to 164/170.

DDeepInfer: DeepSeek's Own Inference Engine, Read Honestly

On 20 October 2026 DeepSeek open-sourced DeepInfer, the engine behind its production service — 16.3k stars in a day, 2,251 points on Hacker News. This page covers the four things the headlines missed: it is a vLLM fork, not a from-scratch engine; 'NVIDIA SM100' excludes every H100 and every consumer GPU you own; its decode throughput is published under full-acceptance accounting, which counts drafted tokens whether or not they were accepted; and its genuinely important feature — batch-invariant inference — is mutually exclusive with its own speculative decoder.

LLatent Reasoning: Agents That Think Without Words

HybridAgent is the first open-weight agent that does most of its thinking as hidden states instead of tokens — 170-183 latent steps per task against ~300 text tokens, 35.5-54.7% fewer tokens, 2.7-3.7x cheaper. The mechanism (a ~10M-parameter adapter), the reason it loops forever if you load it in llama.cpp, and why 'inspectable latent thoughts' is the most contested claim in the paper.

TThe Third-Party Memory Layer: How claude-mem Actually Works

claude-mem has 103,000 GitHub stars and solves the complaint every coding-agent user has: the agent forgets everything when the session ends. Underneath, it is a second model running in a background daemon that reads every tool call you make — and on the hosted path its own cost dataset maps that observer to a DeepSeek model, not Claude. The six hooks, the exit-code contract that can block your session, the three-layer recall that makes the token math work, where the data lands, who pays for the compression, and a checklist for evaluating any memory layer before you install it.

WWhy a 35 KB System Prompt Breaks on a Local Model

Moving a big Claude/GPT system prompt to Ollama or another local runtime: the context-budget math, Ollama's silent middle-of-prompt truncation, the failure signals of context exhaustion, and the restructuring that actually works.

Next

• Put it to work: Playbooks
• Go deeper: Power User

이 페이지 공유

• X
• LinkedIn
• Reddit
• Hacker News