There is no single 'best LLM for research' — there's a best for each research job. Here's the one-screen answer for the five things a founder actually does research for: reading a stack of papers at once, web research with citations, rigorous reasoning over technical material, cheap high-volume triage, and private work on confidential docs. Plus the trap in each — big context windows aren't perfect recall, and 'cited' answers routinely cite fewer sources than they read.
There is no single 'best LLM for coding' — there's a best for each job. Here's the one-screen answer for the four things a founder actually hires a coding model to do: hard agentic work, cheap high-volume work, self-hosting, and huge-codebase refactors. Plus a warning: the benchmark scores you'll find on most 'ranking' pages contradict each other by 20+ points, and here's how to read them.
The whole reserved-vs-on-demand question collapses to one number: your break-even utilization equals the reserved discount. Here's the rule, the worked math, and when a solopreneur should sign.
Collecting traces isn't the job — closing the loop is. Here's the runnable three-step pipeline that turns a flagged production failure into a human-labeled, versioned regression case, using only Langfuse's SDK and one REST call.
A runaway agent loop bills tokens as fast as the API answers. Here is how to set a real spending ceiling at the gateway — one that rejects the call before it costs you — in LiteLLM and OpenRouter, with the caveat nobody mentions.
NVIDIA's open-source NOOA framework collapses an agent into one plain Python class: methods are its actions, fields are its state, docstrings are the prompt, and type hints are the contract. Here's the full build — install, generation vs deterministic methods, typed state, running it, and the SQLite memory that lets it drop context compaction — with copy-paste code.
Two small lines in the changelog fix two things that used to fail as a mystery. A gateway spend cap now shows the developer the limit, its reset time, and who to ask — and `claude agents` finally prompts for workspace trust in an untrusted directory, the same as `claude` always has. Here's what each one closes and how to set it up.
Every month a cheaper model ships and the group chat says 'switch.' The rate card is the wrong number to switch on: an agent's real cost is tokens-per-task times price times a retry penalty, and only one of those three is on the pricing page. Here's the reusable test — freeze your tasks, measure completed-task cost, decide in an afternoon — with Gemini 3.6 vs 3.5 Flash as the worked example.
Spot GPUs are the same H100s at 60–90% off — until the provider reclaims one mid-job. The discount isn't the number that matters. The notice window is.
A copy-paste setup that wires three coding agents to the same searchable memory over MCP — so a fact one of them learns is a fact all of them know. Ten minutes, one npm package, no API key required.
Kimi K3 tops the Frontend Code Arena but is a rack to self-host and priced like a flagship. The right way to capture the win is task-based routing: send only your UI calls to K3, keep everything else where it is. Here's the router, the cost guardrails, and the math.
The Aug 4 release makes Ollama's /v1/chat/completions streaming match OpenAI's wire format byte-for-byte: role on the first chunk, finish_reason on its own chunk, usage in a separate one. If you kept a fork of your streaming parser for local models, you can delete it.
A $1B acquisition just made 'non-human identity' a real budget line. Here's what it means when your AI agent needs credentials — and the five moves that give it access without handing it a password you can't revoke.
Kitesurf throws out Chromium and runs the whole browser in V8 isolates on Workers. It's 3–7× cheaper on CPU and memory and ~1.7× slower per page. For an agent firing thousands of short page loads, that trade is the point.
Every tool you register rides in the model's context on every turn, so verbose schemas quietly inflate your input bill. Trim each description to its load-bearing job, measure the drop, and A/B for accuracy — the same move that cut a Deep Agents turn's input tokens ~65%.
The complaint is never that skills give bad instructions — it's that they never fire. The one field that decides whether a skill loads is the description, and most are written too vague and too polite. Here's how to write one Claude reliably picks up.
Claude Code's August build moves any main-conversation MCP tool call that runs past two minutes into a background task, so a slow database query or deploy call stops locking up your shell. Here's exactly what changed, the one environment variable that controls it, and when to turn it off.
Qwen3.8-Max shipped on August 3 speaking both the Anthropic and OpenAI wire formats, so you can run your existing agent CLI on it by changing three environment variables. Here's the exact setup — plus the one Codex gotcha that will waste your afternoon.
llm 0.32 shipped a primitive that most agent frameworks make you build by hand: a tool can raise llm.PauseChain to stop the loop before it does something irreversible, hand control back to you, and resume later without re-running the calls that already finished. Here's the exact pattern — pause, persist, approve, resume — in about 40 lines.
Whole-value masking hides a bare token fine — but it corrupts a JWT your code decodes or an AWS key the SDK signs with. Claude Code v2.1.224 adds three structured fields (extract, decode: jwt, awsPairs) that keep the tool working while the agent still never holds the plaintext. Here's the exact config for each.
v2.1.224 (August 7) deleted the hard per-session ceiling that made long orchestrations fail at agent 201. It didn't make fan-out unbounded — it moved the real limits to concurrency, nesting depth, and a budget cap that finally halts running background agents. Here's the new mental model and the three env vars that set it.
OpenAI's Atlas browser stops working August 9 with no automatic data migration. If you wired an agent to it, here's the export checklist and the honest decision between ChatGPT's desktop app, Comet, Claude in Chrome, and the open-source escape hatch.
Sticker prices lie about coding-agent cost, because a single autonomous task burns one to three million tokens — and most of them are input. Here's the real per-task math across the models a founder would actually point an agent at, with verified prices, the two levers that move the bill 5–10x, and which model wins at each budget.
The whole decision comes down to duty cycle — how many hours a day your GPU is actually busy — and how much cold-start latency you can stomach. Here's the break-even line.
Meta's new 'contributor' price for Muse Spark 1.2 is roughly an order of magnitude cheaper than standard — because you pay the difference in training data. Here's the actual math, and a five-question test for whether that trade is fine or a mistake on your codebase.
Two model names that live in older Kimi and Moonshot integrations stop resolving at the end of August. The fix is one string per call — but the like-for-like replacement isn't K3, it's the model you probably overlooked.
On August 5, Meta dropped its first terminal coding agent — Muse Code, powered by the new Muse Spark 1.2 — straight into the space Claude Code and Codex CLI already own. Here's the what, the install line, the benchmarks, and the pricing catch that's getting the most attention.
K-EXAONE 2.0 is Korea's largest model — 750B parameters, 262K context, 10 languages — and the lab that used to ship the most restrictive license in the business just made it Apache 2.0. That's the first frontier-class open weight you can legally fork, fine-tune, and sell without asking anyone. Here's the self-host math and when to actually use it.
On August 5, 2026, Anthropic hard-retired Claude Opus 4.1 — requests to it now error. DeepSeek did the same to deepseek-chat and deepseek-reasoner on July 24. If a model ID is hard-coded in your app, a provider's calendar is your outage calendar. Here's the runbook that keeps a retirement from becoming a page.
Your agent emitted eight tool calls in one turn. Running all eight at once is how you turn a fast turn into a 429 storm. The fix is a bounded semaphore, backoff that honors Retry-After, and returning every result in one message — about 30 lines.
🎧 Listen
The Stack
