The price of intelligence became a stack you engineer, not a bill you pay.

A hosted chat interface and OpenAI-compatible API that treats model choice as a per-request decision: for each prompt, Echo decides how much computation to spend, which open-weight models from its pool — GLM-5.2, Kimi K2.7, others — should participate, and how their work gets combined. The origin experiment is the honest part: the author first measured a hindsight oracle, where knowing in advance which models would help made the combination beat every individual model in the pool, and Echo is the attempt to recover that margin without the hindsight. On the author's own eval mix it matches Fable's aggregate result at roughly a third of the inference cost, and the author states plainly that coding and agentic tasks are still unproven territory. 402 points on Show HN, with a rough reception: sign-up before you can try anything, commenters filing it next to OpenRouter Fusion and Sakana Fugu in a crowding ensemble aisle, one reporting the privacy policy permits training on your traffic. On the 07-21 question — if the weights are free, what do you sell? — this is the fifth answer we've logged: sell the allocation. Reach for it when your workload is extraction, summarization, or chat-shaped work billed at frontier prices out of habit. Delete the hardcoded frontier model ID in the config nobody has re-examined since January. Tradeoff: hosted and closed, the evals are the author's own with no agentic evidence yet, and an ensemble you can't inspect turns every wrong answer into a murder mystery — the thread's words, not ours.

Full disclosure: this sat in yesterday's footer as half of a routing pair; today it topped Show HN at 183 points and the README earns the full slot. Cactus post-trains Gemma 4 E2B — the smallest Gemma — with probes inside the checkpoint, so every answer ships with a structured confidence score between 0 and 1, returned as data, never parsed out of the answer text. The routing program is three lines: answer on-device when confidence is high, ask a bigger model when it isn't. By the authors' benchmarks that matches Gemini 3.1 Flash-Lite while routing only 15–55% of queries to it, and the handoff table carries its own bad news in the open: at 4-bit quantization, MMLU-Pro needs roughly 90% handoff — quantize the checkpoint and you quantize the savings away. Install paths cover the cactus SDK, MLX, and Transformers (with a pinned version range because newer releases segfault on the checkpoint), plus a compile-in patch for llama.cpp. MIT, repo four days old, weights on Hugging Face. Beside Echo upstairs, the routing story's two halves shipped the same day: Echo allocates from outside by watching models work; Cactus escalates from inside because the model reports its own doubt. Reach for it when on-device-first is the architecture and the open question is when to pay for the cloud call. Delete the keyword heuristics and regex guardrails bolted onto your current local-model fallback. Tradeoff: one model family so far, the calibration behind that confidence number is the vendor's own claim, and `if confidence < 0.85` means the threshold — and everything it silently swallows — is now yours to own.

An Apache-2.0 coding agent tuned specifically for small local models: pi — the minimal harness with four tools and a ~1,000-token system prompt — plus twenty extensions, thirty skill files, and a benchmark harness, installed with one `npm install -g little-coder`. The research claim, written up on the author's Substack and stated as the author's own: scaffold–model fit is a real variable, and with the right fit a 9.7B Qwen beat frontier entries on Aider Polyglot. The canonical setup is llama.cpp serving Qwen3.6-35B-A3B, but the same launcher drives Ollama, LM Studio, and cloud models when you want them. The architecture note that matters: it doesn't fork pi — pi is a plain dependency, and everything little-coder-specific lives in droppable extension directories, so the tuning is inspectable and swappable rather than baked in. The front page carried the counter-argument the same day — HumanLayer's essay says harness engineering is not enough — and this repo is the empirical bet that it's at least underpriced: the harness is the cheapest layer of the stack to re-engineer and the only one you fully control. Reach for it when you want agentic coding without a per-token meter — on a plane, on private code, or on principle. Delete the assumption that a coding agent starts with a frontier subscription. Tradeoff: tuned means tuned — the extensions and skills are fitted to specific small models and inherit pi's model of the world; Node 22.19+ is required; and the Polyglot result is one author's benchmark harness until someone else reproduces it. 1,873 stars since April.

A YC-backed platform for training small language models whose one load-bearing idea is that the price arrives before the run: Freesolo quotes the full cost of a fine-tuning or reinforcement-learning job upfront, and the workflow is API-first, so an agent holding a budget can adjust the dataset and algorithm until the quote fits. That's the procurement-by-agent pattern (07-16, 07-17) arriving at the training layer, and the training-side sibling of Otari's pre-spend inference budgets (07-07). The vendor's own numbers claim 8x cheaper supervised fine-tuning and 5.5x cheaper GRPO than Tinker, with an environment-hub SDK for building the task environments RL needs. #10 on Product Hunt at 96 votes. In this slate it's the fourth layer down: Echo rents the pool, Cactus escalates out of it, little-coder refits the harness around it — this is for when the right answer is a specialist nobody sells, trained on your data. Reach for it when a narrow, repeated task is burning per-token frontier calls that a trained small model could absorb. Delete the GPU invoice that arrives after the money is spent. Tradeoff: hosted, closed, and enterprise-leaning; the 8x/5.5x figures are vendor benchmarks against a single competitor; and the platform prices the run, not the outcome — whether the specialist is any good remains your eval to write.

The kicker, for the frontier bill you keep: an MIT proxy that wraps Claude Code (`uvx claude-thermos` instead of `claude`) and does exactly one job — stops your prompt cache from expiring while the main agent waits on a long-running subagent. The mechanism is precise: a subagent has a different cache prefix, so its traffic never refreshes the main agent's; past the TTL the cached history dies, and the next turn re-encodes 200–500K tokens at the 1.25x write rate instead of reading them back at 0.1x. Across roughly 185 of the author's sessions, those rebuilds were about 22% of the bill. Thermos watches traffic through a loopback proxy, and when the main lineage idles while a subagent runs, it replays the last request with `max_tokens: 1` — a one-token heartbeat whose entire point is the cache-refreshing prefill. The HN thread did the QA, per the Bento precedent: a commenter who inspected their own traffic reports Pro/Max plans currently carry one-hour cache TTLs, and API users can already pay for one-hour caching — so the tax lands mainly on API-rate sessions at the default five-minute tier — alongside the commons objection that warming spends the provider's capacity to save your money. Reach for it when API-billed sessions run long and subagent-heavy and the invoice has a mystery line that scales with conversation length. Delete the re-encode tax — or at least measure it first, since the tool logs every save it makes. Tradeoff: the premise is provider-configurable and one TTL change upstream obsoletes the tool — the thread's first comment gives it until Anthropic notices, and the best realistic outcome, as another commenter put it, is that it forces the pricing fix that makes it unnecessary.

Off the thread but worth knowing: **pi-web** (MIT, 2,548 stars, plus 315 today) is a local web UI over pi's session files — browse sessions by project, chat beside a file tree, switch git worktrees, and fork a session from any message into a separate route. Two notes: the harness under today's little-coder pick and this UI is the same pi (badlogic's pi-mono — answering a question we left open on 07-16), making this the first day one harness's ecosystem trended twice at once; and fork-from-any-message is the fork verb from the replay watch (07-14) shipping in a real harness UI — for pi sessions, that is; forking a Claude Code transcript from minute twelve remains unshipped. **Pushary** (closed, $9.99/month, iPhone and Android, #3 on Product Hunt at 138 votes) and **Remux** (MIT, TestFlight beta, an iOS-native tmux client built on Ghostty's terminal core, with Claude Code and Codex shortcut palettes) are the pocket pair: the doorbell thread's approval half gone commercial — per-tool policies, QR pairing to your terminal, an exportable audit trail, six harnesses covered — while the terminal half goes native and open. Pushary's per-tool auto-approve policies are the closest anything has come to the scoped permission ceilings the watch has been asking for (07-23). **OneCLI** (Apache-2.0, 2,732 stars since March, 101 points on Show HN) is a credential gateway where agents hold placeholder keys and a Rust proxy swaps in the real ones at request time, AES-256-GCM at rest — enforcement-beats-instruction at the secrets layer: agent-vault hid the keys at exec time (07-06), valv walled the database (07-23), this walls every outbound API call. **Palmier Pro** (GPL-3.0, 11,592 stars since April, Show HN 164 points) is an open-source macOS video editor with built-in AI generation and a local MCP server for your agent — the video-incumbent second-interface thread (FableCut 07-10, aftr 07-13, OpenChatCut 07-21) gaining its biggest open entrant. And **Buzz** completed the launch circuit: Tuesday's unreachable news item, Wednesday's plus-3,252-star repo and our confirmation pick, today's Product Hunt #8 at 105 votes — three days, three distribution surfaces, same product.