Field notes on frontier open models.

• 2026-09-09 Why we built Wallaby — and where we're going Wallaby exists because access to frontier open-weight models is harder than it should be. What we run, what we charge, and reliability you can verify. PinnedCompany
• 2026-10-14 We Paid Claude $10 to Review Our Spec. It Sent Back a Test Harness. One spec, four isolated AI reviewers. Claude Opus 5.5 reviewed it, then built the trigger-test harness it had recommended: 1,600 lines, 30 passing tests, $10.04 on the meter. Playbook
• 2026-10-14 · upd 2026-10-14 We Sent the Same Spec to Four AI Reviewers. They Agreed on Exactly Six Things. One brief, four isolated AI reviewers, six points of convergence, one public disagreement, and why convergent AI opinion is still not validation. Playbook
• 2026-10-14 A $51.77 Refactor: What a Full-Throttle Claude Opus Session Actually Buys Claude Opus 5.5 at full reasoning throttle rewrote our 1,203-line operations script in about an hour: 8/8 equivalence checks, 17 passing tests, one blown $50 credit cap. The itemised account. Playbook
• 2026-10-14 Same Skill, Opposite Effects: It Lifts Weak Models and Trips Strong Ones Four rounds of dual-model, dual-judge experiments on when structured agent skills help, when they do nothing, and when they get in the way. Playbook
• 2026-10-14 Your Skill File Is Too Big: The 6KB Line (Measured) We varied only skill file size across 28 executions. The empty-output rate jumped from 0% to 25% at 6KB, and hit 50% for the naturally bloated file. Measured rules for skill and AGENTS.md authors. Playbook
• 2026-10-14 · upd 2026-10-14 We Pre-Registered a Kill Rule for Our Own Product. It Fired. A falsification report from building a lifecycle tool for AI agent skills: pre-registered predictions, two failed rounds, and the one fix that went 0-for-20 to 1.8. Playbook
• 2026-10-13 Kimi K3 in the Vercel AI SDK: one npm package, two lines of code Add Kimi K3 to the Vercel AI SDK with one npm package: install, first call, streaming, and the actual bill. Verified end-to-end on ai 7.0.138. Tutorial
• 2026-10-13 Muse Remembers. Who Governs the Memory? Meta's Muse Wonder ships Active Memory and Active Focus. Memory is now a model feature. Four things built-in memory still cannot do, and who owns them. Industry
• 2026-10-11 · upd 2026-10-13 Kimi K3 CLI setup: every terminal and coding tool that works, in one table Kimi K3 works in Codex CLI, Claude Code, OpenCode, pi, Crush, Goose, Hermes and more through one OpenAI-compatible endpoint. The full tool table, one-line configs, and links to each verified walkthrough. Guide
• 2026-09-12 · upd 2026-10-12 Kimi K3 in Cline: 5-minute BYOK setup Cline's built-in catalog has included a Wallaby provider since v4.1.19: select Wallaby, paste your key, pick kimi-k3. Verified end to end on Cline v4.1.22. Tutorial
• 2026-10-12 Kimi K3 in the Codex Desktop App: the Custom Provider Setup Nobody Shows in the GUI The Codex desktop app has no settings page for custom model providers — but it shares the CLI's config.toml. Here's the exact setup for Kimi K3, why the model picker says Custom, and the daemon restart that makes it stick. Tutorial
• 2026-10-12 GPT-5.5 Leaves Codex on October 14: The Official Fix, and the Option the Announcement Skips GPT-5.5 retires from Codex on October 14. What to use instead: OpenAI's official migration checklist, the custom-provider door the note skips, and what each option costs per million tokens. Guide
• 2026-10-04 · upd 2026-10-12 Open-weight LLM pricing Oct 2026: Kimi K3, DeepSeek V4 Official per-1M-token prices for the four major open-weight model APIs, fetched from the official pricing pages on Oct 4, 2026 — plus the peak-window, context-tier and cache traps, and what a coding-agent workload actually costs on each. Guide
• 2026-10-11 · upd 2026-10-12 TRAE BYOK, tested: point TRAE at your own OpenAI-compatible endpoint in five minutes TRAE's merged release lets you bring your own model endpoint, Free plan included. We wired ours to an OpenAI-compatible API running Kimi K3 and hit every pothole on the way: the provider list that silently sends you to api.openai.com, a 4054 error that only dies when you abandon the chat, and a retry storm we could see line by line in our own bill. The working setup, the verbatim errors, and the fixes. Guide
• 2026-10-12 REA (Reverse Engineer Anything) on Kimi K3: a verified Claude Code setup Wire the REA reverse-engineering toolkit into Claude Code running Kimi K3 through Wallaby Token, with a real analysis run as the receipt. Tutorial
• 2026-10-05 · upd 2026-10-11 Where to get Kimi K3: every access route compared Kimi K3 access routes compared: official API, inference clouds, gateways, subscriptions, free tiers and self-hosting, with current prices and who each route fits. Guide
• 2026-10-11 ClinePass models vs BYOK: what a month of Kimi K3 in Cline actually costs ClinePass is a flat $9.99/month for curated open models including Kimi K3; BYOK is pay-as-you-go on your own key. With four measured same-task runs, here is the honest breakeven math. Guide
• 2026-10-11 Beam vs Kimi K3: what the challenger's own table shows Reflection AI launched Beam with a 14-benchmark table aimed straight at Kimi K3. We sell Kimi K3, so we read that table closely: both columns, who ran the numbers, and the three things nobody can check yet. Industry
• 2026-10-11 · upd 2026-10-11 TRAE + MCP, tested end to end: setup that works, a loop bug, and shared memory between agents ByteDance just merged TRAE into one app with an Agent workspace. We mounted a zero-dependency MCP server the same night and ran it through IDE chat, SOLO mode, and custom agents. Everything worked except one read instruction that sent the built-in chat into a loop. The setup, the verbatim error, the fix, and the four-tool memory server we built. Guide
• 2026-10-11 Kimi K3 in Crush: config with live cost tracking Run Kimi K3 in Crush, Charm's terminal agent, via an openai-compat provider block, with per-token prices registered so the TUI shows what each session actually cost. Verified on v0.98.1. Tutorial
• 2026-10-11 Kimi K3 in Goose: custom provider setup in 5 minutes Run Kimi K3 in the goose CLI through any OpenAI-compatible endpoint. Two config lines, two env vars, one real 404 pitfall. Verified end to end on goose v1.54.0. Tutorial
• 2026-10-11 The pause tax: what is still leaking from your pi agent's Kimi K3 bill pi 1.0's codemode cut prompt tokens by roughly 40%. What remains is usage-side: cache misses after pauses, untiered thinking, and context that only grows. A metered look at one real session. Playbook
• 2026-09-13 · upd 2026-10-09 Kimi K3 API pricing (October 2026): $3/$15 vs 18 providers Kimi K3 pricing, straight: $3.00/M input, $0.30/M cached, $15.00/M output at official list. What 18 API providers actually charge, and what one task costs. Guide
• 2026-10-08 · upd 2026-10-09 Cheapest way to run Kimi K3 API (October 2026): paid floors, free options, catches The cheapest paid Kimi K3 API access we can verify is $2.70/$0.27/$13.50 per million, an ongoing rate 10% under official list. Free routes exist and every one has a boundary. Here is where each ends. Guide
• 2026-10-08 · upd 2026-10-09 Kimi K3 cached input pricing: $0.30 per million, and when caching pays (October 2026) Kimi K3 context caching, priced: $0.30/M cache hits vs $3.00/M misses, Moonshot's write fees at $3.00 (5m) and $6.00 (1h), the 93% hit-rate math, and when caching does not help. Guide
• 2026-10-08 · upd 2026-10-09 Kimi K3 API price per 1M tokens: the full rate card (October 2026) Kimi K3's API rate card: $3.00/M input, $0.30/M cached, $15.00/M output at official list, who sells it below list, and the arithmetic on one real request. Guide
• 2026-10-09 · upd 2026-10-09 We Paid Claude to Audit Our Blog. Here's the Bill — and What We Shipped. We pointed Claude at our blog through our own gateway and told it to find what was broken. The audit cost $2.50, itemised. Within 24 hours we shipped four of its findings and rejected two suggestions with clear reasoning. Playbook
• 2026-10-09 · upd 2026-10-09 Your Rule File Says \"Later.\" That's Why It Never Happens. Our agent skipped its own registration step three times, and we kept blaming its memory. Three audits later the culprit was a single word: 'after.' Temporal wording in a rule file is not rhetoric: it is control flow. Playbook
• 2026-10-09 Same Prompt, Four Thinking Budgets, Four Bills We ran the same code-review task in Cline at four reasoning_effort tiers and metered every token through our own gateway. The cheapest run caught the headline findings; the most expensive run was the one called 'None'. Playbook
• 2026-09-17 · upd 2026-10-09 Kimi K3 in OpenCode: zero-config, in the catalog Run Kimi K3 in OpenCode with zero config — Wallaby is in the models.dev catalog. One env var, one model refresh, K3 is selectable. Verified end to end. Tutorial
• 2026-10-09 \"Free tier can only be used from within OpenCode\": what it means and three ways out OpenCode's free tier went app-only in September 2026. The 403 FreeTierError reproduced, what changed, and three ways out, including BYOK at 10% off list. Guide
• 2026-10-08 What our own internal workloads cost on Kimi K3: a real September bill (2026) We run our own internal workloads on Kimi K3. September 2026: 3,177 requests, 4.2M tokens, $17.69 total, metered per request. The bill, broken down by task, plus the formula to estimate yours. Guide
• 2026-10-07 Layered Memory for Grok Build: Native Notes + Repo Files for Engineering Teams Grok Build's native memory handles cross-session recall. We layered a repo-file system on top for git-auditable team memory and ran both configurations through the same five tasks: what each layer covers, a phantom-memory boundary, and the habit that fixes it. Playbook
• 2026-10-07 Cline + Kimi K3: every error we've hit, and the fix Every Kimi K3 error we have hit in Cline: Invalid API Response, 128K context fallback, $0.00 cost displays, cost gaps vs your bill, and 429s, with fixes. Guide
• 2026-09-09 · upd 2026-10-07 Cursor "Named models unavailable": why Kimi K3 is blocked Cursor's "Named models unavailable" banner (free plans can only use Auto) is one of two walls for third-party Kimi K3. Both walls, and what works. Tutorial
• 2026-10-06 We Ran Our Memory System Inside Hermes. It Already Had One. Hermes Agent grows its own memory: sessions get distilled into skills over SQLite FTS, tended by a Curator. We installed our file-based memory discipline into v0.21.5 anyway. Three verified mechanics, one caveat from the issue tracker, and a clear answer on which layer does what. Playbook
• 2026-10-06 Run pi with Kimi K3 on Windows or Mac: three commands, no config files Install pi with Kimi K3 on Windows or Mac: one npm extension, three commands, no config files. Verified end-to-end on pi 1.0.4, with billing to match. Tutorial
• 2026-10-06 OpenHands Model Router on your own endpoint: what works, what isn't wired yet, what it costs Day-one field test of OpenHands' Model Router on a custom endpoint: routing works at $0.0025 a call, but two early-build gaps are worth knowing first. Guide
• 2026-10-06 OpenHands with your own endpoint: every error we've hit, and the fix The error map for OpenHands on your own endpoint: the cache_creation_tokens crash, first-chat stalls, $0.0000 metrics, and the settings-layer bugs. Guide
• 2026-10-06 OpenHands v1.25.0: The Settings Layer Grows Up OpenHands v1.25.0 ships 17 features and 33 fixes; 18 of 50 user-facing changes land in one place: LLM profiles, provider connections, the Model Router. Industry
• 2026-10-06 Kimi K3 API with per-request itemized billing Wallaby Token bills every Kimi K3 call as its own line item: input, cached input, output and thinking, in your console seconds after the request. Guide
• 2026-10-06 Kimi K3 API with a free trial credit: $0.50, no card, no top-up Every new Wallaby Token account gets $0.50 of free trial credit for the Kimi K3 API — no card, no top-up. Here's what $0.50 buys and how to make your first call in two minutes. Guide
• 2026-10-06 AI Memory Should Have a Probation Period Trust-building mode: new memories get proposed, not written, until you approve them. Plus a conditional code-version module and a full Chinese edition. Playbook
• 2026-10-05 Kimi K3 API that takes card payment (no crypto) Pay for a Kimi K3 API with a normal bank card: USD card checkout via Stripe, prepaid credits that never expire, and no crypto anywhere in the payment flow. Guide
• 2026-10-05 Kimi K3 API with a public status page: live uptime, 90-day history, JSON feed Wallaby Token publishes a public status page for the Kimi K3 API: per-component uptime over 24 hours, 30 and 90 days, live latency, incident history, and a machine-readable status.json. No login required. Guide
• 2026-10-05 AI Memory Benchmarks Measure the Wrong Thing Every major AI memory system posts green benchmark scores, and every agent user reports the same failures. The gap is structural: benchmarks test retrieval, while production memory dies of rot. Playbook
• 2026-10-05 Why your agent's token bill keeps growing, and how to audit it An agent's daily token spend can double with no change in workload. The causes are measurable once you count per request instead of per day. A field guide with real numbers and a cost simulator. Playbook
• 2026-10-04 SkillRefiner: five stages, zero replays — like ours SkillRefiner refines agent skills offline from historical traces. We run a similar loop in production, built independently: five stages, real cases, and what it taught us back. Playbook
• 2026-10-01 AI Memory Pricing: Supermemory vs Mem0 vs Zep vs Letta AI memory pricing compared from all four official pages — Supermemory, Mem0, Zep, Letta — plus the self-hosted route we benchmarked and the free file option. Playbook
• 2026-10-01 AI Memory: Three Routes, Three Bills Three routes to AI memory — hosted API, self-hosted engine, plain files — and what each one bills you. We benchmarked the hot one, then said no. Playbook
• 2026-10-01 Your Agent Needs Sleep, Not a Bigger Context A stale card on our project board exposed the real bug: we had built our AI a perfect memory, and perfect memory was the problem. Human memory works because it forgets — so we rebuilt ours around sleep, pruning, and myelin. Playbook
• 2026-10-01 Memory Consolidation for AI Agents: A Three-Piece Playbook A working recipe for agent memory that gets more accurate over time: a replay job that decides what survives the day, a pruning pass that deletes on purpose, and a reflex layer that turns repeated procedures into muscle. Playbook
• 2026-10-01 Memory Is Not Written. It's Audited. Write discipline always loosens. Three weeks in, the log claims things are done that the state file still shows in flight — and nobody notices. The fix is not stricter discipline; it's a weekly mechanical diff between what your log claims and what your state shows. Playbook
• 2026-09-30 Microsoft put 1,024 AI agents on one task. What did it cost? Microsoft's Agensh paper scales one task from 1 agent to 1,024 with no orchestrator. The gains are real but sharply sublinear — and the token bill appears nowhere in the paper. Industry
• 2026-09-30 Why did reasoning_effort silently stop working? For 13 days we thought the model ignored reasoning_effort. The truth: our own router was dropping the parameter. A postmortem with logs, retests, and the fix. Playbook
• 2026-09-29 Your API key is probably already on GitHub Searching GitHub for OPENAI_API_KEY returns thousands of committed .env files — many still live. Five boring fixes that actually protect your keys. Guide
• 2026-09-29 It Remembers Because You Close Memory systems don't fail at install — they fail at the end of each session, one skipped write-back at a time. The fix is a two-minute closeout ritual, and it now ships as a paste-in prompt. Playbook
• 2026-09-29 AI Reviews AI: We Planted 7 Bugs in a Rate Limiter and Sealed the Answer Key One agent wrote a token-bucket rate limiter with 7 planted bugs. Another agent — Kimi K3 in OpenHands Agent Canvas — reviewed it. It caught 6, spotted the rigged test, and found 2 bugs we didn't plant. Full scorecard and the $0.16 bill. Tutorial
• 2026-09-28 OpenHands Agent Canvas: Scheduled Slack Digests Wire an Agent Canvas cron automation to Slack via an incoming webhook: a self-contained prompt, a real verified run, metered cost, and two environment notes. Tutorial
• 2026-09-28 The Authors Guild filings are an API procurement problem Unsealed briefs in Authors Guild v. OpenAI quote executives admitting they trained on pirated books. Two questions every API buyer should now ask their provider — provenance and retention. Industry
• 2026-09-25 OpenHands 'Failed to load provider connections': fix OpenHands 'Failed to load provider connections'? Two agent-servers sharing ~/.openhands use different OH_SECRET_KEY values, so saved keys read empty. Fix: one value per machine. Playbook
• 2026-09-25 How to Use Kimi K3 in pi (One JSON Block) How to use Kimi K3 in pi: one JSON block in ~/.pi/agent/models.json, one exported key, done. Verified end-to-end on pi 0.86.1 with billing logs to match. Tutorial
• 2026-09-25 Your AGENTS.md Is Rotting (Here's the Fix) AGENTS.md rots by default: stale decisions, invented paths, confident lies. Structure won't fix it — a maintenance layer will. Here's ours, in production. Playbook
• 2026-09-25 OpenHands 'cache_creation_tokens' AttributeError: fix OpenHands: 'PromptTokensDetailsWrapper' object has no attribute 'cache_creation_tokens' — an SDK/LiteLLM version-skew bug, not your endpoint. Cause and fix inside. Tutorial
• 2026-09-23 OpenHands stuck at 'Running task'? The first-chat stall OpenHands stuck at 'Running task' / 'Waiting for task'? The call succeeded — the UI event channel is rate-limited. Fix: stop, start a fresh conversation. Why Metrics shows $0.0000, too. Tutorial
• 2026-09-21 AI remembers, you find: fixing AI memory rot AI memory rots mid-project: settled decisions get relitigated, files get invented. Wallaby-agent-rules v2 moves memory to plain, dated, sourced files. Playbook
• 2026-09-17 One endpoint, one bill: Kimi K3 for dev teams One endpoint, one bill: give every developer their own API key with a hard budget cap, all from one shared prepaid balance — no seat licenses, no surprise. Tutorial
• 2026-09-17 Union Alpha Free in OpenCode: how to try it Union Alpha Free in OpenCode: zero-cost model, 262k context, no key required. How to enable it, what we verified, and honest caveats before real work. Tutorial
• 2026-09-14 How to compare LLM API costs: measure cost per accepted output LLM API cost comparison goes past price per million tokens: retries and human fixes are the bigger bill. Measure cost per accepted output in one afternoon. Playbook
• 2026-09-13 AGENTS.md as a token budget: the forbidden list AGENTS.md is a Linux-Foundation standard read by 30+ coding agents. Used well it's a token budget — our full production playbook, forbidden list included. Playbook
• 2026-09-13 Kimi K3 in Codex CLI: profile-file setup Run Kimi K3 in OpenAI Codex CLI through Wallaby's API, using the profile file — one config file, verified end to end with a real coding task. Tutorial
• 2026-09-12 Kimi K3 in VS Code Copilot via BYOK: setup Point VS Code's Bring Your Own Key endpoint at an OpenAI-compatible API and run kimi-k3 in Copilot Chat agent mode — full setup with working configuration. Tutorial
• 2026-09-10 Kimi K3 in Hermes Agent: 4 lines of YAML Point Hermes Agent at the kimi-k3 reasoning model through Wallaby Token's OpenAI-compatible API. Verified end-to-end — chat and tool use both work. Tutorial
• 2026-09-09 Top up Wallaby Token: card, Apple & Google Pay How to add credit to your Wallaby Token account with a card, Apple Pay, or Google Pay — plus the pitfalls we hit ourselves when testing. Guide
• 2026-09-09 Kimi K3 in Claude Code: 5-minute setup Configure Claude Code to run Moonshot AI's kimi-k3 model through Wallaby Token's Anthropic-compatible endpoint, verified end-to-end. Tutorial

No posts match. Try a different keyword or tag.