video youtu.be Watch ↗

Ollama + Claude Code = 99% CHEAPER

TL;DR

Claude Code is a harness, not a model — "the car, not the engine" — and the engine is swappable through documented environment variables, so the same tool loop can be driven by a local Ollama model or an OpenRouter-hosted open-weight model instead of Anthropic's API. The video demonstrates both paths end to end and then, more usefully, demonstrates their costs: a 9B local model takes four minutes to answer "what do you know about my project," silently drops its tool-call visibility, and lies about its own context window; and a partial env-var override leaves Claude Code's sub-model slots pointed at Anthropic, so tool calls and file searches keep billing paid Haiku while the operator believes they are running free — the presenter's own OpenRouter logs show exactly that. The two durable ideas underneath the tutorial are that a harness encodes an implicit contract the substituted model must satisfy (tool training, context window large enough for the system prompt, JSON protocol conformance) and that "free" only relocates the bill — into hardware, a subscription, a VPS, latency, or lost observability. The presenter's own landing place is not the title: he ends recommending a cheap-but-paid model (a claimed 14¢/40¢ per million tokens against Opus 4.6's $5/$25) for "50 to 100x cheaper" rather than free, and scopes open models to low-stakes, high-volume work — first-pass triage, summarize-before-passing-up, repetitive scaffolding — which is Model-Tier Routing with a new, self-hostable bottom tier.

Concepts introduced

Harness / Model Separation — the agent harness (system prompt, tool definitions, loop, permissions) is a separable artifact from the model that drives it; "the car and the engine."
Harness / Model Fit — separability is not free. A harness encodes an implicit contract — tool training, a context window that fits its system prompt, protocol conformance — and a substituted model can violate it while still appearing to run.
BYO Inference Endpoint — the config seam that makes the swap real (ANTHROPIC_BASE_URL + ANTHROPIC_AUTH_TOKEN), and its sharp edge: a partial override silently leaves sub-model slots on the vendor's paid defaults.
Open-Weight Capability Gap — the closing-but-nonzero distance between open-weight and frontier closed models, and the capability-per-parameter axis that decides what you can actually self-host.
Total Cost of Inference — "there's really no such thing as free": eliminating the token bill relocates the cost into hardware, a subscription, latency, capability, or observability.

Held, not dropped — themes the capture touches that do not warrant a concept page yet:

Agent observability as a model-dependent property. On the 9B local model, Claude Code stopped streaming tool calls ("it basically just spins forever"); on larger models the calls reappeared. Whether step-visibility is a harness feature or a model capability is a real question — held pending a second source.
Context-window labels that lie. Ollama reporting 200k while serving far less, fixed by a Modelfile rebuild. Sharp and useful; currently one vendor's default, not yet a general claim.
Rate limits as the real free-tier currency. OpenRouter: 50 requests/day at zero balance, 1,000/day once $10 is loaded — money buys throughput, not tokens.
Loss of routing control as the price of a free router. openrouter/free dodges rate limits by choosing the model for you.
Privacy as an independent reason to go local, distinct from cost. Mentioned once, not developed.
Using the coding agent to configure its own substrate — asking Claude Code which model size fits your hardware, and to write the Modelfile command. A small, recurring self-reference worth watching for.
The $5 Anthropic credit minimum required to reach the API-key path before you can override it away.

Key claims

An agent harness and the model that drives it are separable artifacts; the harness supplies the tools, the loop, and the instructions, and the model is swappable underneath it. principle — Harness / Model Separation
A harness encodes an implicit contract on the model — trained on its tools, a context window that fits its system prompt, conformance to its JSON tool-call protocol — and a model that misses it will "seem to misbehave" rather than fail cleanly. principle — Harness / Model Fit
Removing the token bill does not remove the cost; it relocates it into hardware, a subscription, a VPS, latency, capability, or observability. principle — Total Cost of Inference
When redirecting Claude Code to a third-party endpoint, override every model slot — the top model, the Sonnet/Opus/Haiku defaults, the small-fast model, and the subagent model — not just the base URL and token. best practice — context: the vendor's harness falls back to its own paid models for tool calls, file searches, and subagents; the presenter's OpenRouter logs show Claude Haiku 4.5 billed under the app name "Claude Code" while he believed he was running free. The context that makes this "best" is a harness with named sub-model slots — the specific variable names are Claude Code's and will not transfer. BYO Inference Endpoint
Scope open-weight models to low-stakes, high-volume work — first-pass triage, summarize-before-passing- up, grepping a codebase, repetitive scaffolding — and re-check anything consequential with a stronger model. best practice — context: the current capability gap on hard, unrecoverable tasks; the video states this scoping is because the gap exists, so it loosens as the gap closes. Model-Tier Routing
Raise a local model's context window before judging it — an undersized window presents as state loss and lost tool-call visibility, not as an error. best practice — context: Ollama defaults that advertise a larger window than they serve; the same 9B model went from a 4m failure to a 2m 5s success on a -64k rebuild. Harness / Model Fit
Prefer a cheap-but-paid model over a free one when you need predictable routing: the free router picks the model for you, and free tiers are metered by requests per day and per minute. best practice — context: sustained agentic use, where per-request rate limits bind before cost does. Total Cost of Inference
Capability per parameter, not capability alone, decides what you can self-host — a small model at a given Elo is strictly more useful to a local operator than a large one at the same Elo. principle — Open-Weight Capability Gap
The video states that on SWE-bench Verified (charted from Vals.ai, SWE-rebench, and official tech reports, early 2026), Qwen3.5-397B (open weight) scores above GPT-5.1 and below Claude Sonnet 4.6/4.5, and Claude Opus 4.6 leads; and that Sonnet 3.7 now falls out of the top five, below Kimi K2 Thinking. observation — the source's charted claim; groundable, not verified here. Open-Weight Capability Gap
The video states that Google's Gemma 4 models (26B-A4B and 31B, "thinking" variants) reach ~1440–1452 Elo at 20–30B parameters, matching models 10–30× larger, and that this is what prompted the video. observation — the source's charted claim; groundable, not verified here. Open-Weight Capability Gap
The video states that Gemma 4 31B Instruct is priced at $0.14 per million input tokens and $0.40 per million output on OpenRouter, against Opus 4.6 at $5 and $25, yielding "50 to 100x cheaper" Claude Code. observation — the source's on-screen claim; groundable, not verified here. Total Cost of Inference
The video states that OpenRouter allows 50 free-model requests per day at zero account balance and 1,000 per day once $10 is loaded, and that a $5 Anthropic credit purchase is required to reach the API-key path at all. observation — the source's claim; groundable, not verified here.
The video states that pointing Claude Code at a non-Anthropic model is not against Anthropic's terms of service, on the reasoning that "we're using their agent harness, we're just plugging in a different model." observation — asserted on a title card with no citation; a licensing question, groundable, explicitly not adjudicated here. Worth a targeted check before an operator relies on it.

Why this is novel — and what it builds on

Nothing in this vault yet names the harness/model seam. Agent Loop describes the loop, and Agent-Computer Interface (ACI) describes the tools as the agent's interface to the world, but both assume a model that was trained for the harness it sits in. This source pulls the two apart and shows what leaks out of the joint: the tools survive the swap, the capability behind the tools does not (the Web Search tool fires, returns "Did 0 searches in 42ms," and the model gives up). That is a new node, Harness / Model Fit, and it is the reason Harness / Model Separation is interesting rather than trivial.

Secondarily this builds on Model-Tier Routing. That page routes sub-roles across tiers within one vendor's ladder plus a cross-vendor bridge; the "when to actually do this" slide is the same discipline with a self-hostable bottom tier, and the same task shape ("summarize files before passing to a smarter model," "grep a codebase," "repetitive scaffolding") appears independently here. Two sources now converge on "route each sub-role to the cheapest tier that clears the bar." What this source adds — and what tier routing did not previously carry — is that the cheapest tier has a floor: at 9B parameters the model was cheap and useless, and the operator paid in four minutes of wall-clock and total loss of step visibility. Total Cost of Inference is the sharpened form of that.

It also corroborates Execution Commoditization from a third direction. That page argues a cheap model tying an expensive one on ordinary work is a fact about the task; this source, independently, shows an open-weight model displacing a model that was frontier one generation earlier (Sonnet 3.7, now out of the charted top five) and a 31B model matching 400B ones on Elo. Same conclusion, different evidence — convergence from below rather than a tie at the top. Note the presenter does not draw the strategic corollary; he stops at the cost saving, which is precisely the layer Execution Commoditization calls table stakes.

No claim here contradicts an existing concept page.

Illustrated walkthrough