// open-source explorer

Decisions open-source code could run on Jev.

// methodology

How we calculate this.

Static estimates from each call's source, list prices and public latency stats. A shadow run on real traffic measures the real numbers.

• state_tokens (the user input each call sends) is unknown from source, so each figure is a range over 50-2000 tokens; the same count is used on both sides.
• prompt_tokens counts only prompt text found in the call's source, without interpolated input. Prompts loaded from files or other modules are missed (promptInSource: false).
• output_tokens is the smallest JSON the schema allows. Reasoning tokens, schema/tool definitions, tool-use system prompts and chat framing are left out of llm_cost, so llm_cost is a lower bound and the reduction is conservative.
• jev_request_tokens = 259 + 31 per question, and state_factor = 0.71: Jev's billed input tokens, fitted from 45 live Jev calls on 2026-09-24 (jev-swap oss measure-jev).
• Prices are list prices from OpenRouter (https://openrouter.ai/api/v1/models, fetched 2026-09-24); Jev's from its OpenRouter page. Calls whose model can't be read or isn't listed there are shown but left out of totals.
• Only production code counts. Calls in tests, examples/demos/samples/docs/cookbooks/templates, evals/benchmarks/scripts/experiments, and in repositories that are sample collections are left out.
• Calls that also return free-text fields Jev can't produce are 'partial-swap' and left out of totals: the LLM call would still be needed.
• Absolute dollars are never shown: traffic volume is unknown, and it cancels out of a percentage.
• Speed: the LLM's time for a call is estimated as OpenRouter's p50 latency plus output tokens / p50 throughput, from the model's busiest provider on its OpenRouter page (a 30-minute window at fetch time). Jev is taken at 327 ms, its p50 over 45 live calls measured end to end on 2026-09-24 (TypeSafe reports 70-500 ms).
• Verdicts: 'cheaper' when Jev is at least 5% cheaper at the low end of the input range; 'about the same cost, faster' when within 15% of the LLM's cost and at least 1.5× faster; otherwise no saving.
• Calls whose model is chosen at runtime are compared with likely models: ones the same repo names in its other decision calls, else the provider's small default models (data/reference-models.json: hand-picked, not usage data), which are also the cheapest, so the savings shown are the conservative ones. Only models where Jev is cheaper, or about the same cost and faster, are shown; these conditional figures stay out of the headline totals.

Jev at $0.042 per 1M input tokens, output free (OpenRouter, checked 2026-09-24). Input-size range: 50–2000 tokens per call.