[Bug] gpt-4.1 / -mini / -nano are priced as legacy gpt-4 — 15x to 300x over list #13184

Before submitting your bug report

• I've searched open issues — the closest is OpenAI cost calculation ignores cached input tokens (over-costs cache hits) #13104 (cached input tokens not discounted). This is a different defect in the same function, and it survives the fix proposed there.

Relevant environment info

• OS: any
• Continue version: main @ 5522c6f
• Model: gpt-4.1, gpt-4.1-mini, gpt-4.1-nano (any gpt-4.1* id)
• Surface: CLI cost display + telemetry, and the GUI usage panel

Description

calculateOpenAICost picks a pricing row by longest-prefix startsWith. "gpt-4.1" does not start with "gpt-4o", so the loop falls through to the legacy "gpt-4" row — the most expensive row in the table.

core/llm/utils/calculateRequestCost.ts:159-185:

"gpt-4o-mini": { input: 0.15, output: 0.6 },
"gpt-4o":      { input: 2.5,  output: 10 },
"gpt-4-turbo": { input: 10,   output: 30 },
...
"gpt-4":       { input: 30,   output: 60 },   // <- GPT-4 (2023) list price

const sortedKeys = Object.keys(pricing).sort((a, b) => b.length - a.length);
for (const prefix of sortedKeys) {
  if (normalizedModel.startsWith(prefix)) { modelPricing = pricing[prefix]; break }
}

"gpt-4.1".startsWith("gpt-4") is true, and nothing longer matches first. The gpt-4o family is unaffected — "gpt-4o" is checked before "gpt-4" and wins.

To reproduce

The table and matcher below are copied verbatim from calculateOpenAICost; real prices are OpenAI's list rates from https://developers.openai.com/api/docs/pricing, checked today.

const pricing = {
  "gpt-4o-mini": { input: 0.15, output: 0.6 },
  "gpt-4o": { input: 2.5, output: 10 },
  "gpt-4-turbo": { input: 10, output: 30 },
  "gpt-3.5-turbo-0125": { input: 0.5, output: 1.5 },
  "gpt-3.5-turbo-1106": { input: 1, output: 2 },
  "gpt-3.5-turbo": { input: 1.5, output: 2 },
  "gpt-4": { input: 30, output: 60 },
};
const sortedKeys = Object.keys(pricing).sort((a, b) => b.length - a.length);
const match = (m) => {
  const n = m.toLowerCase();
  for (const p of sortedKeys) if (n.startsWith(p)) return [p, pricing[p]];
  return [null, null];
};
for (const m of ["gpt-4.1", "gpt-4.1-mini", "gpt-4.1-nano", "gpt-4.1-2025-04-14", "gpt-4o"])
  console.log(m, "->", match(m)[0], JSON.stringify(match(m)[1]));

gpt-4.1            -> gpt-4  {"input":30,"output":60}
gpt-4.1-mini       -> gpt-4  {"input":30,"output":60}
gpt-4.1-nano       -> gpt-4  {"input":30,"output":60}
gpt-4.1-2025-04-14 -> gpt-4  {"input":30,"output":60}
gpt-4o             -> gpt-4o {"input":2.5,"output":10}

Expected / actual

A 100k-token gpt-4.1-mini prompt is reported as $3.00 against a real charge of $0.04.

The three date-stamped ids (gpt-4.1-2025-04-14 and friends) hit it too, so pinning a snapshot doesn't avoid it.

Suggested fix

Add the three rows. sortedKeys is longest-first, so "gpt-4.1-mini" (12) and "gpt-4.1-nano" (12) are checked before "gpt-4.1" (7), which is checked before "gpt-4" (5) — no matcher change needed:

"gpt-4.1-mini": { input: 0.4,  output: 1.6 },
"gpt-4.1-nano": { input: 0.1,  output: 0.4 },
"gpt-4.1":      { input: 2.0,  output: 8.0 },

Worth considering separately: "gpt-4" is a bare-prefix catch-all at the highest price in the table, so any future gpt-4.x inherits $30/$60 silently. Requiring the next character to be - or end-of-string would send unknown variants to the null path (no cost shown / CLI fallback) instead of to the most expensive row. That is a behaviour change for unknown ids, which is why I've kept it out of the fix above.

Unrelated to this, but visible from the same table: gpt-5* ids match nothing and return null, so the GUI shows no cost for them at all.

Note on a PR

I can't sign the CLA, so this is an issue rather than a PR — the three lines above are the whole change, and the prices are linked to source. Happy to add the vitest cases as a patch in a comment if that's useful.