feat: add OAuth 2.1 support for Claude.ai Connectors

[acedatacloud-dev]

Problem

OpenAI models return cached_tokens in prompt_tokens_details, but _chat_models.json had no discount terms for cached tokens on any OpenAI model. This caused overcharging when the upstream returned cached tokens:

Model	cached_tokens	Expected Cost	Actual Billed	Overcharge
gpt-4o	3712	~0.000303	~0.000583	92.4%
gpt-4.1-mini	3840	~0.000153	~0.000204	33.1%

Root Cause

OpenAI's prompt_tokens includes cached_tokens (unlike Claude where prompt_tokens excludes them). So the discount must be applied as a negative offset — subtracting the discount portion of cached tokens from the total.

Changes

cost/api/_chat_models.json — 19 model rules modified

Each rule's consumption formula now includes a third term:

cached_mult * var("prompt_tokens_details.cached_tokens", 0)

Where cached_mult = -(prompt_mult * discount_rate), so it subtracts the discounted portion.

50% discount (matching OpenAI pricing):

• gpt-4o, gpt-4o-mini, gpt-4o-2024-05-13
• o1, o1-mini, o1-pro, o3-mini

75% discount (matching OpenAI pricing):

• gpt-4.1, gpt-4.1-mini, gpt-4.1-nano
• o3, o3-pro, o4-mini
• gpt-5, gpt-5-mini, gpt-5-nano, gpt-5.1, gpt-5.4, gpt-5.4-pro

Also added official_price.cached_tokens to each rule for transparency.

tests/api_cost/test_claude_cache_tokens.py

• Replaced test_non_claude_model_unaffected (tested OLD behavior) with two new tests verifying cached_tokens discount IS now applied
• Added TestOpenAICachedTokens class with 14 test cases covering all discount tiers, edge cases (zero cached, all cached, cached_creation_tokens ignored)
• All 642 tests pass

Backwards Compatibility

When cached_tokens = 0 (or absent), the new term evaluates to 0, so billing for non-cached requests is completely unchanged.