[007harshmahajan]

Design at: https://github.com/openobserve/designs/blob/main/llm-obervability/model-pricing-tier-rules.md

Pairs with openobserve/openobserve#15237, which adds rules and priority to pricing tiers.

Landing order

Do not merge until a release containing openobserve#15237 is the oldest supported release. Released editors up to v1.1.0 offer Clone on built-in rows and turn every rule-only tier into an input > 0 tier (dropping rules), so a clone made after this merges would price every span at the first rule tier's rate: 2x on OpenAI Fast mode entries, 0.5x on GPT-5 Nano / GPT-5.4 Nano (Flex), 1.8x on Gemini Flash (Priority). No catalogue-only layout avoids it. Tier selection on old binaries is unaffected (see the design's Compatibility section).

What changes

Tier rules (the purpose of this PR)

• The 77 draft param_conditions tiers (OpenAI Flex / Fast / Ultrafast, Gemini Priority, Anthropic fast mode) become rules. service_tier reads the response keys first (what was billed), then the request keys, then param.service_tier; Anthropic speed likewise. None of these tiers are on main today.
• Three combined tiers settle cross-dimension ties: GPT-5.5 and GPT-5.4 "Fast mode · Extended Context", GPT-5.5 Pro "Flex · Extended Context". OpenAI does not list long-context Fast/Flex rates for these; the prices follow the existing pattern (extended × rule tier ÷ default) and are only reached by such a span.
• Ten entries now list the unrestricted default tier first (GPT-5.6 Sol/Terra/Luna, Claude Sonnet 4.6/4.5/4, Gemini 3.1 Pro/3 Pro/2.5 Pro, Gemini 1.5 Flash 8B); the editor treats tier 0 as the default. Old-binary selection is unaffected.

Catalogue updates carried in the same PR

These are live on every deployed binary as soon as this merges (they need no rules support). All checked against the providers' pricing pages on 2026-10-15/16.

• New models (10): GPT-6.1 Sol, GPT-6 Astra, GPT-6 Sol, GPT-6 Luna, Claude Opus 5.5, Claude Sonnet 5.5, Claude Fable 5.1, Claude Mythos 5.1, Gemini 3.8 Flash, Gemini 3.7 Flash.
• Price changes:
  • GPT-5.6 Sol: $5/$30 → $4/$20 (Extended $10/$45 → $8/$30). This is OpenAI's promotional price, documented "through at least November 21, 2026".
  • Gemini 3.6 Flash: $1.50/$7.50 → $0.75/$3.75. Google lists $1.50 again from 2027-01-01.
• New rule-less "Extended Context (272k+)" tiers: GPT-5.5, GPT-5.4, GPT-5.5 Pro, GPT-5.4 Pro (2x input, 1.5x output, per OpenAI).
• match_pattern changes (17): GPT-5.5, GPT-5.4 (stop matching their -mini/-nano/-pro siblings), GPT-5.2, GPT-5, GPT-4.1, Gemini 2.0 Flash (dated snapshots), o1 (-preview), GPT-4 Turbo (-preview variants), Claude Opus/Sonnet/Fable/Mythos 5 (anchored so 5.x does not match), Claude Sonnet 4 (-latest), Claude 3.7/3.5 Sonnet and 3.5 Haiku (. as separator), Gemini 1.5 Flash (anchored).

Validator and CI

scripts/validate_llm_pricing.py, run by .github/workflows/validate-llm-pricing.yml on every change to llm_pricing.json:

• Mirrors the backend's parsing and validate_tiers (types, priority within ±1000 as an integer, window bounds, condition fields, rule caps in bytes, bare usage. keys, presence operators on usage. keys, a single unrestricted tier), so an entry the sync would silently skip fails CI instead.
• Enforces the old-binary layout: default tier first; a tier with rules plus a condition follows a rule-less tier with the same condition.
• Checks tier selection in the definition's own mode: with ranking (any rules or priority), no two tiers may match one span with equal rank unless a covering tier outranks both; with first-match, no tier may be unreachable behind an earlier one.

Follow-ups (dated)

Built-in entries cannot carry valid_from, so these need catalogue edits on the day:

• GPT-5.6 Sol back to $5/$30 (Extended $10/$45) when OpenAI ends the promotion (no earlier than 2026-11-21).
• Gemini 3.6 Flash back to $1.50/$7.50 on 2027-01-01.

Known gap

Gemini reports a Priority → standard downgrade only in the x-gemini-service-tier HTTP response header, which spans do not carry, so a downgraded request is priced at the Priority rate (documented in the design).

🤖 Generated with Claude Code