stack 2/5: price long-context requests at the published long rate (#908)

[lidge-jun]

Stack

2/3 — long-context pricing tiers

Base: codex/bug-stack-plan (#951)
Next: carried contributor bug fixes (#952)

Summary

Several vendors reprice the entire request once the prompt crosses a token threshold. Cost4 is flat and resolveMatchedPrice() never saw a token count, so there was nowhere to express "this rate depends on how big the prompt is" — every request billed at the short rate, including the long ones, which are the expensive ones.

• new ContextTier registry with exact provider+model rules, each carrying its source URL and verifiedAt
• tier selection between base-price resolution and calculateCost(); resolveMatchedPrice() stays token-independent so its provider/model memoization is untouched
• contextTier surfaced on AttemptCostEstimate and CostEstimate, propagated to combo results
• base prices for the three -pro virtual aliases

Model	Threshold	Operator	Effect
gpt-5.6-sol / -terra / -luna (+ -pro)	272,000	>	2× input, 2× cached, 2× cache write, 1.5× output
grok-4.5	200,000	>=	2× uniform
MiniMax-M3	512,000	>	2× uniform

Verified 2026-08-03 against the published tables — OpenAI's >272K is exclusive, xAI's ≥ 200k is inclusive.

Three things worth reviewing closely

The threshold reads raw usage.inputTokens, not normalized input. normalizeCostTokens() subtracts cache read/write to produce billable input, so a 280k prompt with a 200k cache read has 80k billable input and still crosses OpenAI's boundary. Deciding after normalization would have silently under-billed exactly the cache-heavy long requests. Covered by L3.

Long context and Fast are mutually exclusive, not composable. An earlier draft multiplied both. OpenAI's Fast guide states plainly that "Long context, fine-tuned models, and embeddings are not supported", so that product cannot exist. Exclusivity keys on the response-confirmed tier: a >272k request merely tagged priority was necessarily downgraded and must still bill long — suppressing the tier there would under-bill the downgraded request. That required passing tier provenance rather than the collapsed effectiveServiceTier() scalar to all four estimator call sites. Covered by L8.

MiniMax casing is exact on purpose. The bundle carries both minimax-m3 (0.6/2.4/0.12/0) and MiniMax-M3 (0.3/1.2/0.06/0); case-folding would select the wrong base row. Covered by L5.

A separate bug this surfaced

gpt-5.6-sol-pro, -terra-pro, -luna-pro had no base price at all — the virtual resolver keeps the selected id in the usage log while cost resolution deliberately does not fall back through resolvedModel. A probe against the real resolver returned null for all three, meaning -pro usage rendered no cost estimate whatsoever, and a context-tier row alone could never have been reached. Base rows added; L10 covers both halves.

Not fixed here

Terra and Luna still carry pre-price-cut base rates — that is #907, and it cannot be fixed in this repository (canonical models.json lives in lidge-jun/jawcode). Note for whoever lands it: PRIORITY_MULTIPLIERS stores Fast pricing as ratios calibrated against the stale bases, so correcting them without recomputing those ratios would fix an overcharge and introduce an undercharge.

Verification

• bun x tsc --noEmit — exit 0
• bun test tests/usage-cost.test.ts — 51 pass, 0 fail
• red-green: ablating applyContextTier() fails 8 of the new tests
• bun run test — 7691 pass, 8 skip, 0 fail, 507 files
• bun run privacy:scan — passed

The existing Fast fixture used a 1M-token prompt that now crosses the threshold; it moved to 200k so those cases still isolate the Fast multiplier, with totals recomputed.

Fixes #908

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

🗂️ Base branches to auto review (2)

• ^dev$
• ^preview$

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: ASSERTIVE

Plan: Pro Plus

Run ID: a44debeb-e3ff-4023-b54b-b28bf56ec1ba

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

_{Comment @coderabbitai help to get the list of available commands.}

[lidge-jun]

Stack navigation

1. stack 1/5: triage the open issue surface and lock the bug plan #951 — triage the open issue surface and lock the bug plan (base dev)
2. stack 2/5: price long-context requests at the published long rate (#908) #952 — long-context pricing tiers, Cost estimates ignore published long-context pricing tiers (OpenAI >272k, xAI >=200k) #908 (base stack 1/5: triage the open issue surface and lock the bug plan #951)
3. stack 3/5: carry six contributor bug fixes with authorship intact #953 — carry six contributor bug fixes (base stack 2/5: price long-context requests at the published long rate (#908) #952)

Review and merge bottom-up. Each PR targets the preceding stack branch, so its Files changed view contains only that layer.

The layers touch disjoint files — devlog/, src/usage/, and the carried contributors' paths — so any layer can be retargeted to dev and taken independently without a rebase conflict if you prefer.

Carried in #953, with authorship preserved: #939, #942, #943, #944, #945, #948.

[chatgpt-codex-connector]

💡 Codex Review

opencodex/src/usage/cost.ts

Line 334 in c72dc99

if (isConfirmedFast(tier)) return [cost4, undefined];

Scope Fast suppression to the priced attempt

In a mixed-provider combo, the request-level tier context is passed to every attempt even though attempts carry no service-tier provenance. If the final OpenAI response confirms priority, this unconditional return also suppresses long-context pricing for earlier xAI or MiniMax attempts; for example, a preceding 600K MiniMax-M3 attempt is charged at the short rate despite OpenAI Fast restrictions being irrelevant to it. Gate suppression on the current provider/model being eligible for OpenAI Fast, or persist and use service-tier provenance per attempt.

---

opencodex/src/usage/cost.ts

Lines 381 to 382 in c72dc99

	const [tieredCost4, contextTier] = applyContextTier(
	price.cost4, attempt.provider, attempt.model, attempt.usage.inputTokens, serviceTier,

Select context tiers before aggregating upstream calls

The usage passed here can represent multiple upstream requests rather than one pricing decision: runWithImageBridge() explicitly sums every hidden and final iteration into one OcxUsage, then stores that aggregate on the active attempt. Applying the threshold to that sum overcharges media-loop requests when, for example, two xAI calls each use fewer than 200K input tokens but together exceed 200K; neither individual call qualifies for the long rate, yet the estimator doubles both. Preserve per-send usage and apply the context tier to each upstream call before summing its cost.

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Include the generic GPT-5.6 alias in context tiers

The OpenAI API catalog exposes gpt-5.6, and the repository's API contract records that this upstream alias routes to Sol, but this exact-match list begins at gpt-5.6-sol. Consequently, findContextTier() returns no rule for API-key requests selecting the generic alias, so prompts above 272K still use the short rate—for example, 300K input plus 20K output remains $2.10 instead of $3.90. Include gpt-5.6, preferably by deriving this list from the canonical provider registry rather than maintaining another independent model list.

AGENTS.md reference: src/AGENTS.md:L18-L18

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Document the user-visible long-context pricing rules

This registry changes the estimated amounts displayed by the Logs and Usage surfaces whenever OpenAI, xAI, or MiniMax prompts cross the new thresholds, including the non-obvious Fast-mode exclusivity rule, but the commit contains no docs-site/ update explaining those calculations. Add the thresholds and estimation behavior to the English documentation and keep the translated versions consistent so users can interpret the changed dashboard totals.

AGENTS.md reference: AGENTS.md:L212-L213

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Apply Fast multipliers to newly priced Pro aliases

These rows make the *-pro selections priceable while their selected suffix remains in the usage log, but applyPriorityMultiplier() consequently looks up names such as gpt-5.6-sol-pro in PRIORITY_MULTIPLIERS, which only contains the base slugs. The API-key provider still sends service_tier=priority after rewriting the virtual selection to its base wire model, so a response-confirmed Fast Sol Pro request with 200K input and 20K output is reported as $1.60 instead of the $3.20 Fast rate. Resolve the multiplier through the virtual model's base ID, or derive corresponding alias entries and cover them with a Fast regression test.

AGENTS.md reference: src/AGENTS.md:L18-L18

Useful? React with 👍 / 👎.

[lidge-jun]

Stack navigation

1. stack 1/5: triage the open issue surface and lock the bug plan #951 — triage the open issue surface and lock the bug plan (base dev)
2. stack 2/5: price long-context requests at the published long rate (#908) #952 — long-context pricing tiers, Cost estimates ignore published long-context pricing tiers (OpenAI >272k, xAI >=200k) #908 (base stack 1/5: triage the open issue surface and lock the bug plan #951)
3. stack 3/5: carry six contributor bug fixes with authorship intact #953 — carry six contributor bug fixes (base stack 2/5: price long-context requests at the published long rate (#908) #952)
4. stack 4/5: keep an explicit thinking disable through translation (#545) #954 — Claude Desktop classifier thinking round-trip, Claude Desktop 3P Auto Mode classifier retries after 64-token Anthropic OAuth outputs #545 (base stack 3/5: carry six contributor bug fixes with authorship intact #953)

Review and merge bottom-up. Each PR targets the preceding stack branch, so its Files changed view contains only that layer.

The layers touch disjoint files — devlog/, src/usage/, the carried contributors' paths, and src/claude/ + src/adapters/anthropic.ts — so any layer can be retargeted to dev and taken independently without a rebase conflict.

#954 needs human security review per MAINTAINERS.md: it changes request construction on an Anthropic OAuth execution path. It is last in the stack so the first three can land without waiting on that.

Carried in #953, with authorship preserved: #939, #942, #943, #944, #945, #948.

[lidge-jun]

Stack navigation

1. stack 1/5: triage the open issue surface and lock the bug plan #951 — triage the open issue surface and lock the bug plan (base dev)
2. stack 2/5: price long-context requests at the published long rate (#908) #952 — long-context pricing tiers, Cost estimates ignore published long-context pricing tiers (OpenAI >272k, xAI >=200k) #908 (base stack 1/5: triage the open issue surface and lock the bug plan #951)
3. stack 3/5: carry six contributor bug fixes with authorship intact #953 — carry six contributor bug fixes (base stack 2/5: price long-context requests at the published long rate (#908) #952)
4. stack 4/5: keep an explicit thinking disable through translation (#545) #954 — Claude Desktop classifier thinking round-trip, Claude Desktop 3P Auto Mode classifier retries after 64-token Anthropic OAuth outputs #545 (base stack 3/5: carry six contributor bug fixes with authorship intact #953)
5. stack 5/5: probe reset-derived cooldowns without waiting to be selected (#915) #955 — cooldown early-recovery probe, [Bug]: Reset-derived cooldowns can miss early recovery while another pool account remains eligible #915 (base stack 4/5: keep an explicit thinking disable through translation (#545) #954)

Review and merge bottom-up. Each PR targets the preceding stack branch, so its Files changed view contains only that layer.

The layers touch disjoint files — devlog/, src/usage/, the carried contributors' paths, src/claude/ + src/adapters/anthropic.ts, and src/codex/ — so any layer can be retargeted to dev and taken independently without a rebase conflict.

#954 needs human security review per MAINTAINERS.md: it changes request construction on an Anthropic OAuth execution path. #955 sits above it in the chain but is independent of it in code, so if that review blocks, #955 can be retargeted to #953 without conflict.

Carried in #953, with authorship preserved: #939, #942, #943, #944, #945, #948.