Add GPT-5.6 Sol/Terra/Luna Codex pricing

[0xSMW]

Summary

• Add bundled local-cost pricing for the three released GPT-5.6 tiers and map the unsuffixed public alias to the default tier.
• Cover the complete published surface used by CodexBar: standard short-context, >272K long-context, cache-read, cache-write, output, and Priority/Fast pricing.
• Preserve OpenAI's boundary that Priority excludes long-context requests: traced >272K rows stay attributed to the Priority bucket but use the published Standard long-context base cost, with no invented combined rate.
• Keep models.dev authoritative when it supplies values. Fill bundled short-context omissions only, and use the bundled long-context tuple only when the catalog has no context block at all.
• Fingerprint every effective pricing input. A changed catalog now invalidates unchanged Pi session caches and reparses every message, while a fetched-time-only change remains cache-stable.
• Preserve contributor credit and document the user-visible fix in the changelog. Thanks @0xSMW!

Published rates (per 1M tokens)

Tier	Input	Cached	Cache write	Output	Long input	Long cached	Long write	Long output	Priority input	Priority cached	Priority write	Priority output
Sol	$5	$0.50	$6.25	$30	$10	$1	$12.50	$45	$10	$1	$12.50	$60
Terra	$2.50	$0.25	$3.125	$15	$5	$0.50	$6.25	$22.50	$5	$0.50	$6.25	$30
Luna	$1	$0.10	$1.25	$6	$2	$0.20	$2.50	$9	$2	$0.20	$2.50	$12

Source: OpenAI pricing and the released model cards.

Verification

• Rebased exact head 615276b8 on current main (5bee0172)
• Focused pricing, scanner, models.dev, and unknown-model suites: 112 tests passed
• make check
• Full make test: all 595 selections and 50 groups passed; one slow unrelated group exceeded its batch budget and all 12 selections passed on automatic isolated retry
• Independent exact-head review and range-diff review: clean
• Signed debug package: exact commit prefix embedded, Developer ID signature verified, packaged CLI reports CodexBar 0.41.1
• Exact packaged behavior proof with isolated synthetic Pi JSONL and no credentials, Keychain, browser state, or provider requests
  • 272,000 prompt tokens: $0.59625
  • 272,001 prompt tokens: $1.185005
  • rate-only catalog change repriced an unchanged Pi file from $1.781255 to $3.56251 and changed the pricing key
  • fetched-time-only change kept cost, pricing key, and scan timestamp stable

The changed behavior is local cost calculation, so the packaged CLI is the closest deterministic live surface; no provider or menu interaction is required.

Notes

• Native Codex session events currently expose input, cached-input, and output counts. Cache-write billing applies wherever a source, such as Pi, supplies write counts.
• Public Model Identifier Gate: PASS. Production names are released canonical identifiers; dated test variants are explicitly fictitious.
• No dependencies added or updated.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 1733c703cd

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".

[0xSMW]

Review feedback addressed (e20f9a8)

All four Codex review threads were verified against OpenAI pricing / GPT-5.6 Sol model card and fixed:

Thread	Verdict	Fix
Cache-write 1.25×	Correct — Pi folded writes into input at 1×	Optional cache-write rates on CodexPricing; codexCostUSD(…, cacheWriteInputTokens:); Pi passes writes separately
Long-context >272K	Correct — 2× in / 1.5× out full request	thresholdTokens + above-threshold rates for Sol/Terra/Luna
Priority rates	Correct — Priority tab shows 2× input	Priority input/output/cache-read rates for all three tiers
gpt-5.6 alias	Correct — alias routes to Sol	normalizeCodexModel → gpt-5.6-sol

Verification

swift test --filter CostUsagePricingTests   # 38 passed
swift test --filter PiSession               # 11 passed

Threads replied + resolved. Ready for re-review.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: e20f9a8f16

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".

[0xSMW]

Second review round addressed (8d2d038)

Thread	Verdict	Action
Invalidate stale Pi cached costs	Correct	Bump Pi cost cache v4 → v5 so complete costNanos rows reprice with cache-write rates
Priority cache-write 1.25×	Correct	Priority CodexPricing now sets write = priorityInput × 1.25 when base model has write rates
Wire cache writes into native Codex JSONL	Not fixable yet	Local token_count events only expose input / cached_input / output / reasoning_output — no cache_write_tokens in CLI logs (verified). Pi path already prices writes. Follow-up when CLI surfaces the field.

Verification

swift test --filter 'CostUsagePricingTests|PiSessionCostScannerTests'  # 51 passed

Threads replied + resolved.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 8d2d038560

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".

[0xSMW]

Third review round addressed

Thread	Verdict	Fix
Fall back all GPT-5.6 long-context rates from models.dev gaps	Correct	Bundle above-threshold input/output/cache-read rates when catalog omits context_over_200k but the 272K threshold is still applied

Verification

swift test --filter CostUsagePricingTests  # 40 passed

Thread replied + resolved.

[steipete]

Thanks — I independently verified the public GPT-5.6 Sol, Terra, and Luna model cards and OpenAI pricing table. The model IDs, gpt-5.6 → Sol alias, standard/cache-write/long-context/Priority rates, and 272K threshold are current, so this PR should stay open.

Before merge, please rebase onto current main; #2018 now conflicts in CHANGELOG.md and the generated parser hash. Preserve both changelog entries and regenerate the hash after resolving.

One repository-policy fix is also needed: the official cards currently list only the unsuffixed snapshots, so the *-2026-06-26 identifiers in CostUsagePricingTests.swift are not verified released model IDs. Please use the exact released IDs for pricing coverage, or an explicitly fictitious dated suffix such as 2099-01-01 for normalization-only coverage.

Please also refresh the PR summary, which still says flat rates/no long-context or Priority although HEAD now implements those paths. After rebasing, rerun the focused pricing/Pi tests, make check, and make test. The current head has only GitGuardian; we’ll re-review the exact rebased head once the full CI rollup is green.

[0xSMW]

Thanks for the rate verification and the correction on the model identifiers.

1. Rebased
2. Updated model ID policy
3. Refreshed PR summary
4. Verification green

[steipete]

Exact-head proof for fd72c50fe3b09d83ac4093d648cc37cfbf27a24e:

• Public Model Identifier Gate: PASS. Only the released GPT-5.6 tier IDs and public unsuffixed alias are used in production; dated normalization fixtures use an unmistakably fictitious year. The identifiers, alias, and rates match the current OpenAI model cards and pricing table.
• Pricing semantics: the published short/long/cache-write/Priority rates are explicit; Priority remains unavailable above 272K, so traced long-context rows retain their bucket attribution while using Standard long-context base cost.
• Focused suites: 150 pricing/scanner/cache tests passed, including all-tier boundaries, catalog authority/fallback, unchanged-file repricing, and fetched-time stability.
• make check: passed; generated parser hash current.
• Full make test: 594/594 selections, 50/50 groups, zero failures, retries, or timeouts.
• Independent review: clean after rejecting an unsupported attempt to combine Standard-long and Priority-short pricing.
• Signed packaged proof: debug bundle embeds fd72c50f, bundle ID com.steipete.codexbar.debug, valid Developer ID signature, packaged CLI CodexBar 0.41.1.
• Source-blind isolated CLI behavior: 400,000 tokens priced at $1.00; changing only cache-write rate repriced the untouched Pi file to $1.40 and changed the pricing key; changing only fetchedAt kept $1.40, the pricing key, and scan timestamp stable. No credentials, Keychain, browser state, or provider requests.
• Dependency freshness: N/A; no dependencies changed.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: fd72c50fe3

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".

[steipete]

Exact rebased head 615276b8 is ready for hosted verification.

Proof:

• Public Model Identifier Gate: PASS; released canonical production names only, with explicitly fictitious dated test variants.
• Focused pricing/scanner/models.dev/unknown-model coverage: 112 tests passed.
• make check: passed.
• Full make test: all 595 selections and 50 groups passed. One unrelated slow group exceeded the batch timeout; all 12 selections passed on isolated retry.
• Independent exact-head range-diff and code review: clean; every PR code/test blob is byte-identical to the previously proven head, and the only rebase difference is retained changelog context from Fix: Keep cost history chart steady while hovering #2016/Add Kimi K2 usage dashboard shortcut #1919.
• Signed debug package: embedded 615276b8, bundle ID com.steipete.codexbar.debug, Developer ID Peter Steinberger / Y5PE65HELJ, packaged CLI 0.41.1.
• Exact packaged live proof: 272,000-token boundary row $0.59625; 272,001-token long-context row $1.185005; a rate-only catalog change repriced the unchanged Pi file $1.781255 → $3.56251; fetched-time-only change kept cost, pricing key, and scan timestamp stable.
• Live harness was fully isolated with synthetic Pi JSONL and a local cache; no provider, browser, Keychain, or ambient credentials.
• Dependency freshness: no dependencies added or changed.

Hosted CI/security now applies to this exact synchronized head.