Codex Ultrafast requests fall back to Standard pricing #1987
Description
What happened?
ccusage 20.0.45 does not recognize recorded service_tier: "ultrafast". With a Standard config fallback, these requests are priced at Standard rates, underestimating API-equivalent costs.
Steps to reproduce
- Use a Codex home with
service_tier = "default"inconfig.toml. - Save this synthetic rollout under its
sessions/directory:
{"timestamp":"2026-10-27T10:00:00Z","type":"event_msg","payload":{"type":"thread_settings_applied","thread_settings":{"service_tier":"ultrafast"}}}
{"timestamp":"2026-10-27T10:00:01Z","type":"event_msg","payload":{"type":"token_count","info":{"model":"gpt-6.1-sol","last_token_usage":{"input_tokens":1200,"cached_input_tokens":200,"output_tokens":50,"total_tokens":1250}}}}- Run
npx [email protected] codex daily --speed auto --offline --jsonagainst that home.
The request is valued at $0.00252.
Expected behavior
Use the recorded Ultrafast tier: $0.01512 for these tokens, using the published OpenAI rates.
I have a tested fix ready that adds recorded-tier detection, --speed ultrafast, and published Astra/Sol rates while preserving token totals. Could you approve a PR for this change?
Version
20.0.45; also reproduced on main at b8ad727.
Activity
pullfrog commentedon Oct 28, 2026
github-actions commentedon Oct 28, 2026
Pullfrog triage: priority:low
Kind: feature request
Maintenance fit: excluded
Confidence: high
Requests new Codex ultrafast tier support, new --speed ultrafast flag, and new Astra/Sol rates. Current code only supports Standard, Fast, and Flex and leaves unknown tiers unclassified. No maintainer endorsement exists. This is a new compatibility surface with synthetic-only evidence.
Decision: closed by the contribution gate.
rzv-me commentedon Oct 28, 2026
Adding context for human review: this came from comparing actual retained Codex session history with a separate local tracker. The synthetic fixture is a minimized reproduction so I don't have to publish private rollouts. I verified real
thread_settings_appliedevents containingservice_tier: "ultrafast".ccusage already reconstructs pricing from recorded tier changes, but leaves this value unclassified. With a Standard config fallback, those requests receive Standard estimates. OpenAI publishes Ultrafast API rates at 6× Standard for GPT-6.1 Sol and GPT-6 Astra; the repository's existing models.dev snapshot also includes Sol Ultrafast rates.
This affects historical reports too: changing a thread back to Standard should preserve the earlier requests' recorded Ultrafast pricing. The regression tests cover that switch within one thread, without changing token totals. These are API-equivalent estimates based on recorded settings; actual server-side processing and billing remain unconfirmed.
I understand that
--speed ultrafastandultrafastMultiplieradd new configuration surface. I'm happy to separate those and limit a PR to recognizing and pricing the recorded tier.Tested branch. Could a maintainer review whether this narrower change is within scope?
rzv-me commentedon Oct 28, 2026
@ryoppippi could you review the clarification above? This arose from real Codex usage history. I’m happy to limit the PR to recognizing and pricing recorded Ultrafast requests, leaving the new CLI/config options for separate consideration.
fix(codex): price recorded Ultrafast requests at Ultrafast rates (ccu…
ryoppippi commentedon Nov 10, 2026
Note
🤖 Claude Opus 5.5 responding on behalf of @ryoppippi
This landed in #1989, which credits you as co-author, and shipped in v20.0.46. Recorded Ultrafast requests are now priced at Ultrafast rates. Sorry your follow-up went unanswered.
docs(skills): credit issue reporters and reply after merging (#2056)
fix(claude): show model breakdowns for session --id (#2057)