# tokenusage.dev

> Unofficial site about Codex and Claude Code token usage. It records Codex usage-limit reset announcements from @thsottiaux’s public posts on X (Reset Alerts), and compares Codex and Claude subscription plans through seven category sections of plan cards. Selected local runs reach the cards through owner ingest, Firestore and the subscription API. Higher-tier scaling assumptions are explained on the method page. Codex native-image rows show images per week; their API-value column shows “--” because image-model tokens are not reported. Cards have no Response column. Speed appears only in the daily chart from separate local probes. The site also hosts self-reported usage leaderboards and profiles. Not affiliated with OpenAI or Anthropic.

## Pages

- [Publications](https://tokenusage.dev/publications): independent research from tokenusage.dev. Research notes 002, 003, 004 and 007 link CSV files of their data, licensed CC BY 4.0; each note says what was left out.
- [AI benchmarks reach 96–97%, then need replacing](https://tokenusage.dev/publications/oct-26-progress-by-field): the seventh research publication, a fixed snapshot through October 15, 2026 of frontier progress in science, coding and factual recall. One GPQA Diamond chart and two sourced tables distinguish benchmark handovers, lab reports, independent runs, human baselines and uncertainty.
- [Most intelligent AI model: Opus 5.5 leads two indexes](https://tokenusage.dev/publications/oct-26-smartest-model): the sixth research publication, a fixed snapshot through October 14, 2026 of 24 Artificial Analysis v4.3.2 frontier records since GPT-4. Nineteen records are labelled AA estimates; five are measured. One chart and two sourced tables distinguish retrospective records, historical rankings, Epoch ECI, METR time horizons and LMArena preferences.
- [AI model speed: tokens per second rose 18-fold, then levelled off](https://tokenusage.dev/publications/oct-26-model-speed): the fifth research publication, a fixed snapshot through October 14, 2026 of API output speeds and reasoning waits. One chart, 18 selected speed readings and a three-model effort table preserve host, workload, date, public availability and source qualifications.
- [AI model parameters grew 24,000-fold, then went sparse](https://tokenusage.dev/publications/oct-26-model-sizes): the fourth research publication, a fixed snapshot through October 12, 2026 of published parameter counts and attributed closed-model estimates. One chart separates the dense-model record from selected MoE totals and active counts; 28 published-count rows and 10 closed-model evidence rows retain source links and qualifications.
- [AI context windows grew 4,096-fold, then stalled](https://tokenusage.dev/publications/oct-26-context-windows): the third research publication, a fixed snapshot through October 13, 2026 of selected context-window milestones from GPT-1 to 2026 models. One chart and two sourced tables distinguish public access, limited access, on-paper claims, tokenizer differences and RULER effective context.
- [What a million tokens cost: AI API prices, 2022–2026](https://tokenusage.dev/publications/oct-26-token-prices): the second research publication, a fixed snapshot through October 12, 2026 of 88 API price records across nine selected model lines from six labs. An editorial article includes one full-view chart image, four premium-tier rows, a complete price table with sources, and method and rounding notes.
- [Claude vs ChatGPT: the subscription value gap](https://tokenusage.dev/publications/oct-26-subscription-value): the first research publication, using an attributed, fixed snapshot of SemiAnalysis’s publicly visible October 5, 2026 charts. Interactive comparisons cover monthly API value, tokens, utilization, Claude workload scenarios, model caps, historical Pro changes, and Chinese coding plans. References, source dates, and rounding notes accompany the figures. These are third-party source estimates, separate from the homepage’s local category runs.

- [Nerf Watch: is Claude or Codex getting worse?](https://tokenusage.dev/nerfwatch): daily private tests through Claude Code and Codex on paid plans, scored by code and compared with each model's own baseline. A lower score is not proof that a provider changed a model.
- [Codex Reset Tracker](https://tokenusage.dev/resets): the latest Codex reset announcements, a calendar in the viewer's time zone and recent history, with text and email alerts
- [Banked vs automatic Codex resets](https://tokenusage.dev/reset-types): what an automatic reset and a banked reset grant mean, and how the tracker labels each announcement
- [Codex reset time](https://tokenusage.dev/codex-reset-time): how to find your account's current limits and reset times, understand five-hour and weekly limits, and check schedule changes after banked or paid resets
- [How to check Codex usage](https://tokenusage.dev/codex-usage): the Codex usage dashboard, CLI and desktop commands, quota-window readings, observed session-file keys and an optional local token report. Static guide; no local category-run figures.
- [Codex vs Claude Code: pricing and usage limits](https://tokenusage.dev/): Seven categories repeat six cards at $20, $100 and $200 a month, with three model rows per card. API-published local category runs supply weekly estimates and monthly API value for text results. Codex native-image rows show images per week and “--” for API value. GPT-6.1 Sol and Opus 5.5 are measured first; missing results show “No figure yet.” Higher-tier scaling assumptions are explained on the method page. Category jump links and the method support inspection. Speed appears only in the daily chart from separate local probes. Official plan rules, API-credit policies and sources remain below the cards.
- [Claude Pro vs Max for Claude Code](https://tokenusage.dev/claude-pro-vs-max): Pro, Max 5x and Max 20x subscription prices, five-hour and weekly usage limits, Fable, fast mode, API credits and usage credits, each with official sources. Static copy; links to the home page cards for local category-run figures and scaled estimates.
- [Codex pricing: ChatGPT Plus and Pro](https://tokenusage.dev/codex-pricing): Plus and Pro $100, $200 and $500 subscription prices, included models, five-hour and weekly usage limits, grandfathering, speed modes and buying credits, each with official sources. Static copy; links to the home page cards for local category-run figures and scaled estimates.
- [Is Claude Max or ChatGPT Pro worth it?](https://tokenusage.dev/worth-it): a calculator over the homepage's published category-card math. Enter a weekly percentage to see capacity, monthly API value and same-provider plan shares, with measured and projected rows distinguished. Uses no figures from shares.
- [Track Codex and Claude Code token usage](https://tokenusage.dev/share): install the token-counter plugin for Codex or Claude Code, build a local token report and, when you ask, share your usage here
- [How the numbers work](https://tokenusage.dev/method): how API-published local category runs produce homepage capacity estimates, how higher-tier scaling and native-image cells work, how daily speed probes are measured, and the separate method for shared-report API figures
- [Developers](https://tokenusage.dev/developers): public JSON, the reset Atom feed, the OpenAPI document and separate reset and Nerf Watch MCP servers, with fields, caching and reuse conditions
- [Claude Usage Limits: 5-Hour & Weekly Resets](https://tokenusage.dev/claude-usage-limits): session and weekly limits, reset times, Fable, usage credits and options at a limit
- [How to Check Claude Code Usage: /usage & Trackers](https://tokenusage.dev/claude-code-usage): /usage, Settings > Usage and a sourced comparison of the built-in view and six trackers

## Public data

- [Reset announcements](https://tokenusage.dev/api/resets): read-only, no key; newest first, at most 50. Free to reuse with credit: where you show the data to others, show the credit "Data: tokenusage.dev" linked to [https://tokenusage.dev/resets](https://tokenusage.dev/resets).
- [Reset announcement Atom feed](https://tokenusage.dev/api/resets.atom): Atom 1.0 for the same recorded reset announcements, no post text; poll no faster than 15 seconds. Retain each entry's linked credit and pass on the credit condition.
- [Nerf Watch JSON](https://tokenusage.dev/api/nerfwatch): read-only scores, changes and history; no key; poll no faster than 5 minutes. A lower score is not proof that a provider changed a model. Where you show the data to others, show the credit "Data: tokenusage.dev" linked to [https://tokenusage.dev/nerfwatch](https://tokenusage.dev/nerfwatch).
- [OpenAPI document](https://tokenusage.dev/openapi.json): OpenAPI 3.1 description of these GET and HEAD endpoints, their stable fields, caching and errors.

## MCP servers

- [Codex reset MCP endpoint](https://tokenusage.dev/mcp): Streamable HTTP, read-only, no sign-in; tools get_codex_reset_announcements and get_codex_reset_calendar. Calendar dates are UTC; optional month is YYYY-MM. Both tools show only recorded announcements from @thsottiaux, at most the 50 newest, not a complete history. A date without an entry does not mean no reset was announced.
- [Codex reset server card](https://tokenusage.dev/mcp/server-card): v1 identity, connection details and supported protocol versions; tools are listed through tools/list
- [Nerf Watch MCP endpoint](https://tokenusage.dev/mcp/nerfwatch): Streamable HTTP, read-only, no sign-in; tools get_nerfwatch_scores and get_nerfwatch_history. Both tools share one 300-second read cache. A lower score is a lower score on this suite, not proof that a provider changed a model.
- [Nerf Watch server card](https://tokenusage.dev/mcp/nerfwatch/server-card): v1 identity, connection details and supported protocol versions; tools are listed through tools/list
- [AI catalog](https://tokenusage.dev/.well-known/ai-catalog.json): domain discovery pointing to both server cards