Grok and Codex actual versus recorded cost
This audit matches the three requested Bullswarm attempts to their durable provider records, reads the provider-reported token fields, and compares them with the estimate stored in the completion artifact. The measured dollar amounts below are API-equivalent published-rate calculations. They are not a claim about a Grok Build or ChatGPT subscription debit; each recorded completion artifact says api-equivalent rate; subscription debit may differ (/home/dev/.bullswarm/workflows/wf-mu5ul9j7-4a3a73/completion-verify.json, /home/dev/.bullswarm/workflows/wf-mu6db6m7-bee921/completion-widget.json, and /home/dev/.bullswarm/workflows/wf-mu6db6m7-bee921/completion-integrate.json, .verdict.meta.usage.cost.basis).
Result
| Attempt and model | Recorded estimate | Measured provider usage | Measured API-equivalent USD | Measured / recorded |
|---|---|---|---|---|
fuyyca / verify-1, Grok 4.6 | 4,316 tokens; $0.009668 (/home/dev/.bullswarm/workflows/wf-mu5ul9j7-4a3a73/completion-verify.json, .verdict.meta.usage) | 9,400,999 billed tokens (rtk node scripts/cost-audit/grok-session-cost.mjs --log /home/dev/.grok/logs/unified.jsonl --session-id ee40deee-8f12-4dfc-98d8-4bb85d60bfef, output below) | $7.783282 (same reader command) | 2,178.174004x tokens; 805.056061x USD (ratio command output below) |
2yrcxi / widget-1, GPT-5.6 Luna | 1,757 tokens; $0.0006554 (/home/dev/.bullswarm/workflows/wf-mu6db6m7-bee921/completion-widget.json, .verdict.meta.usage) | 37,677,530 tokens (rtk node scripts/cost-audit/codex-session-cost.mjs /home/dev/.codex/sessions/2026/09/18/rollout-2026-09-18T11-12-47-01a0b280-9b02-7e30-a869-6ca18d6677f4.jsonl) | $0.96434956 (Codex arithmetic command output below) | 21,444.240182x tokens; 1,471.390845x USD (ratio command output below) |
2yrcxi / integrate-1, GPT-5.6 Sol | 1,766 tokens; $0.016424 (/home/dev/.bullswarm/workflows/wf-mu6db6m7-bee921/completion-integrate.json, .verdict.meta.usage) | 8,853,907 tokens (rtk node scripts/cost-audit/codex-session-cost.mjs /home/dev/.codex/sessions/2026/09/18/rollout-2026-09-18T12-08-59-01a0b2b4-0b9a-7161-b59d-4c7476cf26b9.jsonl) | $4.5087544 (Codex arithmetic command output below) | 5,013.537373x tokens; 274.522309x USD (ratio command output below) |
The recorded attempt identity, model, and time window are in /home/dev/.bullswarm/workflows/wf-mu5ul9j7-4a3a73/state.json for verify-1, and /home/dev/.bullswarm/workflows/wf-mu6db6m7-bee921/state.json for widget-1 and integrate-1. The token and dollar values in the table come from the completion artifacts named in the table, not from the human-readable durations in the task prompt.
The widget rollout was selected by action identity, not by the first file with a matching start second. The workflow event log records widget-1 starting at 2026-09-18T03:12:47.767Z and finishing at 2026-09-18T03:56:36.031Z (/home/dev/.bullswarm/workflows/wf-mu6db6m7-bee921/events.jsonl, sequences 18 and 24). The matching rollout is /home/dev/.codex/sessions/2026/09/18/rollout-2026-09-18T11-12-47-01a0b280-9b02-7e30-a869-6ca18d6677f4.jsonl: its durable task_complete message closes that action ([other project's task text removed]) and its last timestamp is 2026-09-18T03:56:34.042Z. The other nearby rollouts are separate actions of the same workflow (/home/dev/.codex/sessions/2026/09/18/rollout-2026-09-18T11-12-47-01a0b280-997e-7972-bfbe-d2c7308f18af.jsonl, /home/dev/.codex/sessions/2026/09/18/rollout-2026-09-18T11-12-48-01a0b280-9c4c-7950-8d2e-a26fe7d057e6.jsonl, and the later /home/dev/.codex/sessions/2026/09/18/rollout-2026-09-18T11-56-36-01a0b2a8-b815-79f3-a1c9-2743bebd3559.jsonl). They are deliberately not added to the widget total.
The remaining boundary checks are also inside their recorded windows: the Grok state starts at 2026-09-17T21:15:23.757Z and finishes at 2026-09-17T21:41:21.693Z, while the selected Grok session spans 2026-09-17T21:15:25.266Z through 2026-09-17T21:41:19.704Z (/home/dev/.bullswarm/workflows/wf-mu5ul9j7-4a3a73/state.json and the Grok reader command below). The integrate state spans 2026-09-18T04:08:58.795Z through 2026-09-18T04:24:40.637Z, while its rollout spans 2026-09-18T04:08:59.351Z through 2026-09-18T04:24:38.410Z (/home/dev/.bullswarm/workflows/wf-mu6db6m7-bee921/state.json and the integrate reader command below).
The recorded token estimates are the UTF-8 fallback components in those same artifacts: Grok 4,057 standard-read + 259 output = 4,316 total-known; widget 1,453 + 304 = 1,757; integrate 1,181 + 585 = 1,766. The source field is estimated:utf8-bytes/4 in each .verdict.meta.usage object (/home/dev/.bullswarm/workflows/wf-mu5ul9j7-4a3a73/completion-verify.json, /home/dev/.bullswarm/workflows/wf-mu6db6m7-bee921/completion-widget.json, and /home/dev/.bullswarm/workflows/wf-mu6db6m7-bee921/completion-integrate.json).
Codex measurement
project-n documents the Codex storage layout as ~/.codex/sessions/<YYYY>/<MM>/<DD>/rollout-<ISO>-<sessionId>.jsonl (/home/dev/project-n/packages/parser/src/codex.ts:4-13). Its parser reads the cumulative event_msg/token_count total under payload.info.total_token_usage and maps input, output, and cached-input fields at lines 150-158. It explicitly treats those values as cumulative and attaches the final cumulative total once at lines 288-309; summing all token_count events would overcount. The parser's ordinary turn counter is based on visible event_msg/user_message records at lines 161-169.
The new dependency-free reader is scripts/cost-audit/codex-session-cost.mjs. It prints the last cumulative total, counts unique root/turn IDs from token_usage_record, and reports both possible user-message shapes. These are the exact rollout paths and command outputs used:
rtk node scripts/cost-audit/codex-session-cost.mjs \
/home/dev/.codex/sessions/2026/09/18/rollout-2026-09-18T11-12-47-01a0b280-9b02-7e30-a869-6ca18d6677f4.jsonl
=> model gpt-5.6-luna, turns 1, tokenCountEvents 259,
input 37,603,121, cached input 36,919,808, cache write 0, output 74,409,
reasoning output 34,405, total 37,677,530,
event_msg user messages 0, response_item user messages 2,
malformedLines 0
rtk node scripts/cost-audit/codex-session-cost.mjs \
/home/dev/.codex/sessions/2026/09/18/rollout-2026-09-18T12-08-59-01a0b2b4-0b9a-7161-b59d-4c7476cf26b9.jsonl
=> model gpt-5.6-sol, turns 1, tokenCountEvents 72,
input 8,834,905, cached input 8,669,696, cache write 0, output 19,002,
reasoning output 5,771, total 8,853,907,
event_msg user messages 0, response_item user messages 2,
malformedLines 0The rollout scan independently inspected token_usage_record usage and found, for widget and integrate respectively, 259 and 72 records, maximum input requests of 240,477 and 158,477 tokens, and 0 requests above the OpenAI 272K long-prompt threshold. That validation was run with:
rtk node --input-type=module <<'NODE'
import fs from 'node:fs';
const files={widget:'/home/dev/.codex/sessions/2026/09/18/rollout-2026-09-18T11-12-47-01a0b280-9b02-7e30-a869-6ca18d6677f4.jsonl',integrate:'/home/dev/.codex/sessions/2026/09/18/rollout-2026-09-18T12-08-59-01a0b2b4-0b9a-7161-b59d-4c7476cf26b9.jsonl'};
const out={}; for(const [name,p] of Object.entries(files)){let n=0,max=0,over=0,sum={input:0,cached:0,output:0,reasoning:0,total:0};const roots=new Set(),turns=new Set();for(const l of fs.readFileSync(p,'utf8').split(/\r?\n/)){if(!l)continue;const r=JSON.parse(l);if(r.type==='token_usage_record'){if(r.payload?.root_turn_id)roots.add(r.payload.root_turn_id);if(r.payload?.turn_id)turns.add(r.payload.turn_id);const u=r.payload?.usage??{};n++;max=Math.max(max,u.input_tokens??0);if((u.input_tokens??0)>272000)over++;for(const [k,f] of [['input','input_tokens'],['cached','cached_input_tokens'],['output','output_tokens'],['reasoning','reasoning_output_tokens'],['total','total_tokens']])sum[k]+=u[f]??0;}}out[name]={tokenUsageRecords:n,sum,maxInputTokens:max,requestsOver272K:over,uniqueRoots:roots.size,uniqueTurns:turns.size};} console.log(JSON.stringify(out,null,2));
NODEFor the dollar calculation, output tokens include the reasoning subset in the Codex cumulative total, so reasoning output is reported for diagnostics but is not added a second time. The calculation uses uncached input as input - cached_input, cached input at the cached-input rate, and output at the output rate:
rtk node --input-type=module <<'NODE'
import fs from 'node:fs';
const rows=[
{name:'widget', rollout:'/home/dev/.codex/sessions/2026/09/18/rollout-2026-09-18T11-12-47-01a0b280-9b02-7e30-a869-6ca18d6677f4.jsonl', completion:'/home/dev/.bullswarm/workflows/wf-mu6db6m7-bee921/completion-widget.json', rate:{input:.2,cached:.02,output:1.2}},
{name:'integrate', rollout:'/home/dev/.codex/sessions/2026/09/18/rollout-2026-09-18T12-08-59-01a0b2b4-0b9a-7161-b59d-4c7476cf26b9.jsonl', completion:'/home/dev/.bullswarm/workflows/wf-mu6db6m7-bee921/completion-integrate.json', rate:{input:4,cached:.4,output:20}},
];
const parse=p=>{let u;for(const l of fs.readFileSync(p,'utf8').split(/\r?\n/)){if(!l)continue;const r=JSON.parse(l);if(r.type==='event_msg'&&r.payload?.type==='token_count'&&r.payload.info?.total_token_usage)u=r.payload.info.total_token_usage;}return u;};
const out={};for(const x of rows){const u=parse(x.rollout), rec=JSON.parse(fs.readFileSync(x.completion,'utf8')).verdict.meta.usage, uncached=u.input_tokens-u.cached_input_tokens, cost=(uncached*x.rate.input+u.cached_input_tokens*x.rate.cached+u.output_tokens*x.rate.output)/1e6;out[x.name]={recordedTokens:rec.tokens.totalKnown,recordedUsd:rec.cost.estimatedUsd,actualTokens:u.total_tokens,actualUsd:Number(cost.toFixed(8)),uncachedInput:uncached,cachedInput:u.cached_input_tokens,output:u.output_tokens,tokenRatio:Number((u.total_tokens/rec.tokens.totalKnown).toFixed(6)),usdRatio:Number((cost/rec.cost.estimatedUsd).toFixed(6))};}console.log(JSON.stringify(out,null,2));
NODEThe command output was:
widget: recordedTokens 1757, recordedUsd 0.0006554,
actualTokens 37677530, actualUsd 0.96434956,
uncachedInput 683313, cachedInput 36919808, output 74409,
tokenRatio 21444.240182, usdRatio 1471.390845
integrate: recordedTokens 1766, recordedUsd 0.016424,
actualTokens 8853907, actualUsd 4.5087544,
uncachedInput 165209, cachedInput 8669696, output 19002,
tokenRatio 5013.537373, usdRatio 274.522309The vendor pages name the exact models. OpenAI's GPT-5.6 Luna page says “Per 1M tokens — Input $0.20, Cached input $0.02, Output $1.20” (model page, lines 888-902) and applies the >272K input multiplier at line 923. The GPT-5.6 Sol page says “Per 1M tokens — Input $4.00, Cached input $0.40, Output $20.00” (model page, lines 887-901) and gives the same long-prompt rule at line 924. Therefore the two Codex sessions use the short-context rates: no request in the rollout scan crossed that threshold.
Grok measurement
The Grok provider module confirms that Bullswarm's meter is a weekly credit endpoint, not a per-session token ledger: src/providers/grok/provider.mjs defines CREDITS_URL and returns seven_day utilization in fetchGrokUsage (lines 14-18, 168-213). The CLI does retain a session record. The relevant safe inventory under ~/.grok is:
/home/dev/.grok/logs/unified.jsonl
/home/dev/.grok/sessions/%2Fhome%2Fdev%2FRepo%2Fbullswork%2Fbullswarm-dashboard/ee40deee-8f12-4dfc-98d8-4bb85d60bfef/
summary.json
events.jsonl
chat_history.jsonl
updates.jsonl
prompt_context.json
resources_state.json
rewind_points.jsonl
signals.jsonNo credential-bearing files were printed. The target session's summary.json identifies cwd /home/dev/Repo/bullswork/bullswarm-dashboard and model grok-4.6; events.jsonl contains one turn_started, 58loop_started events, and one completed turn_ended. The session ID is ee40deee-8f12-4dfc-98d8-4bb85d60bfef, selected by matching the verify-1 window in /home/dev/.bullswarm/workflows/wf-mu5ul9j7-4a3a73/state.json.
scripts/cost-audit/grok-session-cost.mjs filters ~/.grok/logs/unified.jsonl to that session ID and sums each shell.turn.inference_done record. Its exact command output was:
rtk node scripts/cost-audit/grok-session-cost.mjs \
--log /home/dev/.grok/logs/unified.jsonl \
--session-id ee40deee-8f12-4dfc-98d8-4bb85d60bfef \
--session-dir /home/dev/.grok/sessions/%2Fhome%2Fdev%2FRepo%2Fbullswork%2Fbullswarm-dashboard/ee40deee-8f12-4dfc-98d8-4bb85d60bfef
=> model grok-4.6, turns 1, inferenceLoops 58, inferenceRecords 58,
prompt 9,311,859, cached prompt 9,046,912,
uncached prompt 264,947, completion 58,164,
reasoning 30,976, totalBilledTokens 9,400,999,
short-context requests 42 / $3.393098,
long-context requests 16 / $4.390184,
apiEquivalentUsd $7.783282, malformedLines 0Grok's model page names the exact rate card: “Input Tokens $2.00/1M tokens,” “Cached tokens $0.50/1M tokens,” and “Output Tokens $6.00/1M tokens” (Grok 4.6 pricing, lines 261-277). The model page says requests exceeding the 200K context window use higher-context pricing (lines 285-287). The xAI API pricing table names the exact long-context row for grok-4.6: short 2 / 0.50 / 6 and long 4 / 1 / 12 USD per million input/cached/output tokens, with the threshold shown as ≥ 200k (API pricing, lines 226-238). xAI's usage/pricing page specifies that reasoning tokens use the full completion rate and that, when the total prompt exceeds the threshold, cached and non-cached prompt tokens use their respective long-context rates (usage and pricing, lines 306-312). The Grok reader therefore prices 42 prompts below 200,000 at input/cache/output rates 2 / 0.5 / 6 USD per million, and 16 prompts at or above that threshold at 4 / 1 / 12, charging reasoning at the completion rate. The exact arithmetic and ratio command output was:
rtk node --input-type=module <<'NODE'
import fs from 'node:fs';
const sid='ee40deee-8f12-4dfc-98d8-4bb85d60bfef';
const log='/home/dev/.grok/logs/unified.jsonl';
const grokRecorded=JSON.parse(fs.readFileSync('/home/dev/.bullswarm/workflows/wf-mu5ul9j7-4a3a73/completion-verify.json','utf8')).verdict.meta.usage;
let total={prompt:0,cached:0,completion:0,reasoning:0,cost:0,requests:0,long:0,short:0};
for (const line of fs.readFileSync(log,'utf8').split(/\r?\n/)) { if (!line) continue; const r=JSON.parse(line); if(r.sid!==sid||r.msg!=='shell.turn.inference_done') continue; const c=r.ctx??{}; const p=Number(c.prompt_tokens)||0, cache=Math.min(p,Number(c.cached_prompt_tokens)||0), comp=Number(c.completion_tokens)||0, reason=Number(c.reasoning_tokens)||0; const long=p>=200000; const rates=long?{input:4,cached:1,output:12,reasoning:12}:{input:2,cached:.5,output:6,reasoning:6}; total.prompt+=p; total.cached+=cache; total.completion+=comp; total.reasoning+=reason; total.cost+=((p-cache)*rates.input+cache*rates.cached+comp*rates.output+reason*rates.reasoning)/1e6; total.requests++; total[long?'long':'short']++; }
total.totalBilled=total.prompt+total.completion+total.reasoning; console.log(JSON.stringify({recordedTokens:grokRecorded.tokens.totalKnown,recordedUsd:grokRecorded.cost.estimatedUsd,actual:total,tokenRatio:Number((total.totalBilled/grokRecorded.tokens.totalKnown).toFixed(6)),usdRatio:Number((total.cost/grokRecorded.cost.estimatedUsd).toFixed(6))},null,2));
NODEThe command output was:
recordedTokens 4316, recordedUsd 0.009668,
prompt 9311859, cached 9046912, completion 58164, reasoning 30976,
requests 58, long 16, short 42, totalBilled 9400999,
cost 7.783282000000002, tokenRatio 2178.174004,
usdRatio 805.056061The extra precision in the arithmetic command is ordinary floating-point representation; the report uses the reader's rounded $7.783282 output.
Validation and boundaries
Focused validation passed:
rtk node --check scripts/cost-audit/codex-session-cost.mjs
rtk node --check scripts/cost-audit/grok-session-cost.mjs
=> exit 0 for both filesBoth readers were then executed against the exact provider records shown above; both reported malformedLines 0. The arithmetic commands reproduced the table's actual USD values and ratios. No files under src/, tests/, ~/.bullswarm, ~/.claude-acme, ~/.codex, or ~/.grok were modified, and no commit or other git-changing command was run. The only delivered files are the two readers and this report.
rtk npm test was also run as a repository-wide check. It exited 1 with 1,228 tests, 1,200 passing, and 28 failing. The failures are existing terminal-rendering assertions in tests/dash-kit.test.js (Unicode glyphs were rendered as ASCII); the test output also reports that NO_COLOR is ignored because FORCE_COLOR is set. None of the failures exercise either new reader, so this unrelated environment-sensitive suite failure remains unfinished.
The unresolved boundary is billing identity: the durable records expose provider token counts and published-rate equivalents, but neither CLI log exposes a subscription debit. A provider billing export would be required to turn the API-equivalent USD values into account-charge values.