Verdict

Agent loops can consume more tokens than a single response, and compaction can be followed by costly rereads. The evidence does not establish a universal Codex inflation multiplier, a five-file/50K context rule, or an OpenCode efficiency advantage.

Use the labels below precisely:

  • validated — current primary documentation or reproducible source evidence supports the claim.
  • version-specific — a report concerns named software versions and should not be generalized.
  • account-level — evidence describes one account, plan, or usage meter.
  • community-observed — dated user reports exist without first-party confirmation of scope or cause.
  • unresolved — evidence is incomplete or conflicting.
  • speculative — a labeled hypothesis or forecast is not established provider behavior or intent.
  • incorrect — current evidence contradicts the claim or its attribution.

Archive state is separate. validated means a concrete replayable capture exists; archive-pending means it does not. A confidence label never manufactures archive proof.

Claim ledger

ClaimStatusFindingArchive
Agent and subagent loops add model workvalidatedCodex documentation says subagents do their own model/tool work and use more tokens than comparable single-agent runs. No universal multiplier is documented.archive-pending
“Codex inflates tokens by 40–80%”incorrectOpenAI’s 40–80% figure is an internal-test improvement in cache utilization for Responses versus Chat Completions. It is not Codex inflation or waste.validated
Codex automatically includes five recently edited files or 50K tokensincorrectCurrent Codex documentation and searches of the pinned 2026-07-26 source revision do not document this rule. Treat it as incorrect or misattributed unless a versioned source is produced.archive-pending
OpenCode is more token-efficient than CodexunresolvedHarness, model, provider adapter, prompt shape, tools, compaction, retries, and task success all affect usage. No controlled same-task dataset establishes a general winner.n/a
Prompt caching makes prior context freeincorrectCaching discounts eligible exact-prefix reuse. Cached input still occupies context, and stateful Responses chains still bill prior input.validated
GPT-5.6 cache writes cost more than uncached inputvalidatedOpenAI documents cache writes at 1.25× uncached input for GPT-5.6 and later families, with reads billed at the cached-input rate.validated
Codex v0.118 lowered the compaction thresholdversion-specificIssue #16812 alleged a regression, but a maintainer found no threshold or estimator change and the reporter’s controlled test reversed the original result. The reporter closed it as not a bug.validated
A July 24 Desktop session looped through compaction and rereadscommunity-observedOpen issue #35226 records one GPT-5.6 Sol/Desktop 26.721.31836 session. Its token, quota, and causal estimates are reporter-supplied; no maintainer diagnosis is published.validated
The June 26 usage incident proved service-wide token inflationincorrectOpenAI attributed some reports to fraud/abuse systems incorrectly rate-limiting certain accounts, said impact appeared limited, and reported no broader Codex degradation.validated
July 25 resets compensated for caching defectsincorrectA first-party announcement says the reset reached all Codex and ChatGPT Work users after an almost global outage. It does not attribute the reset to caching or compaction.archive-pending
July 28 through August 13 broad resets fixed caching defectsincorrectDirect announcements establish reset scope but do not identify prompt caching or compaction as the cause. Efficiency, product, calendar, and milestone framing do not establish causality.archive-pending
Codex has a documented and universally active five-hour windowunresolvedOpenAI pricing documents shared five-hour windows and possible weekly limits, while current account reports show weekly-only meters. No first-party deprecation or completed-restoration notice resolves the difference, and account behavior may vary.pricing validated; current behavior community-observed
The August 8, 11, and 13 announcements created a permanent reset scheduleincorrectThe posts establish three resets, not a standing entitlement. The August 12 post says the earlier every-million promise ended at 10M; the 15M reset is a discretionary milestone event.archive-pending
A new broad banked-reset grant was announced after July 13unresolvedThe bounded review found no later direct announcement. OpenAI still documents offer-specific banked reset rewards, so absence of a broad announcement does not exclude account-specific grants.archive-pending
Weekly capacity plus discretionary resets is the near-term equilibriumspeculativeThis is a falsifiable AIHackers forecast, not a provider commitment. It fails if dependable five-hour enforcement, routine broad banked grants, or a different published system replaces the current pattern.n/a
Purchased credits refill five-hour or weekly countersincorrectOpenAI says available credits let work continue after included limits. It does not describe credits as restarting the shared five-hour window or refilling a weekly counter.validated
Some accounts can purchase a full weekly resetcommunity-observedDated Plus and Pro 20x reports show an account-scoped purchase control. OpenAI has not published universal availability, eligibility, plan coverage, or pricing.validated community captures
Purchased resets cost $8 for Plus and $80 for Pro 20x universallyincorrectThose are observed prices on some accounts, not a universal price sheet. Pro 5x, Business, tax, refunds, eligibility, and future pricing remain unresolved.validated community captures
A purchased reset stacks a bonus week without moving scheduled recoveryincorrectMultiple reports say redemption starts a new seven-day window and moves the next weekly reset date. This is account-level mechanism evidence, not a first-party universal rule.validated community captures
Paid resets replaced free promotional resetsunresolvedThe bounded evidence shows a paid control and separate historical provider-wide or banked grants. It does not establish that OpenAI ended or replaced future promotional resets.n/a
Sol allowance now lasts 18% longer for everyoneincorrectOpenAI projected around 18% longer allowance during typical Sol use after improvements. This is provider-reported, workload-dependent, and not an AIHackers benchmark or universal account result.archive-pending
A banked reset count applies to every planaccount-levelBanked resets and their visible count depend on eligibility, promotion, plan, workspace, and account state.archive-pending
One referral reward or expiry applies universallyincorrectCurrent ChatGPT Desktop terms make the benefit, eligible experience, qualifying action, cap, cooldown, redemption window, and expiry offer-specific. Referral credits are not API credits unless the applicable offer explicitly says otherwise. See the extra Codex usage referral guide; AIHackers publishes no OpenAI referral link.validated; prior Codex-titled revision captured 2026-07-10

What current Codex evidence actually says

The current Codex manual describes /compact as a way to summarize visible chat and free context, and /status as the place to inspect token usage and remaining capacity. Current source exposes configurable context, auto-compaction, and tool-output token limits.

At pinned commit 61a4488, source search finds compaction thresholds and context accounting. Searches for “recently edited,” “recent_files,” and “50K” do not establish a five-file preload contract. Absence from a source search is not proof that no private or future surface can behave that way; it is enough to reject the claim as documented current Codex behavior.

Read the compaction reports narrowly

Closed April report: version-specific

Issue #16812 compared v0.116 with v0.118 and initially attributed higher usage to more frequent compaction. A maintainer reported no compaction-threshold or token-estimator change. The reporter then ran an identical-prompt test in which v0.116 compacted three times and v0.118 twice, contradicting the original regression theory, and closed the issue as not a bug.

The report remains useful evidence that long, noisy tasks can reread material after compaction. It is not evidence of a confirmed v0.118 defect or a universal multiplier.

Open July report: community-observed

Issue #35226, opened July 24, describes one Desktop session repeatedly rereading files after automatic compaction near a full context window. The issue is open as of July 26. Its estimated 10–15% paid-usage impact is a user estimate, not an audited refund amount, root cause, or platform-wide rate.

Incident and reset chronology

  • June 26, investigating: OpenAI investigated reports of faster-than-expected Codex usage-limit consumption.
  • June 26, monitoring: OpenAI attributed some reports to incorrect fraud/abuse rate limiting of certain accounts, described the impact as limited, and said it had not observed broader degradation.
  • June 29, resolved: OpenAI marked the incident recovered.
  • July 25: OpenAI’s product lead announced a reset for all Codex and ChatGPT Work users after an almost global outage.
  • July 28: he announced a reset for all paid Codex and ChatGPT Work users.
  • July 29: he announced a reset for all Codex and ChatGPT Work users, projected around 18% longer typical Sol allowance, and said the temporarily paused five-hour limit would be restored July 30.
  • August 1: he announced another reset for Codex and ChatGPT Work with efficiency and Luna messaging; the post does not identify the efficiency work as the reset’s cause.
  • August 8: he announced a reset for all paid ChatGPT Work and Codex users.
  • August 11: he announced another reset for all paid ChatGPT Work and Codex users.
  • August 12: he said the earlier every-million promise ended at 10M and teased a surprise after growth passed that milestone.
  • August 13 at 01:01 UTC: he said Codex crossed 15M active users and announced a reset for “everyone.”
  • August 11 onward: dated account reports showed purchased full-reset controls at $8 on some Plus accounts and $80 on some Pro 20x accounts; availability remained uneven.

Keep scheduled limit recovery, provider-wide hard resets, incident remediation, manually applied banked resets, purchased full resets, and purchased usage credits in separate buckets. Reported purchased resets start a new seven-day window; credits continue work after included limits and are not documented as refilling counters. No cited announcement ties these resets to caching defects.

OpenAI’s June 11 release notes say banked resets are usable for 30 days after grant. A July 12/13 direct announcement described a 500,000-user grant and the first broad banked grant for the 7M milestone. Dated reports of grants disappearing around August 12–13 are community-observed; they do not prove universal expiry behavior, notification behavior, or formal withdrawal.

Reproducible comparison standard

To compare Codex with OpenCode, hold constant:

  1. model and provider;
  2. repository commit and clean starting state;
  3. prompt and project instructions;
  4. tools, MCP schemas, permissions, and network access;
  5. reasoning effort, timeout, retries, and subagent budget; and
  6. tests and acceptance criteria.

Record input, cache-write, cache-read, output/reasoning, tool calls, compactions, rereads, human review, and whether the result was accepted. Compare cost per accepted result, not raw tokens from mismatched tasks.

Sources

Global cache/source/issue review: July 26, 2026. Reset, pricing, paid-reset, credit, banking, and five-hour rows were rechecked August 20; unchanged compaction rows retain their earlier evidence dates.