The Shared Agentic Credit Pool: How OpenAI's Quota Architecture Silently Drains Your Codex CLI Budget — and What the 5-Hour Limit Removal Actually Changes

Updated

codex-cliquotacredit-poolbudget5-hour-limitcost-managementconfig.tomlprofilesmodel-routing

The Shared Agentic Credit Pool: How OpenAI’s Quota Architecture Silently Drains Your Codex CLI Budget — and What the 5-Hour Limit Removal Actually Changes

On 12 July 2026, Tibo Sottiaux announced three simultaneous changes to Codex usage limits: the rolling 5-hour window was temporarily removed, every subscriber’s weekly quota was reset to 100%, and GPT-5.6 Sol efficiency improvements were rolling out 1. The community celebrated. What most developers missed is that the structural problem — a single shared credit pool feeding ChatGPT, Work, Codex, ChatGPT for Excel, and Workspace Agents — remained entirely unchanged 2.

This article dissects the quota architecture as it stands in late July 2026, maps it to the Codex CLI configuration surface, and provides a defensive configuration playbook for developers who need predictable agent capacity.

The Shared Pool Problem

ChatGPT Work and Codex draw from the same metered usage bucket 2. A product manager drafting a strategy document in Work and a developer running a multi-file refactor in Codex are competing for the same credits. When Workspace Agents became credit-priced on 6 July 2026, that pool expanded further 3.

The consequences are not theoretical. GitHub Issue #30918 documents a Plus subscriber whose quota jumped from 9% to 70% consumed with no visible session activity, then reached 100% in under six minutes — a drain rate of approximately 5.7% per minute 4. The reporter questioned whether the problem involved “incorrect usage accounting, hidden/background usage attribution, or a serious regression in per-turn weighting.” OpenAI acknowledged a usage-drain fix on 30 June but the behaviour persisted into July.

graph LR
    subgraph "Shared Agentic Credit Pool"
        A[ChatGPT Work] --> P[Weekly Credit Quota]
        B[Codex App/CLI] --> P
        C[ChatGPT for Excel] --> P
        D[Workspace Agents] --> P
    end
    P --> W{Weekly Limit<br/>Reached?}
    W -->|No| OK[Continue]
    W -->|Yes| STOP[Hard Stop]

The UX does not make this pooling obvious across surfaces 2. A developer working exclusively in the CLI has no visibility into whether a colleague’s Work session has already consumed a substantial portion of the shared allocation.

What the 5-Hour Removal Actually Changes

Before 12 July, the quota system enforced a dual-layer constraint: a rolling 5-hour window and a separate weekly cap 1. Developers could exhaust the 5-hour bucket while their weekly quota remained largely untouched, creating frustrating mid-session hard stops.

The temporary removal collapses this to a single constraint: the weekly cap becomes the sole governing limit 5. For CLI-heavy developers who burned through rolling windows during deep coding sessions, this is a genuine improvement — sessions can now run longer without arbitrary interruption.

However, two critical caveats apply:

  1. The removal is explicitly temporary. No reinstatement date has been published 1.
  2. The weekly cap itself has not increased. Removing the 5-hour window simply allows developers to consume their weekly allocation faster, not to consume more of it.

Quota Numbers by Tier and Model

Credit consumption varies dramatically by model family. The following table reflects the subscription-tier limits as of July 2026 6:

Plan GPT-5.6 Sol (5h) GPT-5.6 Terra (5h) GPT-5.6 Luna (5h) Multiplier
Plus ($20) 15–90 20–110 50–280 1×
Pro $100 75–450 100–550 250–1,400 5×
Pro $200 300–1,800 400–2,200 1,000–5,600 20×
Business Plus-equivalent Plus-equivalent Plus-equivalent 1×

⚠️ With the 5-hour window temporarily removed, these ranges now represent approximate consumption capacity within a rolling period rather than hard boundaries. The weekly cap remains the binding constraint.

Token-based credit pricing reveals the cost asymmetry between models 6:

Model Input (credits/M tokens) Output (credits/M tokens)
GPT-5.6 Sol 125 750
GPT-5.6 Terra 62.50 375
GPT-5.6 Luna 25 150

Sol consumes 5× the credits of Luna per million tokens. For routine tasks — file organisation, test scaffolding, documentation updates — routing to Luna or Terra rather than Sol is the single highest-impact budget defence available.

Defensive Configuration for CLI Developers

Named Profiles for Tier Routing

Since Codex CLI 0.134.0, profiles live in separate files rather than [profiles.name] tables 7. Create task-appropriate profiles that route to cheaper models by default:

# ~/.codex/scout.config.toml
# Lightweight tasks: file ops, docs, simple refactors
model = "gpt-5.6-luna"
model_reasoning_effort = "low"

# ~/.codex/deep.config.toml
# Complex multi-file refactors, architecture decisions
model = "gpt-5.6-sol"
model_reasoning_effort = "high"

# ~/.codex/ci.config.toml
# CI/CD pipelines: predictable, low-cost
model = "gpt-5.6-terra"
model_reasoning_effort = "medium"

Invoke with --profile:

# Routine task — Luna at low effort
codex --profile scout "organise these test files by module"

# Deep refactor — Sol at high effort
codex --profile deep "refactor the payment module to use the new billing API"

# CI pipeline — Terra at medium effort
codex exec --profile ci "run the full test suite and fix failures"

Auto-Compaction as Budget Guard

The 272K context window pricing threshold means requests exceeding 272,000 tokens incur a 2× input pricing multiplier 8. Set model_auto_compact_token_limit below this boundary to prevent sessions from silently crossing it:

# ~/.codex/config.toml
model_auto_compact_token_limit = 250000

This forces compaction before the session reaches the pricing threshold, keeping per-request costs within the standard band.

Reasoning Effort as a Quota Lever

Reasoning effort directly affects token consumption and therefore credit drain. The available levels — minimal, low, medium, high, xhigh — map to progressively more expensive inference 7. For the vast majority of CLI tasks, medium provides an adequate quality-to-cost ratio:

# ~/.codex/config.toml
model_reasoning_effort = "medium"

Reserve high and xhigh for security reviews, architectural decisions, and tasks where reasoning depth genuinely matters. A profile-based approach makes this switching friction-free.

Banked Resets: The Emergency Valve

Codex app 26.609 (June 2026) introduced rate-limit reset banking for Plus, Pro, and Business subscribers 9. A banked reset is essentially a recovery coupon — when you hit a rate limit mid-session, you can deploy a banked reset to unblock immediately rather than waiting for the natural reset or switching to API-key billing.

Key constraints:

  • Banked resets operate through the Codex app’s usage menu, not via a CLI command 9.
  • Each reset expires 30 days after grant — no carryover 2.
  • Resets do not increase total quota; they give control over timing 2.
  • One free reset is provided as starting allocation; additional resets may be earned through referral campaigns 9.

For CLI developers, the practical implication is clear: when codex exec hits a rate limit mid-pipeline, a banked reset can unblock you without switching to API-key billing. But this is a tactical measure, not a strategic defence.

The Two-Stage Workflow

A practical pattern for conserving Codex credits is the two-stage approach 10:

  1. Stage 1 (ChatGPT): Complete architectural planning, schema design, and framework comparison in the standard chat interface. This draws from the conversational quota pool — separate from the agentic pool.
  2. Stage 2 (Codex CLI): Execute implementation with a finalized blueprint. This draws from the agentic pool but with a clear scope, reducing wasted exploratory tokens.

The embedded chat panel inside the Codex tab retains standard ChatGPT billing rules 10, so lightweight conversations within Codex do not drain the agentic budget.

graph TB
    subgraph "Stage 1: ChatGPT (Conversational Pool)"
        A1[Architecture Discussion] --> A2[Schema Design]
        A2 --> A3[Framework Selection]
        A3 --> A4[Finalised Blueprint]
    end
    subgraph "Stage 2: Codex CLI (Agentic Pool)"
        A4 --> B1[Scoped Implementation]
        B1 --> B2[Test Execution]
        B2 --> B3[Iterative Fixes]
    end
    style A1 fill:#e8f4f8
    style A2 fill:#e8f4f8
    style A3 fill:#e8f4f8
    style A4 fill:#d4edda
    style B1 fill:#fff3cd
    style B2 fill:#fff3cd
    style B3 fill:#fff3cd

GPT-5.6 Sol Efficiency Improvements

OpenAI’s announced efficiency improvements for Sol report approximately 10% less quota consumption for equivalent tasks 9, with some sources citing up to 54% token efficiency gains for standard mode 2. The discrepancy likely reflects the difference between standard and high-compute modes — the latter consumes significantly more credits and does not benefit equally from the efficiency rollout.

⚠️ These figures have not been independently verified by the community. Issue #34225 documents a 3.3× latency regression during 12–19 July 2026, suggesting that the efficiency improvements may have introduced performance trade-offs 11.

What Happens When Limits Return

The 5-hour window will almost certainly be reinstated — OpenAI has been explicit that the removal is temporary 1. When it returns, developers who have optimised their workflows around unlimited session length will face a sharp adjustment.

The defensive configuration described above — named profiles, auto-compaction, reasoning effort tuning — provides durable protection regardless of whether the 5-hour window is active. The shared credit pool problem, however, requires an architectural solution from OpenAI: either separate quotas for Work and Codex, or at minimum, per-surface visibility into credit consumption.

Until then, the only reliable defence is treating your weekly allocation as a budget and routing tasks to the cheapest model that can handle them competently.

Citations

  1. Tibo Sottiaux, OpenAI Codex announcement, 12 July 2026. GitHub Issue #34035 ↩ ↩2 ↩3 ↩4

  2. ChatForest, “OpenAI Pulls the 5-Hour Wall: Codex Limit Temporarily Removed, But the Shared Agentic Credit Pool Problem Remains,” July 2026. ChatForest ↩ ↩2 ↩3 ↩4 ↩5 ↩6

  3. Daniel Vaughan, “Workspace Agents Credit Pricing Starts July 6: A Codex CLI Practitioner’s Budget Preparation Guide,” Codex Knowledge Base, June 2026. Codex KB ↩

  4. GitHub Issue #30918, “Usage limits draining abnormally fast on Plus: 70% to 100% in about 6 minutes on July 2, 2026.” GitHub ↩

  5. ExplainX, “ChatGPT 5-Hour Limit Removed — July 2026.” ExplainX ↩

  6. SimpleMetrics, “ChatGPT Codex Limits 2026: GPT-5.6, Plus, Pro 5x/20x.” SimpleMetrics ↩ ↩2

  7. OpenAI, “Advanced Configuration,” ChatGPT Learn / Codex CLI docs. OpenAI Docs ↩ ↩2

  8. Daniel Vaughan, “The 272K Tripwire: How GPT-5.6’s Context Window Cap Silently Doubles Your Codex CLI Bill,” Codex Knowledge Base, July 2026. Codex KB ↩

  9. KnightLi, “Is the 5-hour limit of Codex temporarily cancelled: shared quota, Banked Reset and usage view,” July 2026. KnightLi ↩ ↩2 ↩3 ↩4

  10. TreeRouter, “ChatGPT vs Codex Quotas: Stop Wasting Agent Credits.” TreeRouter ↩ ↩2

  11. GitHub Issue #34225, “Starting from 2026-July-12, Codex performance dropped 3X.” GitHub ↩