OpenAI resets Codex and ChatGPT Work limits after 9M active users

• OpenAI reset Codex and ChatGPT Work again after saying the combined products reached 9M active users, with the 9M reset post promising weekly usage would return to 100%.
• The growth curve went vertical after GPT-5.6: Codex moved from 6M to 8M active users in roughly three days in haider1's growth chart, while sama said agentic product usage rose 2.5x in a week.
• The 5-hour gate was relaxed, but the weekly pool still mattered: the 8M reset post said the 5-hour limit was still gone, while theo's usage screenshot showed a weekly balance down to 3%.
• OpenAI also patched burn-rate mechanics, since the GPT-5.6 Sol update listed inference optimizations, a 372K-to-272K context rollback, reverted reasoning-effort experiments, and multi-agent efficiency fixes.
• Hands-on reports split between abundance and wall-clock drag: bridgemindai claimed 30 agents used 2% of a weekly limit, while doodlestein's Sol Ultra run spent 5+ hours auditing a plan and then 4 more hours revising it.

The OpenAI help page says Codex, ChatGPT Work, ChatGPT for Excel, and Workspace Agents draw from the same agentic usage and credit pool. The Codex pricing page says ChatGPT Work uses the same pricing, credits, and usage limits as Codex. The reset wave also came with a weird growth loop: OpenAI's GPT-5.6 promo page offered $100 in credits to eligible paid users who posted public examples, then closed submissions after the first wave.

OpenAI said Codex and ChatGPT Work hit 9M active users, then reset limits while teams worked on reliability. That came after the 8M post said the 5-hour rate limit was still temporarily removed so users could "explore the boundaries" of GPT-5.6 Sol.

The public milestone trail looks like this:

• 1M to 5M from Feb. 5 to May 31, per haider1's chart.
• 6M on Jul. 12, 7M on Jul. 13, and 8M on Jul. 15, per the same chart.
• 9M by Jul. 16, according to the latest reset post.

The product move was rationing turned into community theatre. sama said GPT-5.6 Sol growth was "insane," credited the inference team, and warned that hiccups were possible.

OpenAI's accounting model is shared, not product-local. The official help page says usage from Codex, ChatGPT Work, ChatGPT for Excel, and Workspace Agents draws from one agentic usage and credit pool when those surfaces are available.

The reset types became their own feature surface:

• Hard reset: the 8M post reset usage limits for everyone.
• Banked reset: the 7M post added a saved reset users could apply later in desktop or web.
• Self-service reset: testingcatalog's settings screenshot showed a "Use reset" modal inside ChatGPT.
• Terminal redemption: Teknium's Hermes Agent post added /usage reset for Codex/OpenAI subscriptions.

Anthropic joined the reset race the same morning. ClaudeDevs said it reset 5-hour and weekly rate limits for all users, and kimmonismus framed the timing as direct competition.

The most useful detail was not the reset button. It was the accounting bug hunt.

The update listed four changes:

• Inference optimizations, expected to give roughly 10% more GPT-5.6 Sol usage.
• A context-size rollback from 372K to 272K because the larger setting charged more usage than intended.
• Reverted experiments that changed reasoning effort, described as "juice values" under the hood.
• Multi-agent and auto-review efficiency fixes, especially for high and xhigh reasoning effort.

A separate community report from kimmonismus said the context window reduction, subagent adjustments, and other changes should reduce burn rate.

The usage reports did not converge on one experience. They converged on a new bottleneck: long agentic runs can make weekly quota feel either huge or fragile depending on mode, harness, and task shape.

Examples from the evidence pool:

• theo said a 23h+ Ultra goal still left him feeling like he got a lot of usage from the weekly limit.
• bridgemindai said 30 simultaneous GPT-5.6 Sol Max agents consumed 2% of weekly usage, with five banked resets still available.
• doodlestein said Sol Ultra spent 5+ hours on a plan audit, then another 4 hours revising the plan and launching four review agents.
• kevinkern said Codex built a working SaaS v1 in a little over a day, with 3,810,156 tokens and about 32 hours of runtime shown in the attached completion report.

Kevin Kern's loop is the cleanest operational pattern in the pile:

1. Build a task graph from the product spec.
2. Link tasks and define a completion outcome contract.
3. Have a worker implement and a reviewer independently review.
4. Have a verifier run tests plus a live browser session.
5. Create fix tasks for review findings.
6. Repeat until the graph is complete and the audit passes.

The practical response to burn rate was model routing. kevinkern's cheat sheet mapped GPT-5.6 variants to roles instead of treating Sol as the default for every step:

• Driver/Orchestrator, Sol Medium/High: planning, delegation, conflict resolution, progress evaluation, final synthesis.
• Explorer, Terra Low/Medium: repo exploration, relevant files, dependencies, architecture mapping, summaries.
• Worker, Terra Medium/Sol Medium: features, fixes, refactors, tests.
• Mechanical Worker, Luna Low: extraction, formatting, repetitive edits, boilerplate, deterministic transforms.
• File Operator, Luna Low: scoped file search, directory operations, copying, moving, renaming, sorting.
• UI Designer, Sol Medium/High: UI concepts, layouts, design systems, responsive behavior, visual polish.
• Debug Specialist, Sol High: ambiguous bugs, root cause analysis, complex failures.
• Reviewer, Sol High: material-change review, subtle regressions, final implementation review.
• Security/Architecture, Sol High/Max: threat modeling, migrations, high-risk decisions.
• Browser Verifier, Sol Low/Terra Medium: browser flows, responsive behavior, accessibility, E2E checks.
• Visual Judge, Sol Medium/High: rendered UI evaluation and screenshot comparison.

kevinkern's harness split later reduced the pattern to a driver, worker, reviewer, and verifier split: Sol for driving and review, Terra or Sol for work, Luna or Terra for verification.

Two more workflow gotchas surfaced. reach_vb said asking Codex to research and set its own goal can reduce the search space and make verification clearer, while swyx warned that stale AGENTS.md instructions can become self-inflicted prompt injection during long /goal runs.

OpenAI also paid for social proof. The promotion page offered $100 in ChatGPT credits, about 2,500 credits, to the first 10,000 eligible paid users who posted public examples of what they loved about GPT-5.6 or why they switched.

The same page says submissions are now closed. It also says OpenAI may repost and feature participant replies and X handles in marketing, including paid ads, without extra compensation.