Short answer: Anthropic is replacing Claude Code's temporary 50% usage boost with a permanent 25% increase over the May baseline. Because the 50% boost has been in place since May, users today are working at 150% of baseline, and moving to 125% is a 17% cut from current levels. Anthropic confirmed the math. Meanwhile, Codex expanded limits and added banked resets. For Q4 automation planning, Codex has more capacity runway right now.
Table of Contents
- The 17% Math, Explained Without the Spin
- What Codex Has Been Doing Instead
- Codex vs Claude: Capacity Snapshot for Q4
- What This Means for Your Q4 Automation Budget
- Which Tool Wins in Which Situation
- Action Steps Before September 14
- Bottom Line
- FAQ
- Sources
The 17% Math, Explained Without the Spin
Anthropic's announcement said: "We are permanently raising Claude Code weekly limits by 25%."
That sentence is technically true. It is also misleading, and Anthropic later clarified why.
Here is the actual situation. In May 2026, Anthropic introduced a temporary 50% bonus on top of baseline weekly limits for paid Claude Code subscribers. That bonus has been extended multiple times and has been in effect continuously since May. So today, every paid subscriber is working at 150% of the May baseline.
The new "permanent 25% increase" is measured against the May baseline, not against today's working capacity. So starting September 14:
| Reference point | Limit level |
|---|---|
| May 2026 baseline (original) | 100 |
| Current (with 50% summer boost) | 150 |
| New permanent (25% over May baseline) | 125 |
| Change from today to September 14 | -17% |
125 divided by 150 is 0.833. Users lose 16.7% of their current working capacity, rounded to 17%.
When users pushed back on the original announcement, Anthropic deleted the post and reissued a clarification stating: "Compared to today, this works out to a 17% reduction." That confirmation matters: the 17% figure is Anthropic's own official number, published in its clarification.
The permanent increase is still real. The reference point is May, before the bonus was added, rather than the capacity you are actually working with today.
What Codex Has Been Doing Instead
While Claude has been cycling through temporary boosts and then pulling them back, Codex has been moving in a different direction.
Over the summer of 2026, OpenAI introduced banked usage resets for Codex. Unused weekly quota carries forward instead of expiring at reset. For a business owner who has a lighter week and then wants to run a heavier batch of agent tasks, banked capacity means those credits are waiting for you rather than gone. Claude Code's limits do not work this way: unused capacity resets on schedule regardless.
OpenAI also increased baseline Codex rate limits and aligned them more closely with ChatGPT subscription tiers, so users on higher plans see proportionally more agent runway. The company introduced a capacity boost for harness-level tasks and expanded the number of parallel agent sessions available on Pro and higher plans.
The pattern is consistent. Codex has been adding capacity. Claude Code has been borrowing it temporarily and then taking it back.
Codex vs Claude: Capacity Snapshot for Q4
| Factor | Codex (OpenAI) | Claude Code (Anthropic) |
|---|---|---|
| Recent limit direction | Expanded | Cut 17% from current on Sept 14 |
| Banked (rollover) usage | Yes, unused quota carries forward | No, resets on schedule |
| Parallel agent sessions | Up to 4 on Pro/higher (Ultra mode) | Plan-dependent, no stated multi-agent expansion |
| Capacity trend since May 2026 | Growing | Peaked in May, now contracting |
| Rate limit reset mechanism | Weekly + banked carryover | Weekly, no carryover |
| Budget predictability for Q4 | Higher (stable or growing baseline) | Lower (losing 17% of current budget Sept 14) |
What This Means for Your Q4 Automation Budget
Q4 is the highest-demand quarter for most B2B businesses. It is when you run end-of-year client campaigns, close pipeline, plan for the year ahead, and push as much work as possible into automated systems so December does not collapse into chaos.
If you are running agent workflows through Claude Code, here is what September 14 looks like in practice: the same weekly budget of tasks you ran last week now takes 17% longer to complete, or requires you to reprioritize which tasks the agent handles. The models stay the same. The weekly runway shrinks.
Seventeen percent sounds modest. Run the numbers on actual weekly usage and it is more visible than it sounds. If your current Claude Code weekly budget lets you run 60 client deliverables, you will run 50 after September 14 on the same plan. That is one client project per week that either falls off the agent system or waits until the following week's limit resets.
The situation is different if you are on Codex. OpenAI's direction has been toward more capacity, not less, and the banked reset system means a light week adds to next week's budget rather than expiring. For Q4 planning, that is a meaningfully different resource model.
Which Tool Wins in Which Situation
| Situation | Better fit |
|---|---|
| Running high-volume autonomous agent tasks in Q4 | Codex + ChatGPT Work |
| You have existing Claude Code workflows and switching costs are high | Stay on Claude Code, but audit your weekly task budget |
| You want banked capacity for variable-intensity weeks | Codex |
| Your team is already inside ChatGPT Work for other tasks | Consolidate into Codex to avoid managing two agent limits |
| You want to test which agent runs your specific workflow better | Run both in parallel for two weeks, measure outputs, then consolidate |
| Your workflows are primarily long-document or creative tasks | Either works; evaluate on output quality, not capacity alone |
Action Steps Before September 14
You have 13 days before the Claude Code limit change lands. Here is what to do with them.
- Audit your current weekly Claude Code usage. Pull the last three weeks of task logs and identify which workflows use the most capacity. Those are the ones most at risk of hitting the new lower ceiling.
- Identify which tasks are portable to Codex. Any task that runs through a connector Codex also supports (Gmail, Google Drive, Slack, Google Calendar, Notion, HubSpot) can be moved without rebuilding from scratch.
- Run your highest-volume batch tasks before September 14. You are currently at 150% of May baseline. After September 14 you will be at 125%. Use the current capacity to get ahead on anything that can be pre-run: client reports, content drafts, database cleanups, outreach drafts.
- Set up Codex if you have not yet. If your current agent stack is entirely inside Claude Code, this is a good forcing function to add Codex as a second rail. You do not have to choose one. Running both gives you a combined capacity floor that neither alone provides.
- Check your plan tier on both tools. Codex capacity scales with your ChatGPT subscription tier. If you are on Plus, your banked capacity ceiling is lower than Pro. If you are running significant agent workflows, the Pro plan at $200 per month is the tier where Codex's multi-agent and capacity advantages become most visible.
Bottom Line
Treat the Claude Code limit change as a planning signal rather than a crisis. Anthropic gave every paid subscriber more than they had in May. The summer bonus ran longer than originally planned. The permanent tier is still better than where things started. That is the honest read.
What is also true is that the change lands heading into the highest-demand quarter of the year, and it moves in the wrong direction compared to what Codex has been doing. Codex expanded. Claude Code is contracting, at least relative to today's working levels.
For business owners running or planning to run meaningful agent workflows in Q4, the practical answer is: use both tools, understand each one's weekly ceiling, and route your highest-volume tasks through the system with more capacity. Right now that is Codex.
If you have not set up Codex yet, the setup guide is a good starting point. If you are already on Codex and want to add agent tasks that currently live inside Claude Code, the connector list is largely the same. Most workflows transfer cleanly.
FAQ
Is the Claude Code limit cut permanent?
The removal of the summer 50% boost is permanent. The 25% permanent increase over the May baseline is also permanent. What you are losing on September 14 is the bonus that has been stacked on top of that permanent increase since May. Anthropic has not announced any plans to restore a bonus layer.
Does this affect Claude the chat assistant, not just Claude Code?
This specific limit change applies to Claude Code, the autonomous agent product on paid plans (Pro, Max, Team, and seat-based Enterprise). The standard Claude conversation product has separate rate limits not affected by this announcement.
Can I buy extra Claude Code capacity to offset the cut?
As of this writing, Anthropic does not offer on-demand capacity top-ups for Claude Code the way OpenAI offers additional Codex credits through ChatGPT Business. If you need more agent throughput than your plan provides, your current option is to upgrade to a higher tier or move tasks to a tool that has available capacity.
Why did Anthropic frame a capacity cut as an increase?
The 25% increase is real when measured against the May 2026 original limits. The confusion comes from the reference point. If you joined after May and have only ever seen the boosted limits, you have never experienced the baseline the 25% is calculated against. From your perspective the cut is straightforward. For longer-term subscribers it is the same math: 125% of baseline vs. 150% of baseline is a 17% reduction from today.
Will Codex ever cut limits the same way?
There is no publicly stated plan for Codex to cut limits. OpenAI did briefly restore a tighter 5-hour Codex usage cap for Plus subscribers in late August 2026, which it later revised, so OpenAI's limit management has also been inconsistent. The key difference is the directional trend: Codex has added banked resets and expanded baseline limits over the summer, while Claude Code's summer boost is now ending.
Sources
- BleepingComputer: Anthropic is cutting Claude Code's current weekly limits by 17%
- NotebookCheck: Anthropic announces a 25% increase to Claude Code limits, but there's a 17% catch
- Explainx.ai: OpenAI Codex capacity boost and harness fixes, August 2026
- OpenAI Community: Codex rate limits reset for all paid plans