Codex usage limits are running out faster than they did a few weeks ago, and the people reporting it say their work has not changed. Simon Høiberg, a developer who tracks his own token use through an OpenClaw setup, said on X that one of his Codex accounts, which uses only GPT-5.6 Sol, now exhausts its limits 4.8 to 5.9 times faster than two months ago. His token consumption, he said, is around 18% slower, and his reasoning efforts and mix of work are the same.

A GitHub report filed on September 20 goes further. The reporter says a single Sol task on a Plus plan used up the entire five hour allowance in about 14 minutes. These are individual accounts, not proof that OpenAI changed anything, but they describe the same pattern across several models. This post covers what users are reporting, what seems to help, and which cheaper models come close.

What users are reporting

The complaints began with Astra, OpenAI’s newest model, and OpenAI’s help center says Astra can use your allowance faster than GPT-5.6 Sol, with Plus plans including only limited Astra usage. The newest reports involve the models below it. The September 20 reporter says they had already left GPT-6 Astra because of its higher usage, and that on a Pro 20x account sustained Sol use can now exhaust the weekly allowance in roughly 12 hours.

Another report covers Luna, the cheapest tier. Its author compared local telemetry from August 31 with mid September samples and estimates the same model at the same Max reasoning effort now delivers at least 4.2 times less quota throughput, calling that a lower bound because the quota meter covers the whole account.

A third report, also from September 20, describes a Codex Desktop session of about five messages over roughly 20 minutes that effectively exhausted the five hour allowance, with Codex sending requests every few seconds at roughly 140,000 input tokens each, almost all of them cached.

Why the drain happens

Each time an AI coding agent takes a step, it re-reads the whole conversation so far, which is called its context. Parts it read recently are cached, meaning reused at a discount. Codex counts allowance in credits, and cached input burns a tenth of the credits of fresh input, while output burns six times the input rate. A long context still adds up, because it is re-read at every step. One Plus user logged about 7.96M tokens in a 26 minute session, roughly 7.69M of them cached, while the five hour meter rose from about 0% to about 86%. That session ran in Fast mode, which costs more.

Anshu Chimala, a heavy Codex user, had Astra analyze his own threads and reports that about 70% of his spend was cached input reads. That is one account, but it matches what the public reports describe.

As far as I can find, OpenAI has not yet explained this week’s reports. According to a Kingy AI review, Tibo Sottiaux said on July 28 that Sol worked longer and made more tool calls than earlier models, and on August 21 that worse cache hit rates for some users could make usage drain faster.

What you can do

Lower the reasoning effort. OpenAI says Astra at Low effort can outperform Sol at High, and suggests trying Astra at Low or Medium. In Artificial Analysis testing, each step up in Sol’s reasoning level raised average cost per task by roughly 50%.

Turn Fast mode off. For Astra in Work and Codex, Fast mode is charged at 2.5 times the standard rate.

Keep the conversation short. Start a fresh thread between unrelated tasks instead of letting one run for hours. One caution: a Cerebras guide says cache lasts about 30 minutes, so restarting in the middle of a task can cost more.

Watch for repeated checking. While a long command runs, Codex can check on it again and again, and each check sends the whole conversation back to the model. One traced session made 90 checks, each carrying about 240,000 tokens, for roughly 21.6 million input tokens and almost no output. You can tell Codex to wait for the command to finish, but one user reported it kept checking anyway. If your limits drop while a long task runs, that may be why.

Trim what loads every time. Every connected tool (called an MCP server) adds descriptions to the context. Remove tools you do not use, and keep skills and AGENTS.md files short.

Be careful with /goal on Astra. Late August reports said goals could run past their stop condition or retry broken tools, using 15% to 70% of a weekly allowance in one session. OpenAI says it fixed some of these problems, and Chimala still advises avoiding it.

Check before big jobs. OpenAI recommends looking at Settings → Usage before a large task. In the CLI, /status shows the same.

Cheaper models that come close

Astra is hard to beat on cost per task. Artificial Analysis (AA) found Astra at max effort costs $7.09 per coding task, about 15% more than Sol, for a 7 point higher score, because it uses about a third of Sol’s tokens. Claude Fable 5.1 matches Astra’s coding score of 62, and Claude Opus 5 scores 60, but Astra costs about 40% less per task than Fable 5.1 and about 30% less than Opus 5. Two caveats apply. AA measures pay per token cost, not subscription plan usage, and AA said it supported OpenAI with pre release evaluation of the GPT-5.6 models.

For cheaper models, the best comparison is AA’s general Intelligence Index, a combined score across many tests. AA rebuilt it in early September, so only current scores are comparable. Astra and Fable 5.1 lead at 53, Opus 5 scores 51, Muse Spark 1.3 scores 48 and Sol scores 47.

The closest cheaper options

• If you use Sol (47): GLM-5.3 scores about 45 and Grok 4.6 scores 44. In one reference workload, GLM-5.3’s token prices came to roughly 70% below Sol’s. Grok 4.6 lists at $2 per million input tokens and $6 per million output, for conversations under 200,000 tokens, versus $4 and $20 for Sol. Z.ai also sells GLM coding plans at $18, $80 and $168 a month with published credit limits.
• Budget pick: GLM-5.3 Flash scores 42, the same as GPT-5.6 Terra at max effort, at 18% of Terra’s cost per task ($0.25 versus $1.40). That is about 90% of Sol’s score.
• Cheapest OpenAI option: Luna scores 38 at $0.18 per task, 9 points below Sol, and it is not immune to the usage drain described above.

These are general intelligence scores, not coding tests. The coding specific evidence I found is AA’s Coding Agent Index, which covers Astra, Fable 5.1, Opus 5, Sol and Muse Spark 1.3 but none of the cheaper models above. One roundup reports GLM-5.3 at 42% on Terminal-Bench 4.0, a coding test where AA scored Sol at 40%, which is encouraging but is a single result.

FAQ

Why are my Codex usage limits running out faster than before?

Astra uses allowance faster than Sol, and higher reasoning effort, Fast mode, long conversations, repeated checking and extra connected tools all add to it. Some users also report faster drain for unchanged work, which OpenAI has partly linked to cache hit rates, without addressing this week’s reports.

Is there a cheaper alternative to Astra or Sol?

For Sol, GLM-5.3 and Grok 4.6 score within 2 to 3 points at lower token prices, and GLM-5.3 Flash is the budget pick. For Astra, I found no cheaper model at the same level. Most of these scores measure general intelligence, not coding.

How can I make my Codex usage last longer?

Lower the reasoning effort, turn Fast mode off, keep conversations short, watch for repeated checking during long commands and trim connected tools and skills.

Is Claude a cheaper option?

Not per task in AA’s data. Astra costs about 40% less per task than Fable 5.1 and about 30% less than Opus 5, for the same or higher coding score. One comparison also notes Claude Pro can use Fable 5.1 only by paying for usage credits.

Where this leaves things

Everything above is gathered advice on managing context, and I have not tested most of it myself. I tried Astra once on my Pro plan and ran out of my usage session before it finished a single prompt. I have not used it since, so I cannot say whether that was Astra’s cost, my settings, or one of the problems reported above. If you have found something that works for you, share it in the comments or send it to me, along with your plan, model and /status numbers.