OpenAI rolled back the GPT-5.6 Sol context limit from 372K to 272K tokens after discovering the larger window triggered 2x higher pricing. The rollback affects Codex and ChatGPT Work subscribers who experienced unexpectedly high usage burn.
TL;DR: OpenAI confirmed that GPT-5.6 Sol requests beyond 272K tokens are charged at 2x the standard rate. After setting the Codex context window to 372K tokens, many harnesses inadvertently pushed users into the higher pricing tier. The rollback to 272K restores approximately 10% more usage for subscribers.
Why Did OpenAI Reduce the Codex Context Window?
OpenAI reduced the context window because the 372K token limit was directly responsible for unexpectedly high usage burn among Codex and ChatGPT Work subscribers. According to Elaina on X, OpenAI confirmed that the 372K rollout was “one of the main sources of the unexpectedly high usage burn” affecting users. The larger context window caused requests to exceed the 272K threshold where pricing doubles.
The issue stems from how GPT-5.6 Sol handles token pricing. Requests beyond 272K tokens incur a 2x cost multiplier. By setting the context window to 372K, Codex allowed conversations to extend well into that premium pricing tier. Many users saw their quotas depleted far faster than expected.
This was not intentional. Tibo, an OpenAI engineer, explained on X that the team “noticed that by changing the” context window configuration, they could eliminate the unintended cost escalation. The rollback to 272K prevents requests from crossing into the 2x pricing band automatically.
The fix was straightforward. Cap the context at 272K tokens so conversations never reach the premium pricing tier unless explicitly requested.
How Does the 272K Token Limit Affect GPT-5.6 Pricing?
The 272K token limit directly determines where GPT-5.6 Sol switches from standard pricing to 2x pricing. According to Kun Chen on X, “OpenAI just pointed out requests beyond 272k tokens are over-charged.” This means any request that includes more than 272K tokens in its context window gets billed at double the normal rate.
Under the previous 372K configuration, Codex harnesses routinely built conversations that exceeded 272K tokens. Long coding threads, file uploads, and accumulated context pushed requests past the threshold. Users had no visibility into this pricing cliff — the harness simply filled the available context window.
The new 272K cap acts as a guardrail. By preventing the context window from exceeding the pricing threshold, OpenAI ensures that standard-rate usage is the default behavior. Users who need more than 272K tokens must explicitly opt in.
The pricing structure now works as follows: requests up to 272K tokens are billed at the standard rate, and requests exceeding that threshold are billed at 2x. The context window rollback means most Codex sessions will stay within the standard tier.
What Caused Unexpectedly High Usage Burn for Codex Users?
The unexpectedly high usage burn resulted from a convergence of issues that compounded each other. Theo on X described it as “a few issues with Codex that stacked really badly, causing people’s usage to get nuked.” The primary cause was the 372K context window pushing requests into 2x pricing territory.
Several factors made the problem worse. Codex increased its context limit to 372K tokens, which meant long threads naturally accumulated enough context to cross the 272K pricing threshold. Users had no indication that their conversations were now costing twice as much per request.
The pricing multiplier applied silently. There was no warning or prompt when a conversation approached the 272K boundary. The harness simply continued filling the context window until it hit the 372K ceiling.
Additionally, many third-party harnesses configured the GPT-5.6 context window at 372K based on the initial specification. These harnesses inherited the same problem. Users on non-Codex platforms experienced identical usage burn issues.
Theo noted that the issues “lined up to be as rough as possible” — the timing, the pricing structure, and the context window expansion all hit simultaneously.
How Much Usage Do Users Regain From the Rollback?
Users regain approximately 10% more usage from the rollback, according to Tibo on X. The OpenAI engineer stated that the team “landed inference optimizations and are passing down savings to all the subscriptions for GPT-5.6 Sol,” which should “result in around 10% more usage on its own.”
The usage recovery comes from two separate improvements. First, the context window rollback to 272K prevents requests from entering the 2x pricing tier. Second, inference optimizations reduce the computational cost of serving GPT-5.6 Sol requests across all subscription tiers.
Tibo characterized the updates as “no nerfing, only good stuff” — emphasizing that the changes are purely beneficial for subscribers. The combined effect of the pricing fix and the inference optimizations should restore normal usage patterns for affected users.
For Codex and ChatGPT Work subscribers who experienced depleted quotas, the rollback means their existing subscription allocations will last longer. Users do not need to take any action — the changes apply automatically.
Which Harnesses Are Affected by the Context Limit Bug?
Any harness that configured the GPT-5.6 context window at 372K tokens is affected by the bug. Kun Chen specifically warned users who are “using gpt 5.6 but not through codex” — meaning third-party harnesses and custom integrations that adopted the 372K specification.
The affected harnesses include:
- Codex CLI — OpenAI’s own command-line tool had its context window set to 372K
- ChatGPT Work — the subscription tier that provides access to GPT-5.6 Sol
- Third-party API wrappers — tools that interface with GPT-5.6 through the OpenAI API
- Custom integrations — developers who hardcoded the 372K context limit
- Open-source harnesses — community tools that adopted the initial 372K specification
- IDE extensions — editor plugins that integrate with GPT-5.6 for code assistance
- Automated pipelines — CI/CD systems that use GPT-5.6 for code review
- Research frameworks — academic tools configured for maximum context length
Kun Chen’s warning was explicit: “many harnesses right now set gpt 5.6 context window at 372k, which will cause you to lose quota more quickly than it should.” Users of these harnesses need to update their context window configuration to 272K.
| Harness Type | Previous Limit | Corrected Limit | Action Required |
|---|---|---|---|
| Codex CLI | 372K | 272K | Automatic update |
| ChatGPT Work | 372K | 272K | Automatic update |
| Third-party tools | 372K | 272K | Manual update |
| Custom integrations | 372K | 272K | Manual update |
| Open-source tools | 372K | 272K | Manual update |
For official OpenAI products like Codex and ChatGPT Work, the rollback applies automatically. Third-party harnesses require manual configuration changes to avoid the pricing penalty.
What Should Developers Do to Avoid Overcharging?
Developers using GPT-5.6 outside the official Codex harness must manually cap their context window at 272K tokens to prevent accelerated quota consumption. According to Kun Chen’s post on X, many third-party harnesses currently default to the old 372K limit, which causes requests exceeding 272K tokens to be overcharged at roughly double the standard rate. Updating your configuration is simple. Check your API wrapper settings.
The issue stems from a mismatch between what client-side tools report as the maximum context and what OpenAI’s pricing layer actually expects. Theo (t3.gg) confirmed that GPT-5.6 costs 2x more beyond the 272K threshold. Any tool still advertising 372K support will silently push requests into the expensive tier.
To stay safe, developers should take several concrete steps immediately. The following checklist covers the most critical adjustments:
- Update any local harness or wrapper that references the 372K token limit and replace it with 272K
- Audit recent API logs for requests that exceeded 272K tokens to estimate unexpected costs
- Monitor token counts per request using OpenAI’s tokenizer before sending long conversation threads
- Switch to shorter context windows for routine tasks that do not require deep codebase awareness
- Track quota usage dashboards daily until OpenAI confirms the billing correction is fully deployed
- Notify team members using shared API keys about the new 272K hard cap to prevent accidental overages
- Review any automated scripts or CI pipelines that batch large file contents into single prompts
- Consider chunking large repositories into smaller prompts rather than stuffing everything into one call
| Action | Priority | Estimated Effort | Impact on Cost |
|---|---|---|---|
| Update context limit in config | High | 5 minutes | Prevents 2x overcharging |
| Audit historical API logs | Medium | 30 minutes | Identifies past overcharges |
| Implement token pre-counting | Medium | 1 hour | Avoids accidental threshold crossing |
| Split large repository prompts | Low | Ongoing | Reduces average request cost |
How Did the Community React to the Codex Issues?
The developer community reacted with visible frustration as multiple Codex problems converged simultaneously, causing subscription quotas to deplete far faster than expected. Theo summarized the situation bluntly on X, noting that several distinct issues “stacked really badly” and aligned to create maximum disruption for paying users.
The core complaint was not a single bug but a cascade. First, GPT-5.6 doubles in cost beyond 272K tokens. Second, Codex raised its context ceiling to 372K, practically inviting users into the expensive tier. Third, long threads naturally accumulate token counts that push past the threshold without any explicit user action. The result? Quotas evaporated.
Community discussion centered on the lack of proactive communication. Many developers discovered the overcharging only after noticing their usage metrics spiking unexpectedly. By the time OpenAI posted updates about inference optimizations and the context rollback, a significant portion of affected users had already burned through substantial portions of their allocated quota.
Elaina’s post on X confirmed that the 372K rollout was identified as “one of the main sources of the unexpectedly high usage burn,” validating what the community had been reporting anecdotally for days.
What Inference Optimizations Did OpenAI Implement?
OpenAI landed a set of inference optimizations for GPT-5.6 Sol that the company claims will pass savings directly to all Codex and ChatGPT Work subscribers. Tibo (thsottiaux), an OpenAI engineer, announced on X that these optimizations should deliver approximately 10% more usage for existing subscriptions without any price increase.
The technical details of the optimizations remain undisclosed. However, the timing suggests OpenAI identified specific bottlenecks in how GPT-5.6 processes long context windows. By improving inference efficiency at the infrastructure level, the model can handle requests using fewer computational resources per token.
Tibo’s announcement framed the changes positively, stating there was “no nerfing, only good stuff.” Alongside the ~10% usage increase from inference improvements, the context rollback from 372K to 272K further reduces per-request cost. Combined, these two changes should noticeably extend subscription longevity for the average developer.
OpenAI also indicated that additional optimizations are in progress. The company is actively monitoring usage patterns and plans further adjustments based on real-world data from Codex and ChatGPT Work environments.
Will the Context Window Return to 372K Tokens?
OpenAI has not committed to restoring the 372K context window. The rollback to 272K was described as a corrective measure after the 372K limit was identified as a primary driver of unexpectedly high usage burn among Codex and ChatGPT Work users.
The company’s public statements suggest the 372K window was rolled out before adequate cost safeguards were in place. Since GPT-5.6 doubles in cost beyond 272K tokens, extending the context limit to 372K effectively pushed many routine operations into the premium pricing tier. Until OpenAI resolves the underlying cost structure — or adjusts the pricing model for extended context — a return to 372K seems unlikely.
Developers who depend on large context windows for entire-codebase analysis may need to explore alternative approaches. Chunking strategies, selective file inclusion, and retrieval-augmented generation can achieve similar results without crossing the 272K threshold.
How Does This Compare to Other AI Coding Tools?
The Codex context reduction highlights a broader tension in AI-powered developer tools: balancing capability with cost predictability. Competing platforms face the same fundamental constraint — longer context windows require more compute per request, and someone always pays for that compute.
Anthropic’s Claude supports context windows up to 200K tokens with transparent tiered pricing. Google’s Gemini 1.5 Pro offers up to 2 million tokens, though pricing scales accordingly. OpenAI’s GPT-5.6 at 272K now sits closer to Claude’s range, while the abrupt 372K experiment demonstrated the risks of expanding context before optimizing the cost structure.
Most AI coding assistants recommend keeping context windows well below maximum limits for routine tasks. The Codex incident reinforces why: pushing against the ceiling often triggers disproportionate cost increases that compound across long development sessions.
Frequently Asked Questions
Is OpenAI nerfing Codex by reducing the context window?
OpenAI engineer Tibo explicitly stated there is “no nerfing” involved. The rollback from 372K to 272K tokens was a corrective action because the 372K limit caused unexpectedly high usage burn. The company simultaneously delivered inference optimizations that provide approximately 10% more usage for subscribers.
How much extra did users pay at the 372K token limit?
Users whose requests exceeded 272K tokens were charged at 2x the standard rate for GPT-5.6, according to Theo’s analysis on X. Since Codex had raised its context ceiling to 372K, many long conversation threads automatically crossed into the doubled-cost tier without any explicit user action.
Do I need to update my API configuration after the rollback?
Yes. Kun Chen’s post warns that many third-party harnesses still set the GPT-5.6 context window at 372K. Developers should manually update their configurations to cap at 272K tokens to prevent accelerated quota consumption from the overcharging behavior.
Will ChatGPT Work users get refunded for the overcharging bug?
OpenAI has not announced specific refunds for users affected by the 372K overcharging issue. The company’s response focused on corrective measures: rolling back the context limit and delivering ~10% more usage through inference optimizations. Elaina’s post confirmed the rollback applies to both Codex and ChatGPT Work users.
Summary
- Cap your context at 272K tokens. Any harness still defaulting to 372K will trigger 2x overcharging on GPT-5.6 requests, silently draining your subscription quota.
- The 372K experiment caused real financial damage. Multiple issues stacked together — the doubled cost beyond 272K, the raised ceiling, and long conversation threads — to create maximum disruption for paying developers.
- OpenAI delivered partial compensation through inference optimizations. The announced improvements should yield roughly 10% more usage for all Codex and ChatGPT Work subscriptions.
- No commitment exists for restoring 372K. Until the underlying cost structure changes, developers should plan around the 272K limit using chunking and selective context strategies.
- Audit your recent API usage. Check logs for requests that exceeded 272K tokens to understand how much extra quota was consumed during the 372K window period.
If you found this breakdown helpful, subscribe to gikiewicz.com for more technical analysis of AI tools, API changes, and developer infrastructure.