The abnormal quota consumption vulnerability of OpenAI's coding agent Codex remains unresolved. Core product lead Tibo Sottiaux announced over the weekend that after reviewing logs in the war room on Sunday, a second hard reset has been implemented for all users. The initial reset on June 27 failed to resolve the issue, with some users reporting that within just 36 hours after manually resetting their quota, their allocation dropped by 75% again.
Sottiaux Announces Second Hard Reset
After the initial reset on June 27, 2026, some users reported that within just 36 hours after manually resetting their quota, their allocation dropped by 75% again. Sottiaux opened the war room on Sunday to review logs, but the issue remained unidentified, leading to the decision to implement a second hard reset for all users, citing that some users had already performed up to three manual resets of their quota, opting for a forced reset rather than issuing quota reset cards.
Sottiaux said on X: "Since we are still investigating, I have reset the Codex usage limits for all users. Interestingly, this week has been called 'RESET week' at OpenAI, intended for everyone to relax. However, this will be a different kind of reset week." Independent developer nicdunz criticized OpenAI for being too lax in responding to the failure; the specific trigger mechanism is still under investigation.
OpenAI Opens Limited Preview of GPT-5.6 to Trusted Partners
The positioning and pricing of the three GPT-5.6 models are as follows:
Sol (Flagship): Complex coding, research, biology, and cybersecurity tasks; includes Ultra mode (multi-sub-agent handling complex tasks); Input $5, Output $30 (per million tokens)
Terra (General-purpose): Performance comparable to GPT-5.5, API pricing approximately half that of GPT-5.5; Input $2.50, Output $15
Luna (High-speed): Applications prioritizing speed and capacity; Input $1, Output $6
In Terminal-Bench 2.1 testing, Sol scored 88.8%, and Sol Ultra scored 91.9%. GPT-5.6 Sol is also planned to launch on Cerebras in July (up to 750 tokens per second, initially for limited customers); OpenAI plans a wider release in the coming weeks, with no specific date announced yet.
GPT-5.6 Sol Described by OpenAI as the Most Powerful Cybersecurity Model to Date
OpenAI has described GPT-5.6 Sol as the most powerful cybersecurity model to date, but Sol did not reach the "Cybersecurity Critical" threshold under the "Cyber Security Preparedness Framework" and did not autonomously generate a complete end-to-end attack code under test conditions. OpenAI shared relevant model and capability information with the U.S. government prior to the preview release.
Security measures have introduced model-level refusal mechanisms, real-time classifiers, account-level monitoring, and differentiated access permissions; the company has invested over 700,000 hours of A100 GPU in automated red team exercises. GPT-5.6 prompt cache adjustments: cache write cost is 1.25x that of uncached input, cache read enjoys a 90% discount.
Frequently Asked Questions
What is the current status of the investigation into the OpenAI Codex quota vulnerability?
The initial reset on June 27, 2026, failed to resolve the issue, with some users' quota dropping by 75% again within 36 hours. Sottiaux announced a second hard reset after reviewing logs in the war room on Sunday, but the specific trigger mechanism is still under investigation and has not yet been identified.
What are the positioning and pricing of the three GPT-5.6 models?
Sol (Flagship) Input $5, Output $30 (complex coding and cybersecurity); Terra (General-purpose) Input $2.50, Output $15 (comparable to GPT-5.5 at half price); Luna (High-speed) Input $1, Output $6 (speed-prioritized applications); billing is per million tokens.
When will GPT-5.6 be available to the public?
Currently, only a limited preview is available to a small number of trusted partners. OpenAI has stated that it plans a wider release through ChatGPT, Codex, and API in the coming weeks, but has not yet announced a specific public release date.
