On August 9, OpenAI Codex product lead Tibo (@thsottiaux) announced on X a full reset of GPT-5.6 Sol usage limits. The post, which drew 2.63 million views, 14,000 likes, and 1,500 reposts by public metrics, boils the point down to one sentence: "That's right, GPT-5.6 Sol is awesome and can be used pretty much anywhere, including in the CC harness." In other words, he declared that Sol, the flagship of the GPT-5.6 series launched July 9, can be used as-is inside rival Anthropic's terminal agent harness, Claude Code, and that to celebrate, paid users' usage limits would be reset.
OpenAI has handed out resets like this several times, but the timing is different now. Just one day earlier, on August 8, a developer who followed Tibo's guide to the letter and used a "Claude Code CLI + local proxy + GPT-5.6 Sol" combination had his Anthropic account suspended. A record of the full harness dispute shows OpenAI answering with a "limit reset" before the embers had even died down.
A "Full Reset" Button Pressed One Day After an Account Ban
Tibo was quoting a clarification post from Anthropic's Claude Code lead, Boris Cherny. Boris said, "We don't ban accounts for using harnesses with other models," called the suspension "almost certainly a false positive from a different account classifier," and had the account restored. According to a report summarizing the exchange, Tibo replied "That's right," and drove the point home: "GPT-5.6 Sol is awesome and can be used pretty much anywhere. That includes the CC harness." He added, "To celebrate, and with the fact that I'm not going anywhere, I've reset usage limits."
The reset covers all paid ChatGPT Work and Codex users. But right after it went public, voices questioned its real value, because the reset came just two to three hours before the weekly automatic refresh. One user pointed out that his remaining balance of 97% was reset to 100%, so he actually lost 3%, and Tibo shot back that he'd "do another performance reset on Monday." Even Digg's simple write-up of this reset boils it down to two things: promoting the model and resetting limits.
"It Works in the CC Harness" — The Full Story of the Harness Freedom Dispute
The remark didn't come out of nowhere. Tibo had been publicly advising users to do a "brain swap" since July: keep Claude Code as is, but switch only the model backend to GPT-5.6 Sol through a localhost-only proxy (CLIProxyAPI). He even joked that if you don't have the nerve to install the Codex app, you can keep using the "orange crab" (Claude) and still run Sol. Given reports that Codex has passed 8 million active users, "borrowing a harness to use a model" is no longer a stunt. It has become a phenomenon that generates real traffic.
The trouble started when developer Alex Getman, who did exactly that, had his account suspended on August 8. Anthropic cited "suspicious signals," and Boris publicly acknowledged a false positive and unblocked the account. He affirmed proxy-based multi-model workflows like LiteLLM as a "supported approach," while adding the caveat that harness quality depends on per-model tuning. Tibo countered that "freedom of harness" matters and that "users should decide which model is best."
As a result, the August 8 account suspension ended as a false positive and became the occasion for both companies to confirm, officially, the separation of the front end (harness) from the model backend.
What the Reset Is Worth: A Refresh Pulled Forward 2–3 Hours, Who Gains and Who Loses
This reset isn't a standalone event. It sits on top of the "Sol token burn" debate running since July. GPT-5.6 Sol has drawn criticism for burning through limits quickly because, even at the same reasoning level, it uses more tokens, runs tool calls in parallel, and keeps sessions going longer. On July 29, Tibo did a limit reset for the same reason, stressing that "we have never cut quotas on any plan," and said optimizations would extend effective usage time by about 18%.
| Item | July 29 reset | August 9 reset |
|---|---|---|
| Background | Sol burning through limits faster than expected | Claude Code harness account-suspension controversy |
| Who's covered | All ChatGPT Work and Codex users | Paid ChatGPT Work and Codex users |
| Accompanying measures | About 18% efficiency gain, 5-hour rolling limit restored | Immediate reset on the day of the announcement (one-time) |
| Follow-up controversy | Fears of reduced limits | "Performance reset" criticism |
Three practical points come out of this. First, if you're a paid user with plenty of balance left, a full reset like this actually costs you. Until something like the "banked reset" card Tibo mentioned, which lets users trigger a reset when they want, takes hold, you're better off waiting for the automatic reset timing. Second, when you use a proxy path, you have to account for the security risk that your repository, environment variables, and credentials pass through the proxy. Third, "Sol in CC" isn't entirely the same as the native Codex experience, because, as Boris pointed out, harnesses assume per-model tuning.
Cause of the Ban and Further Resets Still Undecided
There are parts of this incident we should explicitly leave as "unconfirmed." First, it hasn't been proven that the account suspension was caused by proxy use. Even a report pointing out that the ban cause is unverified only raises the possibility of a fraud and abuse classifier false positive, with no evidence that the proxy itself was the cause. Second, many reports agree that the reset targets "all paid ChatGPT Work and Codex users," but the coverage of specific plans and the exact timing can't be settled until the full original X post is checked directly. The quotes in this article are partial phrases confirmed through translated reports. Third, it also isn't yet confirmed whether the August 9 reset is a one-off, or whether the "additional reset on Monday" Tibo hinted at will actually happen.
In short, this incident reads as the opening shot of an era in which the harness, not the model, becomes the platform. OpenAI aggressively promoted that its model works in a rival's harness, and Anthropic confirmed model neutrality for harnesses as its official position. Even if the reset itself isn't worth much, the separation of harness and model has effectively been sanctioned, a variable that could change the developer tools landscape in the second half of 2026.