GLM-52 897 —
GPT-56SC 873 —
CL-OP5X 865 -0.9%
GROK-46H 865 -0.9%
GEM-37FH 865 -0.9%
GPT-56T 861 —
GLM-5 856 —
MUSE-SPK 841 —
QWEN-38X 824 -2.3%
GPT-6A 820 —
KIMI-K3X 810 -1%
CL-FAB5H 787 -0.9%
CL-OP5H 764 -0.9%
CL-OP46H 742 -0.9%
CL-OP47H 733 -1.1%
GEM-38FH 676 -1%
CL-OP47 586 -0.5%
INKL 531 —
CL-OP46 497 —
CL-OP48 490 -0.2%
GLM-52 897 —
GPT-56SC 873 —
CL-OP5X 865 -0.9%
GROK-46H 865 -0.9%
GEM-37FH 865 -0.9%
GPT-56T 861 —
GLM-5 856 —
MUSE-SPK 841 —
QWEN-38X 824 -2.3%
GPT-6A 820 —
KIMI-K3X 810 -1%
CL-FAB5H 787 -0.9%
CL-OP5H 764 -0.9%
CL-OP46H 742 -0.9%
CL-OP47H 733 -1.1%
GEM-38FH 676 -1%
CL-OP47 586 -0.5%
INKL 531 —
CL-OP46 497 —
CL-OP48 490 -0.2%
← Back to feed

OpenAI Fixes GPT-5.6 Sol Token Drain: Codex Usage Now Lasts 18% Longer

OpenAI reset usage limits for all ChatGPT Work and Codex subscribers on July 29 and announced a set of efficiency improvements to GPT-5.6 Sol. Expected result: typical usage lasts around 18% longer. Some power users running complex workflows should see significantly larger improvements.

What Caused the Drain

Sol is a more agentic model than its predecessors. Compared to GPT-5.5, it is more willing to extend tool chains, coordinate subagent workflows, and run parallel operations — behaviors that are valuable on hard tasks but expensive on token budgets. Three specific root causes were identified:

  1. Extended agentic willingness. Sol runs more tool calls and coordinates more complex subagent workflows on difficult tasks. This is by design — it’s what makes Sol better at multi-step problems — but it caused some tasks to consume far more tokens than OpenAI anticipated.

  2. Higher effort at equivalent settings. Sol works harder than GPT-5.5 at the same named effort tier. A session set to “High” on Sol consumes more tokens than “High” on GPT-5.5.

  3. Code mode efficiency. Programmatic tool calling (code mode) gave Sol more flexibility to run tool calls in parallel and continue working while waiting for results. This introduced more responses per turn, more cached input tokens, and higher per-session consumption than expected.

The impact was uneven. Median Codex users were largely unaffected. Power users running multi-hour complex workflows — particularly those using code mode — saw quotas drain in hours. $200/month Codex Pro subscribers were hit hardest.

What Changed

OpenAI improved how Sol handles tool-call waiting and web search parallelism, two of the highest-cost behaviors. The company acknowledged it focused on average and median usage before launch and “missed some cases where the long tail could use significantly more usage.”

The five-hour session limit was also restored. It had been paused while OpenAI investigated the root cause.

Context

GPT-5.6 Sol launched as OpenAI’s most capable agentic model, posting records on Terminal-Bench 2.1, SWE-bench Verified, and Agents’ Last Exam. Its stronger task-completion behavior relative to prior models is structural — not a bug — and the efficiency fixes address the token accounting without changing what the model is willing to attempt.

OpenAI said it will continue sharing updates as it improves efficiency. This is the second public fix cycle for Sol since launch; the first addressed usage limit resets after the model’s initial deployment.