OpenAI opens Fast mode to 272K+ token prompts, 2.5x faster
For GPT-5.6 Sol, Terra and Luna, at twice the price. Plus: Vercel's egress firewall reaches the free tier, and grok-voice-latest quietly swaps models.
Copy markdown
The change: big prompts skip the slow lane
Prompts over 272K tokens can now run in Fast mode on GPT-5.6 Sol, Terra and Luna, up to 2.5x faster than the standard tier. Whole-repo reads and long-document calls that were pinned to standard speed no longer are.
The cost: you pay 2x for the speed
Fast mode bills at twice the standard rate, so treat it as a latency dial, not a free lunch. Reach for it on interactive, agent-in-the-loop calls where wall-clock hurts, and leave batch jobs on standard.
Vercel's egress firewall drops to the free tier
Hobby-plan Sandboxes now get the full egress firewall: policies, credentials brokering, request proxying, and fine-grained matchers. Locking down what agent-written or untrusted code can phone home to is now free.
Elsewhere: grok-voice-latest now points at think-fast-2.0
As of Aug 5, xAI's grok-voice-latest alias resolves to grok-voice-think-fast-2.0, its speech-to-speech model. If you pin the alias in a voice agent you're already on the new model, so re-test latency and turn-taking.