1 00:00:00,000 --> 00:00:01,480 Okay, we're rolling. 2 00:00:01,540 --> 00:00:05,520 Quick thanks to Jellypod to help make this daily show a reality. 3 00:00:05,573 --> 00:00:10,640 Today's topic is the Ultrafast service tier for GPT 6.1 Sol, 4 00:00:10,720 --> 00:00:13,280 which landed in Codex on October 8. 5 00:00:13,520 --> 00:00:19,680 We'll also cover the 0.162.1 patch, since it followed right on the heels of the 6 00:00:19,760 --> 00:00:24,880 sandbox fixes in 0.162.0 that we covered earlier. 7 00:00:25,217 --> 00:00:26,517 Let's start with what changed. 8 00:00:27,077 --> 00:00:33,277 OpenAI rolled out Ultrafast mode for GPT 6.1 Sol across Codex, 9 00:00:33,517 --> 00:00:36,037 ChatGPT Work, and the Responses API. 10 00:00:36,637 --> 00:00:38,817 It's a service tier, not a new model. 11 00:00:39,077 --> 00:00:41,517 Same Sol, served on Cerebras hardware. 12 00:00:42,297 --> 00:00:46,137 OpenAI's announcement says up to 8x faster than Sol Standard, 13 00:00:46,477 --> 00:00:48,137 with near Astra intelligence. 14 00:00:48,777 --> 00:00:54,317 The 750 output tokens per second figure comes from the August preview of Ultrafast 15 00:00:54,377 --> 00:01:00,577 on GPT 5.6 Sol, so treat that number as a preview ceiling, not a promise for 6.1. 16 00:01:00,875 --> 00:01:05,435 Why this matters for Codex specifically comes down to how agents spend time. 17 00:01:05,715 --> 00:01:08,075 A single answer isn't the problem. 18 00:01:08,182 --> 00:01:12,555 An agent working through a failing test suite might take eight to twenty sequential 19 00:01:12,648 --> 00:01:18,555 turns: read the output, open a file, run the linter, run the tests again. 20 00:01:18,699 --> 00:01:23,835 Each turn waits on token generation, and that wait compounds into minutes of sitting 21 00:01:23,875 --> 00:01:24,155 there. 22 00:01:24,475 --> 00:01:27,515 A faster tier shrinks every one of those turns. 23 00:01:27,655 --> 00:01:32,635 And the older fast options generally meant dropping to a smaller model like Luna. 24 00:01:32,811 --> 00:01:37,755 Here you keep Sol's coding and tool calling, and you only change how quickly it 25 00:01:37,808 --> 00:01:38,555 comes back. 26 00:01:39,003 --> 00:01:40,283 So, how do you turn it on? 27 00:01:40,823 --> 00:01:47,323 In the CLI, the persistent route is your config file, at ~/.codex/config.toml. 28 00:01:48,003 --> 00:01:50,983 Set service_tier equal to the string ultrafast. 29 00:01:51,563 --> 00:01:54,443 If you'd rather flip it per session, there are the slash commands, 30 00:01:54,563 --> 00:01:56,243 /fast and /ultrafast. 31 00:01:56,923 --> 00:02:00,863 On compatible CLI and Bedrock routes there's also a command line flag for the 32 00:02:00,923 --> 00:02:04,443 service tier, which landed in pull request 50472. 33 00:02:05,043 --> 00:02:09,503 And if you're scripting Codex through the SDK, the per turn setting is 34 00:02:09,583 --> 00:02:13,403 turn_service_tier, set to ultrafast on the turn call. 35 00:02:13,863 --> 00:02:17,343 That one only affects the single turn you start, which is handy for reserving the 36 00:02:17,383 --> 00:02:19,843 expensive tier for the slow, heavy steps. 37 00:02:20,167 --> 00:02:23,767 There's a networking detail that's easy to miss if you automate. 38 00:02:23,847 --> 00:02:30,167 OpenAI's API docs strongly recommend WebSockets, especially for agentic apps that 39 00:02:30,199 --> 00:02:33,047 make many tool calls in quick succession. 40 00:02:33,087 --> 00:02:37,767 Their wording is that without a persistent connection, network overhead can reduce 41 00:02:37,807 --> 00:02:38,967 the latency gains. 42 00:02:39,107 --> 00:02:40,807 The reasoning is simple. 43 00:02:40,914 --> 00:02:46,967 If every tool result goes out as a fresh HTTP request, you pay connection setup each 44 00:02:47,047 --> 00:02:51,127 time, and that overhead can eat the savings from faster generation. 45 00:02:51,340 --> 00:02:58,007 In Python, that means using client.responses.connect(), setting service_tier to 46 00:02:58,100 --> 00:03:03,927 ultrafast on each response.create, and passing the previous response ID forward for 47 00:03:03,980 --> 00:03:06,407 later turns, along with tool results. 48 00:03:06,677 --> 00:03:11,157 The docs example uses model names for Astra, but the same pattern applies to Sol. 49 00:03:11,677 --> 00:03:16,977 And one more Bedrock note: Codex CLI 0.162.0 and 50 00:03:17,357 --> 00:03:22,877 0.162.1 added Ultrafast service tier catalog support for Amazon Bedrock, 51 00:03:23,237 --> 00:03:25,117 covering the Astra and Sol deployments. 52 00:03:25,860 --> 00:03:30,840 Now the caveats, because there are several, and the first is about who can use it at 53 00:03:30,920 --> 00:03:31,100 all. 54 00:03:31,620 --> 00:03:37,820 In Codex and ChatGPT Work, access is on the Pro $500 plan, eligible Enterprise 55 00:03:37,900 --> 00:03:39,900 plans, and credit based Edu plans. 56 00:03:40,460 --> 00:03:46,180 Standard Plus at $20 and Pro at $200 don't have it, even if you've bought extra 57 00:03:46,220 --> 00:03:46,540 credits. 58 00:03:47,000 --> 00:03:50,100 So if a teammate says the setting does nothing, check the plan first. 59 00:03:50,620 --> 00:03:51,480 Then the cost. 60 00:03:52,060 --> 00:03:56,920 Ultrafast burns your included subscription allowance at 8x the Standard rate. 61 00:03:57,560 --> 00:04:01,580 Purchased credits and Enterprise pay as you go are billed at 6x. 62 00:04:02,340 --> 00:04:09,080 For API reference, GPT 6.1 Sol Ultrafast is listed at 12 dollars per million 63 00:04:09,180 --> 00:04:15,000 input tokens and 60 dollars per million output tokens, which the announcement frames 64 00:04:15,060 --> 00:04:17,660 as about 1.2x the cost of Astra. 65 00:04:18,380 --> 00:04:21,260 So don't leave it on as a global default for everything. 66 00:04:21,900 --> 00:04:25,800 Use it where waiting actually hurts: debugging a live outage, 67 00:04:26,360 --> 00:04:30,860 a tight edit and test loop, or an agent working through a big directory tree. 68 00:04:31,277 --> 00:04:35,457 On the admin side, Enterprise workspaces ship with Ultrafast off. 69 00:04:36,037 --> 00:04:40,037 Workspace owners have to enable it, so a developer won't see it just because the 70 00:04:40,077 --> 00:04:40,957 plan qualifies. 71 00:04:41,337 --> 00:04:47,677 And for data residency, GPT 6.1 Sol Ultrafast supports the United States and Europe, 72 00:04:47,977 --> 00:04:49,817 meaning the EEA and Switzerland. 73 00:04:50,297 --> 00:04:53,117 GPT 6 Astra Ultrafast is US only. 74 00:04:53,737 --> 00:04:58,097 If you're in a regulated EU workspace, that difference decides which model you can 75 00:04:58,137 --> 00:04:58,377 pick. 76 00:04:58,667 --> 00:04:59,547 Now the patch. 77 00:04:59,747 --> 00:05:06,587 Codex CLI 0.162.1 shipped October 9, and it 78 00:05:06,680 --> 00:05:09,787 fixes two things that could ruin your morning. 79 00:05:09,840 --> 00:05:15,467 First, pull request 52648, a startup bug with the background daemon. 80 00:05:15,707 --> 00:05:20,427 If a daemon was already running with feature settings that differed from the CLI's 81 00:05:20,489 --> 00:05:23,387 defaults, a new client could fail to launch. 82 00:05:23,473 --> 00:05:28,827 Compatibility checks now apply only to explicit command line overrides, 83 00:05:28,907 --> 00:05:32,747 so a new terminal can attach to a running daemon across updates. 84 00:05:32,987 --> 00:05:36,587 If you got locked out after an upgrade, this is the fix. 85 00:05:37,257 --> 00:05:40,137 Second, pull request 51866. 86 00:05:40,797 --> 00:05:45,197 The terminal UI could panic when an asynchronous clarification question contained 87 00:05:45,357 --> 00:05:46,097 multiple lines. 88 00:05:46,537 --> 00:05:50,917 That's fixed, and newlines and complete hyperlink targets now render properly. 89 00:05:51,577 --> 00:05:55,337 If your agent asks a long, structured question while it works in the background, 90 00:05:55,697 --> 00:05:56,717 you won't lose the terminal. 91 00:05:57,042 --> 00:05:59,002 Two smaller things worth knowing. 92 00:05:59,095 --> 00:06:03,922 In the agent command center, you can now pin tasks with the p key, 93 00:06:03,954 --> 00:06:07,282 from pull request 51500. 94 00:06:07,442 --> 00:06:12,882 And if mouse scrolling in the fullscreen UI feels off, there's a config setting, 95 00:06:12,962 --> 00:06:17,122 tui.mouse_scroll_speed, to tune the wheel. 96 00:06:17,325 --> 00:06:24,165 So the practical takeaway: update to 0.162.1, check your plan and workspace 97 00:06:24,205 --> 00:06:30,005 settings, then try ultrafast on one heavy loop and watch how fast your allowance 98 00:06:30,065 --> 00:06:30,385 drains. 99 00:06:31,045 --> 00:06:34,945 If it's a good trade for that workflow, keep it for that workflow.