1 00:00:00,280 --> 00:00:01,020 Morning, everyone. 2 00:00:01,660 --> 00:00:06,060 Before we get into it, thanks to Jellypod for helping make this daily show a 3 00:00:06,100 --> 00:00:06,660 reality. 4 00:00:07,700 --> 00:00:12,340 Today's topic is a model, but it's one with real workflow consequences. 5 00:00:13,320 --> 00:00:20,100 On October 26, OpenAI announced GPT 6.1 Terra, and it's available right 6 00:00:20,160 --> 00:00:25,080 now in Codex for Plus, Pro, Business, Enterprise, and Edu users. 7 00:00:25,840 --> 00:00:28,680 It's also on Amazon Bedrock and in the API. 8 00:00:29,260 --> 00:00:35,900 The ID is lowercase and hyphenated: gpt, 6.1, terra. 9 00:00:36,840 --> 00:00:41,760 It replaces GPT 5.6 Terra as the mid tier, efficiency focused model. 10 00:00:42,485 --> 00:00:48,965 And this connects to our earlier episode on inline model switching in Codex 0.164.0. 11 00:00:49,605 --> 00:00:53,545 If you remember, that release let you pass the model, reasoning effort, 12 00:00:53,605 --> 00:00:57,865 and permissions directly to the slash model and slash permissions commands. 13 00:00:58,585 --> 00:01:01,005 Well, now there's a model worth switching to. 14 00:01:01,665 --> 00:01:06,365 You can test Terra immediately, either with the inline model command or from a 15 00:01:06,445 --> 00:01:08,905 session preset, without opening the picker at all. 16 00:01:09,610 --> 00:01:09,750 Right. 17 00:01:10,150 --> 00:01:14,330 So let's start with what changed, because the headline numbers are specific. 18 00:01:15,070 --> 00:01:20,770 On Terminal Bench 4.0, Terra scores 50.3 percent at max reasoning effort. 19 00:01:21,550 --> 00:01:25,430 GPT 5.6 Terra scored 21.5 percent. 20 00:01:26,190 --> 00:01:31,070 So that's more than double, on a benchmark that's close to what Codex does all day: 21 00:01:31,610 --> 00:01:32,870 multi step command line work. 22 00:01:33,667 --> 00:01:35,087 Speed is the other half. 23 00:01:35,547 --> 00:01:41,127 OpenAI says Terra finishes multi step computer and terminal tasks over three times 24 00:01:41,167 --> 00:01:45,627 faster on average than GPT 6 Astra and GPT 6.1 Sol. 25 00:01:46,427 --> 00:01:49,847 And the price drop matters for anyone calling the API directly. 26 00:01:50,307 --> 00:01:56,327 Standard pricing is cut by 50 percent compared to both GPT 5.6 Terra and GPT 6.1 27 00:01:56,367 --> 00:01:56,567 Sol. 28 00:01:57,087 --> 00:02:01,867 That's one dollar per million input tokens, five dollars per million output tokens, 29 00:02:02,207 --> 00:02:04,907 and five cents per million cached input tokens. 30 00:02:05,427 --> 00:02:10,467 That cached rate is a 95 percent discount, which is where agent loops tend to live, 31 00:02:10,566 --> 00:02:12,507 since they resend a lot of the same context. 32 00:02:13,255 --> 00:02:18,255 And on a Codex subscription, the claim is that your plan allowance stretches 33 00:02:18,355 --> 00:02:22,095 significantly farther on Terra than on Sol or Astra. 34 00:02:23,235 --> 00:02:27,195 I'd treat that as directional until you watch your own usage meter, 35 00:02:27,775 --> 00:02:31,135 because OpenAI hasn't given a single multiplier here. 36 00:02:31,932 --> 00:02:32,672 Two more specs. 37 00:02:33,112 --> 00:02:37,572 The context window is up to 1,000,000 input tokens, with text and high resolution 38 00:02:37,612 --> 00:02:38,252 image inputs. 39 00:02:38,752 --> 00:02:43,252 And the reasoning controls include low, medium, and max, plus something I think is 40 00:02:43,312 --> 00:02:46,312 underrated: reasoning effort set to none. 41 00:02:47,005 --> 00:02:48,265 Why does none matter? 42 00:02:49,265 --> 00:02:55,625 On GPT 6.1 Sol, the sampling parameters temperature and top p were locked. 43 00:02:56,445 --> 00:03:01,165 On Terra, if you set reasoning effort to none, those unlock. 44 00:03:01,905 --> 00:03:05,825 Developers in the OpenAI community thread confirmed that behavior. 45 00:03:06,765 --> 00:03:12,005 So if you have a pipeline that depends on tuned sampling, say a code formatter or a 46 00:03:12,065 --> 00:03:15,805 classifier step, Terra is the model where you can get that back. 47 00:03:16,923 --> 00:03:17,703 Let's get practical. 48 00:03:18,143 --> 00:03:24,743 In Codex CLI 0.164.0 or newer, type slash model, space, 49 00:03:25,223 --> 00:03:28,023 gpt 6.1 terra, and you're switched. 50 00:03:28,603 --> 00:03:34,283 You can add the effort on the same line: slash model gpt 6.1 terra, 51 00:03:34,743 --> 00:03:36,523 then reasoning equals low. 52 00:03:36,923 --> 00:03:41,783 Or reasoning equals none if you want the fastest path and don't need it to think. 53 00:03:42,343 --> 00:03:46,523 That's the exact pattern the inline arguments in 0.164.0 were built for. 54 00:03:47,428 --> 00:03:49,108 And pair that with presets. 55 00:03:49,988 --> 00:03:54,388 0.164.0 added ten numbered slash presets slots. 56 00:03:55,228 --> 00:03:59,288 A sensible setup is to put Terra in slot one as your daily driver, 57 00:03:59,888 --> 00:04:03,548 with your workspace sandbox permissions attached, so quick edits, 58 00:04:03,808 --> 00:04:07,788 scripts, test writing, and bug fixes are one keystroke away. 59 00:04:08,788 --> 00:04:12,288 Then keep Sol or Astra in another slot for the heavy sessions. 60 00:04:12,988 --> 00:04:15,348 Switching becomes a habit instead of a menu trip. 61 00:04:16,023 --> 00:04:21,643 For AWS shops, GPT 6.1 Terra is generally available on Amazon Bedrock. 62 00:04:22,203 --> 00:04:25,823 You'd use the standard Bedrock endpoints, configured through the Bedrock setup 63 00:04:25,843 --> 00:04:31,303 wizard in Codex, which became enabled by default in 0.164.0. 64 00:04:31,943 --> 00:04:33,103 Now the caveats. 65 00:04:33,783 --> 00:04:39,083 First, OpenAI itself says Sol and Astra remain better for high difficulty 66 00:04:39,243 --> 00:04:43,123 architectural refactoring and subtle algorithmic debugging. 67 00:04:44,023 --> 00:04:46,963 Terra can hit limits on multi layer dependency graphs. 68 00:04:47,603 --> 00:04:51,303 So don't hand it your multi repo migration just because it's cheaper. 69 00:04:51,983 --> 00:04:54,563 On reliability, there's good news with a condition. 70 00:04:55,083 --> 00:05:00,443 On OpenAI's internal error inducing prompt benchmark, low effort error rates fell 71 00:05:00,483 --> 00:05:06,383 from 53.6 percent on GPT 5.6 Terra to 16.7 percent. 72 00:05:06,963 --> 00:05:10,663 At max reasoning, it's 4.3 percent, close to Sol and Astra. 73 00:05:11,283 --> 00:05:14,863 But that's OpenAI's own benchmark, so it's a reason to test, 74 00:05:15,163 --> 00:05:15,803 not a guarantee. 75 00:05:16,563 --> 00:05:20,583 And the system card addendum notes slight variances on some edge case moderation 76 00:05:20,663 --> 00:05:25,383 prompts, hate and extremism tests in particular, which dropped a little relative to 77 00:05:25,423 --> 00:05:30,123 the Astra baseline but stayed within what OpenAI calls acceptable production 78 00:05:30,163 --> 00:05:30,663 thresholds. 79 00:05:31,350 --> 00:05:32,610 One practical gotcha. 80 00:05:33,290 --> 00:05:38,550 If you're on a client older than 0.163.0 and the catalog is cached, 81 00:05:38,950 --> 00:05:40,010 Terra may not show up. 82 00:05:40,650 --> 00:05:44,190 A daemon restart or a manual catalog refresh should populate it. 83 00:05:44,950 --> 00:05:46,650 Easiest fix is to just update. 84 00:05:47,187 --> 00:05:48,207 A few smaller things. 85 00:05:48,827 --> 00:05:53,047 Codex Security 0.7.0 shipped on October 26. 86 00:05:53,607 --> 00:05:58,387 It adds a validate flag, written with two dashes, that runs a supplemental pass 87 00:05:58,467 --> 00:06:02,207 after a scan and saves a separate validation dot md report. 88 00:06:02,847 --> 00:06:07,147 It makes extra model calls, leaves your findings and fail on severity decision 89 00:06:07,207 --> 00:06:11,247 unchanged, and can't be combined with a max cost cap, dry run, 90 00:06:11,467 --> 00:06:12,147 or mock. 91 00:06:12,992 --> 00:06:16,012 That same release has upgrade notes worth reading. 92 00:06:16,772 --> 00:06:18,972 External imports are now opt in. 93 00:06:19,732 --> 00:06:26,492 Set CODEX_SECURITY_ENABLE_EXTERNAL_IMPORTS equal to 1 before using publish findings 94 00:06:26,612 --> 00:06:31,512 or the GitHub advisories import, and only the literal value 1 works. 95 00:06:32,112 --> 00:06:37,132 The inventory helpers moved from Python to Node, so if you called those scripts 96 00:06:37,252 --> 00:06:40,652 directly, switch to the bundled launcher with its helper flag. 97 00:06:41,212 --> 00:06:44,632 And committed diff scans now need a clean, full checkout. 98 00:06:45,503 --> 00:06:51,963 Last one: Codex CLI 0.166.0 alpha 1 and alpha 2 landed October 99 00:06:52,123 --> 00:06:52,643 27. 100 00:06:53,343 --> 00:06:55,323 Alphas, so no need to chase them. 101 00:06:55,723 --> 00:06:58,923 The latest stable is still 0.164.0. 102 00:06:59,803 --> 00:07:06,803 So the takeaway: update to 0.164.0 or newer, run slash model gpt 103 00:07:07,103 --> 00:07:12,043 6.1 terra on a routine task, and put it in a preset slot if it holds up. 104 00:07:12,743 --> 00:07:13,443 That's it for today.