OpenAI DevDay 2026: What Changes Your Work (And What's Just Hype)

0:00 10 million tokens in, two million out. On GPT6 Astra, that job costs $200. On GPT6.1 Soul, the model Open AAI released at dev day on September 29th, the same job cost 40. And OpenAI says Soul

0:15 matches Astra on a coding benchmark called Deepswe. That's OpenAI's own result, so test it before you trust it. Still, it's a bold thing to claim about a model that costs

0:26 a fifth as much. That was one announcement. The same keynote shipped agents that drive a browser, an API that only answers from a list you write, apps that live inside chat GPT, and a stack

0:36 of codeex updates. And each one wears a different label. Public beta, limited preview, research preview, depends on your plan. By the end, you'll know which of these lowers your bill today, which

0:48 takes infrastructure off your plate, and which you can't touch yet. I'll push one team's workload through all of it. And hold on to one question. Why is there a GPT 6.1 soul but no GPT 6.1 Astra? The

1:00 answer matters once agents start clicking things. The job. A support team runs a nightly pass. Read the day's tickets, sort them, check a vendor site for order status, then post a summary

1:11 where the team already works. Call it 10 million input tokens and 2 million output. Step one, pick the model. Astra charges $10 per million input tokens and 50 per million output. 10 million in is

1:23 $100. 2 million out is another 100. Soul charges two in and 10 out. So 20 plus 20, $40. That's list price arithmetic before batch discounts or tool cost. Now the cash input column where the ratio

1:38 changes. Astro charges a dollar per million cash tokens. Soul charges 10 cents, so a tenth instead of a fifth. So if every call in the run starts with the same long system prompt, that repeated

1:49 chunk costs Soul a tenth of what it costs Astra. Cashing makes both bills smaller on the repeated part. Soul's discount just gets deeper. Two more rows while we're here. GPT6 Luna sits at 10

2:00 cents in and 50 cents out for the tiny stuff. And the ultra fast tier, the faster option, is listed only for Astra at $60 in 300 out. $300 per million output tokens. That's a lot of money to

2:13 spend on waiting less. The other number Open AI gives, Soul lands within 2.1 points of Astra on the OSWorld 2.0 0 benchmark at about a seventh of the cost in chat. GBT soul's available in work

2:26 and codecs that's for plus pro business enterprise and edu but it's not in regular chat yet in the API it's on the price list now for a coding heavy team that deep claim is the one to test first

2:39 rerun a job you already trust on both models and compare the diffs yourself the price gap is on the table the quality gap is what you're checking now the missing Astra techrunch reports

2:50 openai dropped a planned GPT 6.1 Astra After internal safety testing, the testing turned up quote higher levels of deception and a tendency to move forward with tasks without asking the user for

3:02 permission. Keep that last part in mind. Permission is what the next announcement is all about. Step two, hand the whole job to the agents API. An agent needs a harness, the loop that calls the model,

3:13 runs a tool, feeds the result back, and goes again. You used to build and babysit that yourself. Now Open AAI hosts and maintains it. You pick the compute environment, the tools, and the

3:24 knowledge. It supports MCP, the model context protocol for plugging in tools, plus custom functions, built-in tools, and parallel sub aents. Picture one sub agent sorting tickets while another

3:35 checks order status, and you're not maintaining the plumbing between them. Pricing is refreshingly boring. No fee for the agents API itself. You pay for the tokens and tools your agents use.

3:45 Status public beta. The order status check is where computer use comes in. The browser runs on Open AI's side, not yours. When the agent wants to reach a new site, your app receives a browser

3:56 origin access event and you approve that origin, meaning the whole site. These computer use details come from a secondary summary of OpenAI's docs, so treat them as provisional. Approving a

4:07 site doesn't add a confirmation before each action. Approve the vendor portal once, and every click and form after that goes through unless your app adds its own check. So, Open AI shelved a

4:18 model for not asking permission, then shipped a browser that asks once per website. Your origin list is the safety net, so keep it narrow. Some small print for the browser, especially if you're in

4:28 a regulated shop. Containers cost 12 cents per 20 minutes. Data residency is US only. There's no zero data retention, and pass keys and QR code login aren't supported. So, if your vendor portal

4:40 signs in with a pass key, the agent stuck at the login screen. Cost check for the agent version of our job. Tokens at sle prices plus tools plus container time. If the vendor check runs for an

4:52 hour, that's three 20 minute blocks, 36. Next to $40 of tokens, the container times a rounding error. What the agent reads still counts as tokens though. The model choice is where the money is. Step

5:04 three, swap out the sorting step itself. Sorting a ticket into billing, bug, or refund doesn't need a paragraph. OpenAI's new decisions API powered by GPT6 Luna lets you define the questions

5:17 and the possible answers and it picks one. You can classify content, route requests, or choose an agent's next action. Status limited preview. If you've ever fished the word billing out

5:28 of a reply that started with great question, you know why this exists. The new stack reports about 150 milliseconds per call against 1.6 seconds for a standard Luna call. And each answer

5:40 comes back with a confidence score. That's from press coverage. There's also no published price inside our agent. The same kind of call can pick the next step. Reply, escalate, or check the

5:50 vendor site. One question, three allowed answers, and the agent can't wander off into a fourth. If you've watched the Jev videos on this channel, this should look familiar. Typesafe launched Jev on

6:01 September 15th. You send application state plus typed questions like pick one of these and you get typed answers with probabilities instead of generated text. Output tokens are free and the context

6:13 window is 32,000 tokens. The new stack even framed OpenAI's version as a response to Jev. Side by side, it's the same idea, a fixed set of answers with a number saying how sure the model is. The

6:25 difference is what you can do with it today. Jev's already live on Open Router and output cost nothing. Open AI's version is preview only with no price for triage you need running this month.

6:36 I test Jev on a day of real tickets and keep an eye on the decisions API. Step four is where the results show up. Plug-in extensions let your plug-in hook into chat GPT's own interface, the

6:48 sidebar, conversation panels, file viewers and editors, rich forms, deep links, and composer mentions. The doc's example is bits and bolts, a CAD parts library that opens STL files in a viewer

7:00 right inside chat GPT. Our triage summary becomes a sidebar panel right next to the chat where people already ask about those tickets. Availability is where the recap and the docs drift

7:10 apart. The recap says available to all plans. The docs say composer mentions only work in the desktop app and web extensions for free and go users are coming soon. So all plans eventually in

7:22 some apps. Step five is the repo behind all of this. Codeex Cloud now runs tasks while your laptop is closed with access across devices. So you can start a fix for the triage script before you leave.

7:34 Shut the lid and check the diff from your phone. The CLI also gets two-way voice and a / agents view. Code review adds summaries and diffs and runtime wire reports. GitHub support is

7:45 generally available with GitLab in preview. Codex itself comes with plus, pro, business, healthcare, education, and enterprise. Of the Codex updates, the one I'd try first if your workspace

7:57 can get in is Codex Security Cloud. You choose a one-time repository scan or commit monitoring that checks new commits as they land. It validates findings in isolated environments,

8:07 removes duplicates across scans, and lets you review a proposed patch before it becomes a pull request. pointed at the triage jobs repo. It runs in the background and someone reviews patches

8:16 instead of waiting through raw alerts. The catch? It's a research preview on web and desktop and it needs workspace access, a connected GitHub repo and a codeex cloud environment. The whole job

8:28 label by label soul on the API price list now and in chat GPT depending on your plan. Agents API public beta decisions API limited preview no price plug-in extensions depends on plan and

8:43 on which app you're in codeex cloud depends on plan codec security cloud research preview my call if you're paying Astra bills rerun your heaviest job on soul this week it's the one

8:54 announcement that can lower your bill today and if you reuse long prompts the cash rate is a tenth of Astras if you maintain your own agent loop pilot the agents API just Check the hosted browser

9:05 before a regulated team switches it on. That summary of the docs says it's US only with no zero data retention. If you're already in the security cloud preview, turn on commit monitoring. The

9:16 rest bookmark and check the labels again next month. Which leaves the question I'd most like answered. When OpenAI finally puts a price on the decisions API, will it undercut Jev, whose output

9:27 tokens are already