AI Models

GPT 5.6 Launching SOON! (What to Expect)

Published
Jul 8, 2026
Duration
11:39
Module
AI Models
Click to load the YouTube player

Reviewed companion

Useful notes, receipts, and next steps

Format
opinion
Reviewed
Jul 18, 2026

TL;DR

  • The video presents GPT 5.6 as three role-based models: Soul for difficult reasoning and long-running agent work, Terra for everyday execution, and Luna for cheap repetitive jobs. (source video BRThrYEiccA, 01:15)
  • Early-access reactions sounded strong on persistence, sub-agent orchestration, computer use, iOS work, and front-end design, but they were tester reports—not Ron’s completed benchmark. (source video BRThrYEiccA, 02:44; source video BRThrYEiccA, 03:02; source video BRThrYEiccA, 03:23; source video BRThrYEiccA, 03:39; source video BRThrYEiccA, 05:39)
  • Ron’s practical routing idea was Luna for intake, Terra for most execution, and Soul for the first plan or work that is ambiguous, long-horizon, or important. (source video BRThrYEiccA, 07:29)
  • Max and Ultra were presented as different tools: reserve Max for deep planning; use Ultra when a complex prompt benefits from sub-agents and decomposition. (source video BRThrYEiccA, 04:12; source video BRThrYEiccA, 04:58)
  • Cybersecurity and biological work were expected to meet more safeguards and possible friction. Ron’s GPT-5.5 fallback idea was explicitly an assumption, not an announced routing rule. (source video BRThrYEiccA, 08:41; source video BRThrYEiccA, 10:52)

Ron’s verdict

Do not replace every model in your stack with Soul on day one. The interesting part of GPT 5.6 was the proposed division of labour: expensive thinking at the top, a balanced execution layer in the middle, and cheap background work at the bottom. Soul looked like the model to test first for planning and orchestration, but the video was an expectations brief built from positioning and early-access reactions. (source video BRThrYEiccA, 01:39; source video BRThrYEiccA, 07:22; source video BRThrYEiccA, 09:59) Companion test plan: the default has to be earned with real tasks, measured cost, and direct checks of the files and actions the agent produces.

Key moments

Useful quotes

“the big thing to understand is that GPT 5.6 is not one model.” — Ron, source video BRThrYEiccA, 00:12

“I think the only time you really want to go for max reasoning effort is on the first prompt when you’re planning it.” — Ron, source video BRThrYEiccA, 04:58

“specify your user story, your user intention, not just the user requirement.” — Ron, source video BRThrYEiccA, 08:13

“Nothing here says they’ll be routing it to Tara.” — Ron, source video BRThrYEiccA, 10:56

Separate the claims from the test plan

This is a pre-launch expectations piece. Its most useful move is to keep vendor positioning, third-party reactions, Ron’s operating advice, and Ron’s speculation in separate boxes.

What the video discussesEvidence status in the videoWhat an operator should do
Soul for hard reasoning and long-horizon agents; Terra for balanced daily work; Luna for fast, cheap scalePresented as the announced family positioning. (source video BRThrYEiccA, 01:15)Test each role independently; do not assume the family label proves task quality.
Strong persistence, orchestration, computer use, iOS knowledge, and front-end designAttributed to early testers Theo and Skirano. (source video BRThrYEiccA, 02:42; source video BRThrYEiccA, 03:39; source video BRThrYEiccA, 05:39)Re-run your own long agent task, computer-use task, and front-end brief.
Soul at $5 input / $30 output per million tokens; Terra at half; Luna at $1 input / $6 outputPrices read and interpreted in the video. This companion does not confirm current pricing. (source video BRThrYEiccA, 07:07)Measure total completed-task cost, including retries and sub-agent work.
Sensitive-domain requests may be slower or more restrictedRon’s expectation from the described safety and monitoring controls. (source video BRThrYEiccA, 08:41)Keep defensive intent explicit and record refusals, delays, or degraded results.
Weak coding might be routed to GPT-5.5Ron’s stated worst-case assumption; he corrects himself that no Terra routing was implied. (source video BRThrYEiccA, 10:38; source video BRThrYEiccA, 10:52)Treat poor output as a diagnosis prompt, not proof of hidden routing.

A practical role-routing decision aid

The family only saves money if the hand-offs are deliberate.

Work arrivingFirst model to testAcceptance check
Intake, classification, or repetitive background workLunaDid it route correctly and preserve every required field? (source video BRThrYEiccA, 07:34)
Normal code execution with a settled planTerraRun the code and inspect the changed files; “done” is not evidence. (source video BRThrYEiccA, 07:36)
First plan, ambiguous architecture, or long-horizon workSoulCheck the plan against the user story and intention before execution. (source video BRThrYEiccA, 07:42; source video BRThrYEiccA, 08:07)
One prompt with many separable componentsSoul on UltraConfirm the sub-tasks recombine into one coherent result. (source video BRThrYEiccA, 05:28)
Defensive cyber or biological-domain workStart cautiously and log behaviorMake benign intent clear; expect additional controls, but do not guess which hidden model answered. (source video BRThrYEiccA, 09:34; source video BRThrYEiccA, 10:52)

What to validate before changing your default

  1. Give Soul one real planning task and one long-running agent task. Record whether it keeps the goal without repeated steering. This tests the early-access persistence claim. (source video BRThrYEiccA, 02:44)
  2. Compare Medium, High, and Max on the same plan. Ron’s starting hypothesis was that the highest effort earns its cost mainly on the first planning prompt. (source video BRThrYEiccA, 04:52)
  3. Use Ultra on a genuinely decomposable build. Inspect both the final integration and any gaps between sub-tasks. (source video BRThrYEiccA, 04:17)
  4. Price the completed workflow, not the headline token rate. Include retries, long reasoning, and sub-agent usage. This extends the video’s expensive-thinking/cheap-execution framing without claiming a measured result. (source video BRThrYEiccA, 06:57)
  5. Test a front-end task and a computer-use task yourself. Those were prominent tester reactions, not results demonstrated in this video. (source video BRThrYEiccA, 03:41; source video BRThrYEiccA, 05:39)

What changed since this video

The video was published July 8, 2026 and spoke about a launch “this Thursday” and “tomorrow.” This companion was source-checked against the immutable full transcript and 571 timestamp segments on July 18, 2026. No post-launch model card, release notes, pricing page, availability check, or fresh benchmark was added. Treat the model names, prices, launch timing, tester reactions, and capability expectations above as a dated record of the video—not confirmation of the product state now.

Watch on YouTube

Prefer the native player? Open it on YouTube: https://www.youtube.com/watch?v=BRThrYEiccA