In partnership with
Reading time: 7 minutes
Greetings from above,
I have been watching AI model launches for a while now and most of them follow the same pattern. Company drops a blog post. Numbers go up on a benchmark. People argue on X for three days. Then everyone forgets about it and waits for the next one.
This one is actually different. And I want to walk you through exactly why so you can decide what it means for your work right now.
Today we will talk about:
What GPT-5.6 Sol, Terra and Luna actually are and which one you probably need
The new voice model that works nothing like anything OpenAI has shipped before
The government angle nobody is talking about and why it matters for how you access this
Let's get into it.
You've seen the AI demos. Viktor does it without you watching.
The AI tool you tried last quarter waited for a prompt, hallucinated a number, then asked if you'd like a summary.
Viktor opened a PR at 2am, rebased it against main, ran your test suite, and posted a note in #eng: "Two flaky tests in payments service, both pre-existing. Recommended merging after fixing them." Then drafted the customer reply for the support ticket the bug created.
That's 619K autonomous actions per day across 20,000+ teams. Not chat replies. Real work shipped to GitHub, Stripe, Linear, Notion, and 3,000+ other tools, from inside Slack and Microsoft Teams.
You don't supervise him any more than you supervise a senior engineer.
SOC 2 certified. Your data never trains models.
"It's what you probably originally thought AI was going to be when you first heard of it in sci-fi movies." Tyler, CEO.
So What Even Is GPT-5.6
Ok so here is the simple version of what just happened.
OpenAI did not just drop one model. They dropped three at once and gave them names instead of just numbers. Sol is the flagship, the most powerful one. Terra is the middle option, balanced for everyday work and 2x cheaper than GPT-5.5. Luna is the fast affordable one that still punches well above its price.
The naming system is actually kind of smart when you think about it. The number tells you the generation, so 5.6 means this is the sixth iteration of the fifth generation. Sol, Terra and Luna are permanent tier names that will keep updating on their own schedule. So next time OpenAI improves the flagship it will still be called Sol, just a newer version of Sol. You do not have to remember a new name every three months which is honestly a relief.
The pricing landed like this. Sol is $5 per million input tokens and $30 per million output tokens. Terra is $2.50 input and $15 output. Luna is $1 input and $6 output. For context Sol is actually cheaper than Fable 5 on input tokens, which costs $10, and roughly comparable on output. If you are already paying for frontier model access this pricing is not going to shock you.
HOW GPT-5.6 CAN ACTUALLY HELP YOUR BUSINESS RIGHT NOW:
Sol's new ultra mode uses multiple subagents working in parallel on complex tasks which means work that used to take hours of back and forth can now run in the background while you do other things
Terra gives you near GPT-5.5 performance at half the cost which makes it the obvious choice for high-volume workflows where you do not need maximum intelligence on every single call
Luna brings real capability at the lowest price point OpenAI has ever offered for a model this capable which makes previously expensive automation actually affordable to run at scale
The Part That Actually Surprised Me
GPT-5.6 Sol is not just faster or bigger than what came before. It introduces two things that are genuinely new.
The first is max reasoning effort. This gives Sol the maximum possible time to think through a problem before responding. Not just more compute, actually more structured reasoning time before the output comes out. On Terminal-Bench 2.1, which tests real command-line workflows involving planning, tool coordination and iteration, Sol Ultra scored 91.9% and standard Sol scored 88.8%. For comparison Claude Mythos 5 scored 84.3% and Fable 5 scored 83.4%. These are not small gaps on a hard benchmark.
The second is ultra mode. This goes beyond what a single agent can do by spinning up subagents that work in parallel and accelerate complex tasks. So instead of one model working through a big problem step by step you get multiple models dividing the work and coordinating the results. This is the kind of thing that previously required you to build your own multi-agent architecture from scratch.
On cybersecurity tasks specifically Sol is doing something remarkable. On ExploitBench it matches Anthropic's Mythos Preview performance while using only about a third of the output tokens. That means you get equivalent results for roughly a third of the cost on security work.
The Voice Model That Works Nothing Like The Old One
Ok so separately from GPT-5.6 OpenAI also launched GPT-Live and this is honestly the more interesting product for most people even if it got less attention in the launch coverage.
Here is the thing about every voice AI before this one. They all worked in turns. You talk, then you stop, then the AI talks, then you stop. The AI had to wait for silence before it knew you were done. If you paused to think for half a second it would jump in. If there was background noise it would get confused and respond to that instead.
GPT-Live works on what they call a full-duplex architecture. This means it listens and speaks at the same time, continuously, the way an actual person does in an actual conversation. It can hear you pause mid-thought and just wait. It can say "mhmm" or "got it" while you are still talking so you know it is following along. You can interrupt it mid-sentence with a question and it adjusts immediately without losing the thread.
On top of that GPT-Live can delegate to a more powerful model in the background when you ask something that needs real reasoning or web search. So you ask a hard question, it keeps talking with you naturally, and a few seconds later it brings the answer back into the conversation without making you wait in silence while it thinks. More than 150 million people use ChatGPT Voice every week. This is a significant upgrade for all of them.
There are two versions. GPT-Live-1 is the main one that becomes the default for Plus, Pro and Go subscribers. GPT-Live-1 mini becomes the default for free users. Both are already rolling out globally on iOS, Android and the web.
⚙️ Here Is Everything Broken Down Step By Step
What Sol is for and when to use it
Sol is the model you reach for when the task is genuinely hard. Complex code migrations across a large codebase. Long-horizon research that needs to hold a lot of context across many steps. Security analysis where you need to find real vulnerabilities. Biology and genomics work where the analysis requires deep multi-step reasoning.
If the task would take a skilled person a full day or more and you want the AI to handle it mostly autonomously Sol with ultra mode is probably the right tool. It is not the model you run for everything. It is the model you escalate to when the cheaper options are not quite getting there.
What Terra is for and when to use it
Terra is probably the model most people reading this should use most of the time. It runs at competitive GPT-5.5 level performance at half the price. For content creation, research summaries, customer communications, standard coding tasks, data analysis, and the bulk of everyday business work Terra is the default smart choice. It is also the one that makes high-volume automation economically sensible in a way that Sol simply is not at $30 per million output tokens.
What Luna is for and when to use it
Luna is the speed play. Fast responses, lowest cost, still genuinely capable. For any workflow where you need a lot of AI calls at low latency and the tasks are relatively straightforward Luna is the one that makes the unit economics work. Think automated customer triage, bulk content classification, quick summarization pipelines, or anything where you are making thousands of calls and cost per call actually matters.
How to actually use GPT-Live right now
If you are on a Plus, Pro or Go plan it is already rolling out to you today. Open ChatGPT on your phone, tap the voice button and you will either already have GPT-Live or you will get it in the next few days as the rollout completes.
Free users get GPT-Live mini which is still a significant upgrade over the old Advanced Voice Mode. You can choose your reasoning level too. Instant for fast casual responses, Medium when you want it to think a bit harder before answering, and High when you genuinely need the best possible answer and are willing to wait a few extra seconds for it.
How to access GPT-5.6 Sol right now if you are not in the preview
Here is the honest answer on this. Right now GPT-5.6 Sol is in a limited preview for a select group of trusted partners. It is available through the API and Codex but not broadly through ChatGPT yet.
OpenAI says general availability is coming in the next few weeks. If you are a developer you can sign up through the API to get notified. If you are a regular ChatGPT subscriber you are likely looking at a few weeks before Sol shows up in your model picker.
The prompt caching upgrade you should know about
GPT-5.6 also introduced something called explicit cache breakpoints. You can now specify exactly where in your prompt you want the cache to break, and the minimum cache life is 30 minutes. Cache writes cost 1.25x the standard input rate but cache reads get the same 90% discount as before. For anyone running long system prompts or large context documents across many API calls this meaningfully reduces the cost of repeated context. Worth setting up properly before you start routing high-volume work through 5.6.
The Cerebras announcement that most people skipped over
OpenAI is launching GPT-5.6 Sol on Cerebras hardware in July at up to 750 tokens per second. For context that is an almost absurd speed for a frontier model. Standard API speeds are nowhere near that. This matters for real-time applications where latency is the constraint not cost. If you are building anything that needs frontier-level intelligence at conversation speed this is the infrastructure story worth watching. Access starts limited and expands as capacity grows.
What the government preview process means for you
This one is worth understanding properly because it is going to affect how future models launch. OpenAI previewed GPT-5.6 with the US government before the public launch at the government's request. They started with a limited group of trusted partners whose participation was shared with the government before releasing more broadly.
OpenAI was pretty direct in their statement that they do not believe this should become the long-term default because it keeps the best tools away from users and developers who need them. But they are doing it now as a short-term step while working with the administration on a cyber Executive Order framework.
The practical impact is that the full public release is a few weeks away rather than immediate. The longer-term implication is that government involvement in model launches is now an established pattern at OpenAI, which is a meaningful change from how things worked a year ago.
GPT-5.6 FULL LAUNCH BREAKDOWN — SUMMARY
Sol, Terra and Luna give you three clear tiers with obvious use cases rather than making you guess which model fits which task and the pricing at $5 / $30, $2.50 / $15 and $1 / $6 per million tokens is competitive enough that switching costs for existing OpenAI users are basically zero
GPT-Live is not just a better voice mode, it is a fundamentally different architecture that makes voice AI actually feel like a conversation for the first time and it is already rolling out to 150 million weekly voice users starting today
The government preview process is real, it is documented, and it is the reason Sol is not in your ChatGPT model picker yet — it will be in a few weeks and the capabilities benchmarks suggest it is worth waiting for
WRAP UP
What you learned today:
GPT-5.6 is a three-model family not a single release and the Sol/Terra/Luna naming system is designed to stay stable over time so you always know which tier you are talking about regardless of which generation you are on
GPT-Live's full-duplex architecture solves a problem that has existed in voice AI since the beginning and the real-world test results show it is preferred over Advanced Voice Mode by a significant margin across naturalness, flow, and listening quality
The limited government preview is a short-term access constraint not a permanent gate and once Sol hits general availability in the next few weeks the benchmark numbers suggest it is genuinely the strongest model currently available for complex coding, cybersecurity and long-horizon reasoning tasks
The AI race has not slowed down even slightly. OpenAI just shipped a family of models with a new naming system, a new voice architecture, new multi-agent ultra mode, government-coordinated launch processes, and Cerebras hardware doing 750 tokens per second. All in one week.
If you are building anything with AI right now you need to know which of these three tiers your work actually belongs in and start testing accordingly. The pricing is accessible enough that there is no real excuse to still be routing everything through one default model when three clearly differentiated options are sitting right there.
Pick your tier. Test the voice mode. Watch for Sol in your model picker.
And as always, thanks for being a part of my lovely community,
Keep learning,
🔑 Robert from God of Prompt
