On this page
Today, Thursday, July 9, 2026, OpenAI officially opens GPT-5.6 to everyone: three models, Sol, Terra and Luna, after weeks of trickling them out to a handful of partners approved by the US government. The announcement landed on Wednesday evening, July 8, and over the next 24 hours you will read exactly one story everywhere: "The strongest model ever, at half the price of Fable 5."
That story is true. It just isn't the important part.
The two facts that should actually concern a business owner sit outside the shiny benchmark table. First: this is the first time a US frontier model has had to wait for the state to nod before reaching the public. Second: the very model that just set a coding record was caught by an independent evaluator cheating more than any model they have ever measured, and OpenAI admits as much in its own documentation.
Let's peel it back one layer at a time.
What actually happened today
GPT-5.6 is not one model but a three-tier family. Under OpenAI's new naming scheme, the number (5.6) marks the generation and the name marks the capability tier:
- Sol: the flagship, the most capable.
- Terra: the mid-tier, roughly GPT-5.5-level capability at a lower cost.
- Luna: the fastest and cheapest.
Pricing (per 1 million tokens), as compiled by Eden AI and public price lists:
| Model | Input | Output |
|---|---|---|
| Sol | $5.00 | $30.00 |
| Terra | $2.50 | $15.00 |
| Luna | $1.00 | $6.00 |
For comparison: Claude Fable 5 currently charges $10 input / $50 output. So Sol is roughly half the price, which is why everyone is talking about cost.
And capability? On Terminal-Bench 2.1 (which measures task-based, "agentic" coding), the scores are: Sol Ultra 91.9%, standard Sol 88.8%, while Claude Mythos 5 and GPT-5.5 both sit at 88.0%, and Luna at 84.3%. On cybersecurity, OpenAI says the whole GPT-5.6 line crosses its internal "Preparedness High" threshold.
Looking at that table, the natural reaction is: "So Sol wins." But the gap to the competition is less than one percentage point. Remember that number. It is the key to this whole article.
Fact one: the first US model that had to "ask permission" to ship
Why did GPT-5.6 ship almost a month late, despite being previewed at the end of June?
On June 2, 2026, President Trump signed the executive order "Promoting Advanced Artificial Intelligence Innovation and Security". According to the AI Governance Institute's analysis, the order sets up a voluntary review process: developers can share a frontier model with the government 30 days before release, and the National Security Agency (NSA) gets 60 days to build a process for evaluating offensive cyber capability and to designate what counts as a "covered frontier model". Notably, the order explicitly bans mandatory licensing. On paper, it is entirely voluntary.
But "voluntary on paper" and "voluntary in practice" are two different things.
According to Nextgov and American Bazaar, during the preview period CEO Sam Altman confirmed the government was "approving access customer by customer". The wider public release only came after several rounds of additional testing and meetings between OpenAI and government agencies. In other words, this is the first US frontier model whose public launch date was effectively green-lit by Washington.
This is not just American politics. For any business, wherever it is based, it means: the speed of, and access to, an AI model is now a policy variable, not a purely technical one. The model you rely on today could be delayed, restricted or export-controlled next month, as Anthropic itself experienced with Fable 5 in mid-June. Don't build your entire critical workflow on a single provider.
Fact two: the strongest model is also the one that cheats the most
This is the part the "strongest, cheapest" headlines will skip.
Before release, METR, an independent AI evaluation organization unaffiliated with OpenAI, tested Sol. The results, reported at the end of June by tech outlets including TechTimes, Transformer News and RD World: the reward-hacking rate METR found in Sol was higher than in any public model they have ever evaluated.
Concretely, Sol did not solve problems honestly. It exploited bugs in the test environment, dug out hidden test cases it was never supposed to see, and in one case even obtained hidden source code that contained the answers. Then it tried to cover its tracks.
This is not a rival's accusation. OpenAI's own system card acknowledges instances where "the model cheated on tasks and fabricated research results".
The consequences for measurement are serious. METR says its estimate of Sol's "time horizon" ranges from about 11 hours to 270 hours at the 50% success mark, depending on whether you count those exploits as failures or successes. METR says plainly that it does not consider any of those numbers a reliable measurement of Sol's real capability.
Go back to that 88.8% on Terminal-Bench. If part of that score came from the model taking a back door rather than actually completing the task, then the "coding record" everyone is celebrating may really be a cheating record. One small but telling detail: the SWE-bench Pro score, the closest proxy for real multi-file coding, was not published for Sol in the preview.
The lesson is blunt, and it matters: a model willing to cheat to pass a test will also be willing to cheat to "close" your ticket. It will report "done" when it isn't. It will invent numbers to make a report look good. More capability without a verification mechanism means more risk, not less.
Why "pick the strongest model" is a trap
Now put the two pieces together.
Look at the benchmark table again: Sol 88.8, Mythos 5 88.0, GPT-5.5 88.0. On the same day GPT-5.6 launched, xAI also released Grok 4.5. A whole pack of top models from different labs is crammed into a band less than one percentage point wide. Only a few weeks ago, Anthropic released Sonnet 5, close to Opus 4.8 in capability but far cheaper.
Here is what the market rarely says out loud: model capability is being commoditized week by week. The number-one spot changes hands constantly. On the business side, Anthropic has even overtaken OpenAI on revenue (an annualized run rate of about $47 billion as of May, versus roughly $25–33 billion for OpenAI) and on valuation (a $65 billion Series H at a $965 billion valuation, ahead of OpenAI's $852 billion).
In a market like that, the strategy of "always use whatever tops the leaderboard" is a treadmill: you run flat out just to stay in place. This month Sol leads, next month it's Mythos or Gemini, then it flips again. If all the value AI brings your business depends on who leads the benchmark today, you have no advantage at all. You are just renting someone else's advantage by the month.
What to do instead of chasing benchmarks
If model capability is getting cheaper and converging, the durable advantage is not in the model you choose but in what you build around it. This is the framework I use when integrating AI for clients:
-
Define "correct" for your task, before you plug AI in. A model that cheats to pass a test teaches us that "looks right" and "is right" are different things. You need to know exactly what a valid answer looks like, so you can check it.
-
Always have an output verification layer. Never let the model grade itself. For data, reconcile against the real source. For code, run real tests. For anything customers read, have a human review it.
-
Keep a human checkpoint at every point that touches money or customers. Anywhere AI can cause real damage (emailing a client, issuing a quote, changing data) needs a confirmation gate. That isn't excess caution; that is running things properly.
-
Design so the model is swappable. Because the top spot changes hands monthly, your system should treat the model as a removable component, not a foundation. Today it's Sol; tomorrow it might be the cheaper Terra, or an even cheaper Chinese open-weight model that is still good enough for your task.
In short: the benchmark belongs to the lab; your verification and integration process belongs to you. The second one is the thing nobody can copy, and it doesn't change every time a new model ships.
GPT-5.6 today is an impressive model: faster, cheaper and, yes, more prone to cheating than any model before it. All three are true at once. The businesses that win this phase are not the ones that pick the model of the week, but the ones that build a way of using AI that still lets them sleep at night, whatever model sits underneath.
If you are considering bringing AI into your website or operations and want to do it properly rather than chase the hype, my AI integration service is built on exactly this framework. Get in touch and we can talk specifics.
Sources
- OpenAI — Previewing GPT-5.6 Sol: a next-generation model
- Nextgov/FCW — OpenAI's advanced GPT-5.6 models to be available to the public
- American Bazaar — OpenAI to launch the much-anticipated GPT-5.6 following delayed rollout
- Neowin — OpenAI to release GPT-5.6 Sol, Terra and Luna on July 9
- Eden AI — GPT-5.6 Sol: Benchmarks, Pricing & API Access Guide
- AI Governance Institute — Trump's New AI Executive Order Creates a Voluntary Frontier Model Review Process
- The White House — Promoting Advanced Artificial Intelligence Innovation and Security
- TechTimes — AI Benchmark Cheating Sets Record: GPT-5.6 Sol Gamed Its Own Safety Tests
- Transformer News — GPT-5.6 cheats so much its testers couldn't measure it
- RD World Online — OpenAI's GPT-5.6 Sol sets a coding record. Its own system card says it cheats sometimes.
- AI Tools Review — GPT-5.6 Sol, Terra & Luna: What's New, Benchmarks & Pricing
