OpenAI Just Killed GPT-5.5. The GPT-5.6 Tier Map Is One Model For Paid, One For Free, 68% Fewer Facts Wrong.
OpenAI Just Killed GPT-5.5. The GPT-5.6 Tier Map Is One Model For Paid, One For Free, 68% Fewer Facts Wrong.
Hey guys, Mr. Technology here.
OpenAI shipped two things overnight and one of them is the actual story. GPT-5.6 Sol powers Plus and Pro now — Instant and deep reasoning both — same model. GPT-5.6 Luna hits Free and Go users tomorrow with unlimited text chats. The 68% factuality improvement is the headline OpenAI is selling. The unification is the architecture.
What Actually Shipped
Two models, not one. The OpenAI account put it bluntly: "GPT-5.6 Sol now powers both Instant and deep reasoning for Plus & Pro users." That is one model doing both jobs. The Instant you have been using for casual prompts and the deep-reasoning endpoint you switched to for hard problems are now the same weights, routed by the same router, paying the same bill. The Plus/Pro tier map just collapsed from "two endpoints, you pick" to "one endpoint, the router picks for you."
GPT-5.6 Luna is the free counterpart. Unlimited text, ships tomorrow, free for Free and Go users. The differentiation OpenAI is selling is no longer "which model." It is "which router, which latency class, which feature set."
The 68% Number Is The Headline
OpenAI's second tweet made a specific claim: in their high-stakes factuality evaluation covering finance, medicine, and law, GPT-5.6 Sol produced 68% fewer responses with factual errors than GPT-5.5 Instant.
That number is doing a lot of work. Three things to note before you change a production contract.
1. It is a relative reduction, not an absolute floor. A 68% drop on a baseline that is already low — say, an 8% factual-error rate on the eval — puts Sol at roughly 2.6%. A 68% drop on a baseline of 30% (which is plausible for high-stakes domains) puts Sol at roughly 9.6%. The post-training improvements are real either way, but the absolute error rate on finance/medicine/law tasks is still the number that matters when you are advising a client on a regulatory filing. 2. It is against GPT-5.5 Instant, not GPT-5.5 deep reasoning. The comparison is the easy one. The interesting comparison — Sol versus the deep-reasoning endpoint of 5.5 — is not in the tweet. If Sol is faster and matches 5.5 deep reasoning on factuality, that is a real win. If Sol is faster and slightly less factual than 5.5 deep reasoning on the hardest prompts, the tier collapse has a hidden trade. 3. It is OpenAI's evaluation. High-stakes factuality evals are notoriously domain-sensitive. The right move is the same as it has been since GPT-4: run your own eval on your own workload before you change a production contract.
What This Is, Strategically
OpenAI has been signaling this for two months. The Instant / deep-reasoning split was always a router decision wearing a UX costume. Now the costume is off. Sol is the model. The router decides how long to think. That is the Anthropic effort-dial architecture applied to OpenAI's tier map, and it gives OpenAI three things at once.
- Lower inference cost. One model, one set of weights, one set of serving infrastructure. The router runs the same model at different thinking depths, which means you do not pay for two separate deployments.
- A cleaner tier story for buyers. "Plus gets Sol, Pro gets Sol with longer thinking, Free gets Luna" is a sentence a CFO can hold in their head. "Plus gets 5.5 Instant for casual, plus gets 5.5 deep reasoning for hard prompts, plus gets 5.5 pro for the hardest prompts" was three sentences and a calculator.
- A factuality moat they can repeat. If the 68% claim holds up under independent benchmarks, this becomes the headline OpenAI repeats for the next six months. The next earnings call. The next enterprise pitch. The next "is GPT better than Claude at X" Reddit thread.
What You Do This Week
If you are running GPT-5.5 Instant in production for any high-stakes workflow — contract review, clinical triage, regulatory work, anything where "the model was wrong" is a legal or financial outcome — this is the moment to re-run your eval on Sol. The router behavior is changing too: prompts that used to silently route to deep reasoning may now route differently. The cost per task could move in either direction depending on how Sol's router weights thinking budget versus Instant's.
If you are on Free or Go, your world changes tomorrow. Luna is unlimited text. That is the most generous free tier OpenAI has shipped, and it puts pressure on every other "free with rate limits" story in the market — Claude, Gemini, Grok. Whether it stays unlimited once the inference bill comes due is a different question, but for now it is the floor.
If you are on Plus or Pro, the user-visible story is "the model is smarter and more factual." The architect-visible story is "the tier map is one model now." Both are true.
GPT-5.6 Sol and Luna are live for paid users today, Luna for free tomorrow. The router is the product. The fact is the marketing.
Sources: