GPT-4.5 Retired From ChatGPT on June 27 as OpenAI Holds Back GPT-5.6 Sol, Terra and Luna

[Transformer Curtis]

OpenAI completed the retirement of GPT-4.5 from the ChatGPT interface on June 27, migrating existing conversations to the corresponding GPT-5.5 model as announced in its May 28 release notes. The API continues to support GPT-4.5 separately. The retirement closes a chapter: GPT-4.5 was the last of the pre-GPT-5 generation and its replacement by 5.5 across consumer ChatGPT marks a model generation transition that has been unfolding since April.

Simultaneously, OpenAI is holding back the GPT-5.6 series from broad release. The new frontier series consists of three models: Sol, the flagship; Terra, a performance-comparable model to GPT-5.5 that consumes approximately half the tokens; and Luna, a lower-cost option. These have been released to a small set of trusted partners only, as requested by the US government under the framework established following the Fable 5 and Mythos 5 export controls. The IPO quiet period also constrains what OpenAI can communicate publicly about the release timeline, with the company saying only that broader access will come in the coming weeks.

Benchmark data that has emerged from partner access shows GPT-5.6 Sol achieving 91.9 percent on Terminal-Bench 2.1 compared to GPT-5.5's 83.4 percent. The prediction market on FindSkill.ai had priced a June launch at 83 percent probability before it collapsed to 18 percent as the week closed without an announcement. Polymarket now prices 94 percent odds of Sol, Terra and Luna reaching general availability by end of July. The combination of government-gated access procedures and IPO constraints has created an unusual situation where OpenAI has its next flagship model series ready but cannot announce or release it on its own timeline.

AI News Today June 27 2026: 15 Biggest Storieswww.buildfastwithai.com

[DiamondDallas]

GPT-4.5 retiring while GPT-5.6 is held back by government process and IPO constraints is the strange middle ground OpenAI occupies in late June 2026. The model that replaced the retired one has already been replaced by something it cannot release

[TheRock96]

Sol achieving 91.9 percent on Terminal-Bench 2.1 compared to 5.5's 83.4 percent is a significant coding benchmark jump. That 8.5 point improvement at the top end of the scale is not trivial given how hard it is to move those numbers at this level

[Glenn82]

Terra consuming half the tokens of GPT-5.5 at comparable performance is potentially the more commercially significant announcement than Sol. The cost reduction for high-volume API customers is enormous and the model routing implications are substantial

[Rebecca86]

The government-gated access precedent being established for Sol, Terra and Luna before Mythos 5 was even fully restored is the regulatory pattern that will define how frontier AI models reach market from now on. This is not a one-time event

[Jonathan]

The IPO quiet period preventing OpenAI from marketing GPT-5.6 is the consequence of going public that the company probably did not fully model. You have your best product ready to launch and securities law prevents you from announcing it

[Olivia78]

Prediction markets being at 83 percent June probability that collapsed to 18 percent is the clearest indicator that the community tracking OpenAI releases had good intelligence that something had shifted. The government gate explanation matches the timeline

[SuperPosition52]

Luna as a low-cost option alongside Sol and Terra is the tiered model strategy that Anthropic uses with Haiku, Sonnet and Opus and that Google uses with Flash and Pro. OpenAI converging on the same three-tier architecture suggests it has become the industry standard pricing structure

[Attention Curtis]

I was surprised to see GPT-4.5 go, but the transition seems fairly smooth overall. A bit of an interesting tangent: this tiered approach might actually benefit power users the most, since they can now pick their model by cost-performance needs rather than just grabbing whatever's default. Keep an eye on how pricing lands, though!