Gemini 3.5 Pro, Qwen3.8-27B, Grok 4.7: planning around models that didn't ship

Today is September 12, the date Elon Musk gave on X on September 2 for Grok 4.7. There is no model ID, no pricing, no model card. That makes three announced models this summer I would have had to explain to a client if I had planned around them: Google's Gemini 3.5 Pro, Alibaba's open-weight Qwen3.8-27B, and now Grok 4.7. None of the three is a scandal. All three teach the same lesson, and it is about your architecture, not the vendors.

Three slips, three failure modes

Gemini 3.5 Pro was introduced at I/O 2026 and expected in June. In July, Google pushed it by a month; internal tests reportedly showed the flagship lagging in coding and long-horizon reasoning, and the model stayed in a limited enterprise preview. This is the classic slip: the vendor is open about it, the capability is real, the date is not.

Qwen3.8-27B is the more instructive case. Alibaba promised the open weights for the week of August 10 on Hugging Face and ModelScope. As of August 10 there were no weights, no license, and no repo. What did exist were community quantizations, FP8 and NVFP4, published ahead of a release that had not happened. Qwen3.8 Max had shipped on August 2 as an API-tier model, so the family was real; the open piece was not. A Qwen3.8-Flash-Next variant later showed up on Hugging Face and in Ollama's MLX builds, which is a different model and a different promise.

Grok 4.7 is the third mode: a 2.1T-parameter model, reportedly about 40% larger than 4.6 and trained partly on internal SpaceX data, with a launch date inferred from a social post rather than a dated commitment. xAI has slipped before. On September 11 there was still no model ID; on September 12 the date passed without a release.

A model that exists only as an announcement is not a dependency. It is a rumor with a parameter count.

Why this hurts more in 2026 than it used to

Model IDs are load-bearing config now, and two DeepSeek examples from the same summer show how fragile that is. On July 24, the legacy names deepseek-chat and deepseek-reasoner were retired at 15:59 UTC; anything still calling them failed at that minute. And from September 14 at noon Beijing time, the deepseek-v4-pro endpoint serves V4.1 Flash at Flash prices until V4.1 Pro ships. Same ID, different model. If a stable ID can change meaning under you, an unreleased one has no business in a roadmap.

What I actually do about it

• Adapter layer, always. Every model call goes through one interface that maps a task tier to a concrete model ID and a fallback. Swapping Gemini 3.5 Pro for whatever ships in its place should be a config change, not a refactor. The tier design in frontier-model-router-2026 and the plumbing in openrouter-cost-routing are exactly this.
• Roadmap on released weights only. If a local-inference plan needs Qwen3.8-27B, the plan is blocked until the weights and the license exist. Plan on what you can download today and treat the newer model as an upgrade path.
• Pin model IDs and read deprecation calendars. The DeepSeek retirement was announced ahead. The failures were in teams that did not read the notice.
• Write down what a slip costs. If a client feature depends on a model that is not GA, the date risk goes in the plan, with an owner.

How to read community quants with no license

A community FP8 or NVFP4 checkpoint of an unreleased model is useful for one thing: checking whether your hardware and serving stack can run that shape at all. It is not something to ship. Without an official license you have no usage terms, no attribution rules, and no answer when a customer's legal team asks where the weights came from. The checkpoint may also differ from the eventual release, so any eval numbers you collect describe a different artifact.

The honest gap

I do not know why any of these three slipped. The coding and reasoning explanation for Gemini 3.5 Pro is reporting, not a Google statement, and the Grok 4.7 parameter count comes from a post, not a card. Some of these models may ship next week and be excellent. That would not change the point: the cost of planning around them was paid by anyone who did, whether or not the model arrived.