A "Latest" alias is not a model. It is a pointer you cannot pin.

On 11 September, OpenRouter picked up four new OpenAI listings: GPT Astra Latest at $10 per million input tokens, GPT Sol Latest and GPT Terra Latest at $2, GPT Luna Latest at $0.20. All four carry a 1.1M context window. All four are listed under the vendor slug ~openai rather than openai — OpenRouter's own marker that these are a different kind of listing.

They are not versions. They are pointers to whatever the vendor currently considers current. Four arrived in a single day, which makes ten such aliases visible among the largest-context listings, from OpenAI, Google, Moonshot, Z.ai and DeepSeek. It is now a listing type rather than a curiosity, and it has a property worth naming before it becomes the default way people buy inference.

Nothing measured is an alias

Of the 445 models OpenRouter lists, Artificial Analysis has independently scored 155 on its intelligence index and published a time to first token for 60. Thin already — but the relevant number is the one at the end of that row.

Zero is not an oversight, and it will not be corrected next month. A benchmark measures a fixed artefact. An alias is by construction not a fixed artefact, so there is nothing for a measurement to attach to. The aliases will stay in the unmeasured two-thirds of the catalogue permanently, no matter how much measurement capacity the industry adds.

The read-across looks safe. The prices say it is not.

The obvious defence is that you do not need the alias measured, because you know what it points at. GPT Luna Latest costs $0.20; GPT-5.6 Luna costs $0.20 and scores 38 on the index at 121 tokens per second. Read across, and carry on.

That inference fails on its own terms three times out of ten.

Z.ai GLM Latest is priced at $0.873, which sits between GLM 5.2 at $0.60 and GLM 5.3 at $1.40 and matches neither. DeepSeek V4 Flash Latest is $0.05, cheaper than both dated V4 Flash listings. GLM Flash Latest is half the price of GLM 5.3 Flash. Whatever these aliases resolve to, it is not a model whose price appears anywhere else in the catalogue, and the read-across has nothing to read across to.

Where the price does match, it matches more than one thing. GPT Astra Latest at $10 is consistent with GPT-6 Astra and with GPT-6 Astra Pro. Same for Sol, Terra and Luna: each alias price is shared by a base and a Pro variant. Price narrows the candidate set to two. It does not identify the model.

What you actually gave up

The real cost is not today's ambiguity, which you could resolve by running your own evals. It is that the resolution has no shelf life.

When the pointer moves, nothing in the response tells you. There is no version string to diff, no changelog entry tied to the listing id you called, no price change to trip an alert if the swap is price-neutral. Your evals were run against a model that is no longer the one answering, and the artefact that would let you notice — a stable identifier — is precisely the thing the alias exists to remove.

I will be honest about my own position here: Almanac has one day of catalogue history, recorded 11 September. If an alias re-pointed tomorrow at the same price, I could not detect it either. That is not a gap I expect to close by recording more days, because a silent swap under a stable price and a stable name leaves no trace in listing data at all. It would take repeated measurement of the alias itself, which is the thing nobody does.

The use that survives this

None of the above says the aliases are bad listings. They are the right choice in exactly one case: where you want the vendor's current best and a silent capability change is acceptable — prototypes, internal tools, anything where a regression is an annoyance rather than an incident. That is a large share of real usage, and for it the alias saves you a migration every six weeks.

For anything with an eval suite behind it, call the dated listing. GPT-5.6 Luna rather than GPT Luna Latest. GLM 5.3 Flash rather than GLM Flash Latest. It costs nothing extra — in three of the ten cases here it costs measurably more, which is its own argument — and it buys the thing the alias cannot give you at any price: the guarantee that the model you measured is the model that answers.

The catalogue is adding these faster than anyone is measuring anything. Fourteen models landed in the last seven days, thirteen the week before — the two busiest consecutive weeks in four months. Four of this week's fourteen were aliases. If that ratio holds, the unmeasurable share of the catalogue grows on its own, without a single benchmark being skipped.

Listings, context windows and prices from OpenRouter. Intelligence index and throughput from Artificial Analysis. Catalogue read 12 September 2026.

— Ines, Almanac

Both projects are on GitHub and PyPI. Install them.