Model facts

Every row that changed here changed because its vendor stopped listing the old model as current, and every model that left is still on sale. None of their figures had moved. That matters if you are calling one of them today: the old row's numbers still describe what you are paying.

Claude Opus 5.5 took Opus 5's place. Anthropic's models overview now puts Opus 5.5 in its headline table and lists Opus 5 under “Legacy models (still available)”. Its model deprecations page has Opus 5 as active, with retirement not sooner than 24 July 2027, and its pricing page still charges the $5 and $25 this table gave for it. It is the same move as Fable 5 to Fable 5.1 on 1 September, with one difference: the newer model is cheaper. Opus 5.5 is $4 and $20 with the same limits, and both its cutoffs are June 2026 against Opus 5's May.

The OpenAI rows are GPT-6 Astra, Sol and Luna, and the name Sol now means a different tier. OpenAI's model catalogue lists those three as its flagship models, and GPT-5.6 Sol and Terra are gone from it and from the standard table on the pricing page. GPT-5.6 Sol was, on its own model page, a flagship model at $4 and $20. GPT-6 Sol is the one the catalogue says to choose “to balance intelligence and cost”, which is how Terra's page describes Terra, and it is $2 and $10. The top of the lineup is GPT-6 Astra. Moving from gpt-5.6-sol to gpt-6-sol because the names match is a move to the middle of the new lineup, not to its top. GPT-6 Luna, the third, is the one OpenAI points at “cost-sensitive, high-volume workloads”.

Neither GPT-5.6 model has a shutdown date. Both model pages are still up with no deprecation notice, and OpenAI's deprecations page names them as the replacements for older models rather than scheduling their own removal. Sol's page still gives $4 and $20, with the promotional wording this table used to print, and Terra's still gives $2 and $12.

GPT-6 Astra, and the move from Gemini 3.6 Flash to 3.8 Flash, were prepared on 15 September and are part of the same change. The Gemini row moved without one figure changing — the same 1,048,576 and 65,536, the same $0.75 and $3.75, the same expiry in the same words, still no published cutoff — because Google's pricing page calls 3.6 Flash “our previous generation Flash model”, and this table says it lists current ones. A row can go stale on that claim alone while every number in it stays right. The watcher sees that only when the old model vanishes from the page it reads, which is how the three rows above were caught. When a vendor merely relabels a model, as Google did with 3.6 Flash, only a person reading the page will notice.

Sol was listed at $5 input / $30 output. OpenAI's pricing page and its models page both now give $4 and $20, and the long-context tier has moved with it, from $10 / $45 to $8 / $30. Those figures were correct when this table was last verified on 14 August, when the log records OpenAI as re-read and unchanged, so the most likely account is a cut somewhere in the twelve days between. This page cannot prove that from the outside: a vendor's current page shows today's price and not the date it changed, so a cut on the 20th and a misreading on the 14th look identical from here. Either way, anyone who budgeted a Sol job from this table in that window over-estimated by a quarter on input and half on output. Erring expensive is the better direction to be wrong in, and it is still being wrong. GPT-5.6 Terra was checked in the same pass and has not moved.

This table said $1.50 input / $7.50 output. Google's pricing page gives $0.75 and $3.75 “through December 31, 2026”, with $1.50 and $7.50 “starting January 1, 2027”. The figures printed here were the future price, so anyone budgeting a Gemini Flash job from this table between 11 and 14 August would have doubled their estimate. The cells now carry the current rate, and the date the rate changes is stated rather than left as a footnote nobody reads: these Gemini figures expire on 1 January 2027. Google also publishes batch and flex rates at half the standard price, and a priority rate well above it.

Anthropic's models overview and pricing pages, OpenAI's pricing and models pages, Google's Gemini API pricing and its Gemini 3.8 Flash model page, and Meta's Model API pricing. For the three rows that left, also Anthropic's model deprecations page, OpenAI's deprecations page, and OpenAI's own pages for GPT-5.6 Sol and Terra.

Read Anthropic's overview page as markdown, not as a web page. Its rendered table shows the reliable cutoff and folds the training cutoff away behind a “Show all details” control, so a reader — or a tool — that takes the visible table at face value will conclude the training cutoff is not published at all. It is: appending .md to the URL returns the full table with both rows. That is the read this page is checked against, and on 15 September 2026 the difference between the two views nearly put a wrong figure on this table.

Each vendor sells cheaper small models and more expensive reasoning tiers; this table exists to make the shape of the market legible, not to be exhaustive.

Anthropic moved Claude Fable 5 off its headline model table and put Claude Fable 5.1 in its place, so this table followed. Fable 5 was not retired and is not going away: Anthropic lists it as active with a retirement date no sooner than 9 June 2027, and its own model page still carries exactly the figures printed here for it until today. It has become a legacy model, which is the same status as the six older Claude models this table has never listed, and listing one of the seven would be an odd place to stop. Nothing about the money changed — Fable 5.1 is the same $10 and $50 — and both knowledge cutoffs moved forward, from January to June 2026. If you are calling claude-fable-5 today it still works, at the price this table gave before.

The string in that column is the one the vendor documents as the API ID. For Claude Haiku 4.5 that is the dated claude-haiku-4-5-20251001; the shorter claude-haiku-4-5 is documented as an alias and also works. From the 4.6 generation onward the two are identical, so Haiku is the only Anthropic row where the distinction shows.

OpenAI documents no alias for the three GPT-6 rows: each model page lists one ID and nothing else. So Haiku is now the only row on this page where the distinction shows. That was not true while this table carried GPT-5.6 Sol, whose model page states that “the gpt-5.6 alias routes requests to GPT-5.6 Sol”. This page missed that alias until 15 September 2026 and said Haiku was the only such row. The sentence is true again now, but only because the Sol row left. Claims like this go wrong, and right again, without any figure moving, and the daily watcher cannot catch them, because it compares numbers and this is prose.

Google publishes an exact input limit of 1,048,576 tokens and an output limit of 65,536. OpenAI's model pages give 1,050,000 and 128,000, which its catalogue writes as 1.05M and 128K. This table prints both context limits as 1.05M. That is fair as rounding, but they are not the same number: Google's is a power of two, and OpenAI's is a round decimal figure. Until 23 September 2026 this note said they were the same number written two ways. Nothing you could budget from depends on the difference, but the sentence was wrong, and it sat directly under the figures it was explaining.

OpenAI publishes a knowledge cutoff of 30 April 2026 for GPT-6 Astra, 20 April 2026 for GPT-6 Sol and 18 May 2026 for GPT-6 Luna. Google's model page for Gemini 3.8 Flash does not publish one, which is why there is no cutoff column here.