An hour apart on the same day, and 218 days
Anthropic and OpenAI shipped frontier models an hour apart on 22 September 2026; Google has not shipped one in 218 days.
Two labs now set their release dates by watching each other: on 22 September 2026 Anthropic shipped Claude Opus 5.5 and OpenAI followed with GPT-6 Sol and Luna about an hour later. With capability too close to separate — Artificial Analysis ranked Opus 5.5 first of 210 models the same week METR called it a modest improvement, and a replay on a production TypeScript monorepo had Opus finish 13 of 20 runs cleanly at $45.81 against Sol's 8 of 20 at $13.16 — the fight moved to price: Opus tokens down 20% to $4/$20 against Sol's $2/$10. Google is not in that exchange. It has shipped no Pro-tier flagship since Gemini 3.1 Pro on 19 February 2026, 218 days, while shipping three Flash models in 43 days.
On 22 September 2026, Anthropic and OpenAI shipped new models an hour apart. The competition showed up in price first, and Google has not shipped a model at this tier for 218 days.
An hour apart
fact Anthropic shipped Claude Opus 5.5 first, and OpenAI followed about an hour later with GPT-6 Sol and Luna (Simon Willison, checked 25 September 2026. The Hacker News posting times are in the same order).
fact The two announcements took the Hacker News front page at the same time, and the size of the reaction was almost the same (Opus 5.5 thread, GPT-6 thread, checked 25 September).
The gap between the announcements was not a day but an hour. There is no reading of that other than each setting its date by watching the other.
The competition showed up in the price
fact Per million tokens, Opus 5.5 is $4 in · $20 out · $0.20 cache read. The previous Opus 5 was $5 · $25 · $0.50 — tokens down 20%, cache reads down 60% (Anthropic pricing docs, checked 25 September).
fact The same day, OpenAI put out Sol at $2 · $10 and Luna at $0.10 · $0.50 (OpenAI API docs, checked 25 September). At those prices, the one going up against Opus 5.5 is Sol.
That said, the numbers in either announcement are hard to take at face value. Anthropic's announcement says it "runs 40% cheaper", but the cut you can check on the price list is 20%. The 40% is a figure the company worked out on its own assumptions about the work. OpenAI's "50% cut" carries the condition that it is against GPT-5.6's promotional price.
fact On top of that, Opus 5.5 cannot have thinking turned off, and thinking tokens are billed as output. Anthropic's own docs note that work you used to run with thinking off may produce more output tokens per request (Anthropic migration docs, checked 25 September). Cut the list price while making the model talk more, and it does not come down by as much as the cut.
Performance comes out differently depending on who measures
fact Artificial Analysis put Opus 5.5 (max) first of 210 models. Intelligence index 58, $5.98 on average per task (Artificial Analysis, checked 25 September).
fact The same week, METR ran a pre-deployment evaluation over ten days on five long-horizon tasks and wrote that it is "not a large leap, but most likely a modest improvement" (METR, checked 25 September. The same page states the work was done under an unpaid agreement).
fact A head-to-head on real code gives yet another picture. On a TypeScript monorepo of 3,700 files, 6 already-merged changes were handed back to each model and run 20 times per model: Opus 5.5 passed cleanly 13 times with 0 test regressions at $45.81, Sol passed 8 times with 5 regressions at $13.16 (paddo.dev, checked 25 September).
Three results in the same week. "First place", "a modest improvement", "three times the price and more accurate" — none of them is wrong. They only measure differently. When the leaders are this close together, one line of a leaderboard does not settle which is better. Anthropic wrote in its announcement that at this level it is hard to judge the real difference from a gap in benchmark scores.
Google has been quiet for 218 days
fact Google's last flagship Pro model is Gemini 3.1 Pro, on 19 February 2026. That is 218 days ago as of 25 September, and the gap before that flagship was 93 days — about 2.3 times as long, and not over yet.
fact Gemini 3.5 Pro was announced at I/O on 19 May for a June release, then pushed back. According to Bloomberg's reporting, the training data was changed at the end of June to raise coding ability, and the result was disappointing (9to5Google, checked 25 September). Today the DeepMind model page still has 3.1 Pro in the Pro slot and lists 3.5 Pro as "coming soon" (Google DeepMind, checked 25 September).
While Anthropic and OpenAI went at each other an hour apart, Google has not shipped a model in the same weight class for over seven months.
Fighting with Flash instead
fact Google kept shipping through the same period. Only, all of it is Flash — 3.6 Flash on 21 July, 3.7 Flash on 13 August, 3.8 Flash on 2 September: three in 43 days. On 22 September, the day the competitors' announcements poured out, it also took 3.8 Flash TTS to general availability (Gemini API changelog, checked 25 September).
fact That Flash is cheap. 3.8 Flash started at $0.75 in · $3.75 out per million tokens (Google blog, checked 25 September). In a market that sells quality you can run in bulk, that price is itself the strategy.
A company that ships three in 43 days has put nothing in the Pro slot for 218 days. Being able to ship fast and being able to ship into that slot are not the same ability.
At the frontier, two companies are the ones setting the price and setting the release dates right now. The day Google puts a model in the same weight class, that arithmetic changes. Seven months in, not yet.