20 November 2026 · AI Model Pricing · 6 min read
Three AI models shipped between November 9th and 17th. Our calculator didn't carry any of them until today — a 3-to-11-day gap between a vendor's announcement and the day you could actually run the numbers here. While backfilling them, I checked our own dedicated Claude pricing page against the same model list the calculator reads from, and found something worse: it was quoting two models that don't exist in the current lineup anymore, both at prices nobody can actually pay anymore.
| Modello | Fornitore | Announced | Added here | Lag | New price (in/out per 1M) |
|---|---|---|---|---|---|
| Muse Ember | Meta | Nov 9 | Nov 20 | 11 days | $0.10 / $0.50 |
| Gemini 4 Hydrogen | Nov 12 | Nov 20 | 8 days | $1.50 / $7.50 | |
| GPT-6.2 Astra / Sol | OpenAI | Nov 17 | Nov 20 | 3 days | $8.00/$40.00 · $2.00/$10.00 |
Muse Ember is the cheapest model Meta has priced to date — $0.10 input / $0.50 output per million tokens, half of Llama 3.1 70B's own base-tier rate ($0.20/$0.60). It's Meta's entry into the classification/routing/extraction tier every other major lab already has a dedicated budget model for.
Gemini 4 Hydrogen is the cheap, fast sibling to Gemini 4 Argon, priced at exactly 75% of Argon's rate on both sides: $1.50 in versus Argon's $2.00, $7.50 out versus Argon's $10.00. Run 10M input and 1M output tokens a month on Argon and you're at $30.00; the same workload on Hydrogen comes to $22.50 — a clean quarter off, no extra markup hidden on the output side.
GPT-6.2 Astra replaces GPT-6 Astra at $8.00/$40.00, a flat 20% cut from the $10.00/$50.00 rate it launched at back in September. On that same 10M/1M-token month, that's $120.00 instead of $150.00. GPT-6.2 Sol, announced the same day, didn't move at all — still $2.00/$10.00, identical to GPT-6.1 Sol before it.
While wiring these three into the model list, I pulled up our own Claude pricing page to sanity-check it against the same registry. It still listed Claude Fable 5 at $10.00/$50.00 E Claude Opus 5 at $5.00/$25.00. Neither model is the current one. Opus 5.5 replaced Opus 5 a while back at $4.00/$20.00 — a 20% cut we actually covered in detail when it happened — and Fable 5.2 replaced Fable 5 the same way, landing at $8.00/$40.00. The live calculator and the model comparison tool both already had the right numbers. The dedicated vendor page, the one that's actually indexed for "Claude API pricing," had two retired models, each overstating the real cost by exactly 25% ($5 vs $4, $10 vs $8 — same ratio twice), with a "last verified" date sitting right next to the wrong numbers that said August.
Fixed it today — the English page and all seven translated copies now show Opus 5.5 and Fable 5.2 at their correct rates.
It isn't an isolated slip either. A quick check of our OpenAI page while I was in there turned up the identical pattern: it still shows GPT-6 Astra at the pre-cut $10.00/$50.00, with no row at all for GPT-6.2 Astra or Sol. Same 25% overstatement, different vendor's page. That one's still open — I'm flagging it here rather than quietly leaving it, because the honest state is "found, not yet fixed."
Both bugs share one cause: the vendor pages — /anthropic.html, /openai.html, and the others like them — are hand-written HTML, built separately from models.js, the file our calculator and the model comparison tool actually read from. Nothing cross-checks the two against each other. A 3-to-11-day lag on a brand-new model launch is one I can live with; nobody prices a model the day it ships. A hand-maintained page quietly drifting for months after the number it shows stops being true anywhere on earth is the one worth catching — and right now the only way I catch it is by tripping over it while doing something else.
Run the current numbers on the AI model price comparison, price a real workload on the Calcolatore dei costi dell'API AI, or see every change we've tracked on the LLM price history page.
Model prices sourced from our own pricing registry (build_models.py), cross-checked against each model's announcement date, 2026-11-20. Anthropic page fixed the same day across all eight language versions; the matching OpenAI page issue is confirmed but not yet corrected. Prices change — confirm current rates on the provider's own pricing page before budgeting.