TL;DR
- On 22 September 2026 Anthropic released Claude Opus 5.5, which it says costs 40% less to run than Opus 5, and OpenAI released GPT-6 Sol and Luna at 50% lower API prices than GPT-5.6 promotional pricing.
- At those prices an English chat reply costs a fraction of a US cent. Which model you choose changes the bill far more than which language you write in, unless Patois turns out to need many times more tokens, which nobody has published.
- The cost argument for a Caribbean language model has weakened. The arguments from accuracy, prompt custody and continuity still stand, and they need measurements the region has not yet made.
- This week, count tokens for 20 real Patois messages against English versions of the same messages and report the ratio by Friday 16 October.
What happened on 22 September
On Tuesday 22 September 2026 Anthropic introduced Claude Opus 5.5, which it says "costs 40% less to run than Opus 5." Input and output tokens are priced at $4 and $20 per million, down from $5 and $25. Cache reads fall from $0.50 to $0.20 per million. Anthropic explains the 40% as two effects stacked: a lower price per token, and fewer tokens used per task.
Within hours OpenAI released GPT-6 Sol and GPT-6 Luna. Its developer announcement promises 50% lower API prices than GPT-5.6 promotional pricing. VentureBeat reporting carried by Yahoo Finance gives the figures: Sol now costs $2 per million input tokens and $10 per million output tokens, down from $4 and $20, and Luna costs $0.10 and $0.50, down from $0.20 and $1.20. An OpenAI spokesperson told VentureBeat the new rates carry no expiry date.
Anthropic's page calls Opus 5.5 its "first release since we called for pacing the frontier", so the price cut arrives with a safety argument attached. Neither launch page reports a result in a Caribbean language. The Opus 5.5 page shows coding, knowledge-work and computer-use benchmarks, and we found no language-specific claim on it.
The numbers behind a Patois message
Start with a typical English exchange, our own illustration: 200 tokens in and 300 tokens out. On the published Luna prices that costs $0.00017, or about 0.017 US cents. On Opus 5.5 at $4 and $20 the same exchange costs $0.0068, or 0.68 cents. Opus 5.5 is 40 times dearer than Luna on output tokens, and that gap dwarfs most other choices a buyer makes.
Now suppose Patois needed more tokens for the same meaning. Petrov, La Malfa, Torr and Bibi found that the same text translated into different languages can differ in tokenization length by up to 15 times, with differences of more than 4 times even for character-level and byte-level models. Their paper does not test Patois. If we borrow 15 times as an outer bound, the exchange costs 0.26 cents on Luna and 10.2 cents on Opus 5.5. A thousand such messages cost $2.55 on Luna and $102 on Opus 5.5, against $0.17 and $6.80 in English. Longer prompts also eat more of the context window, which no price cut refunds.
Supply is the other half. On 5 October 2026 we searched the Hugging Face Hub, the public library where open models and datasets are shared, using its public API (/api/models?search=TERM&limit=100 and /api/datasets?search=TERM&limit=100) with each term below. The counts are what those queries returned that day, and anyone can rerun them.
| Hub search term | Models | Datasets | What the hits were |
|---|---|---|---|
| haitian creole | 8 | 13 | Speech, text-to-speech and translation sets, three Alpaca-derived instruction sets, one word-frequency list |
| guyanese creole | 4 | 0 | Three speech-synthesis models and one fine-tune of a 4-billion-parameter Gemma 3 |
| jamaican | 3 | 2 | JamaicanTinyStories (a small text generator trained on Patois stories, 7 downloads) and its 409,986-row text dataset; jamaican-patwa-whisper (34 audio clips); jamaican-jewel (an image LoRA unrelated to language); dream-interpreter-v1 (an empty repository) |
| patois | 7 | 33 | Five speech models, two audio sets and a Jamaican news-reading set (see below), a restricted proverbs set, two unrelated software models, and 29 French regional dialect collections |
| bajan | 4 | 0 | Two Whisper large-v3 speech models created 11 June 2026 with no model card; two unrelated hits |
| papiamento | 1 | 0 | One Gemma 2 text model with 14 downloads |
| jamaican creole | 0 | 0 | Nothing |
Terms that returned no relevant item: "barbadian", "trinidadian" and "kweyol" returned nothing at all; "trinidad", "saint lucia" and "st lucia" returned only unrelated items or national law collections. These are keyword hits, and they miss anything filed under another name or kept private.
The "patois" search needs a closer reading, because it is where the audio work sits. The author neddamj published patois-sr, 5,110 audio clips with transcriptions (4,088 for training, 1,022 for testing, about 1 GB), created on 19 May 2025. Its card names it "Patois Music Transcription" and does not say which country's Patois it covers, so we cannot call it Jamaican. Its card is cited by neddamj/whisper-small-patois, while neddamj/whisper-base-patois and neddamj/whisper-tiny-patois cite sibling sets named patois-sr-base and patois-sr-tiny, which did not appear in our search. On the cards, whisper-base-patois reports a word error rate of 0.59 and whisper-tiny-patois 0.80. The card for whisper-small-patois, which is titled "Whisper Tiny - Patois" and built on whisper-tiny, reports 120.26, which means more than one wrong word for each word spoken. Two large-v3 test adapters from the same author, dated 13 September 2026, are thinly documented: one has a README, the other none. A separate item, BroadcastingCommission_Patois_Dataset, holds two audio files and two rows of read-aloud Jamaican news from 2022.
So the Jamaican material is small and scattered, not absent: 34 clips in the Patwa set, 2 in the news set, and a 5,110-clip music set of unstated origin. All of it is speech. The one Patois text-generation model, JamaicanTinyStories, has three transformer layers and its card says training was interrupted. We found no public Patois evaluation set for frontier text models, and that is the gap that matters to a buyer. The prices fell for everyone; the data to test them in Patois did not.
Preparation Asymmetry, applied to a price list
Preparation Asymmetry is the term in Adrian Dunkley's published FAQ for the structural gap between nations that build AI systems and nations that inherit them. Caribbean governments and firms mostly inherit, so the rules, the training data and the model values arrive from elsewhere. The price arrives from elsewhere too. Two companies set it on one Tuesday, and a buyer in Kingston learns it from the same announcement as a buyer in Seattle. OpenAI's earlier GPT-5.6 rates were described as "promotional pricing", which tells you a price can be a marketing decision. A budget built on today's rate carries a bet that the rate stays.
The second method is Extracted Intelligence, from the same FAQ: data about a community's language and decisions leaves as raw material and returns as a product the community pays for. Every Patois prompt sent to a foreign API is a small example. A test we suggest, not one Adrian has published: list what data the tool collects, list what the community receives in return, and if the second list is empty, price the data.
Picture a team lead at a Kingston credit union at a quarter to five on a Friday. A member texts "mi nuh see di money inna mi account yet." The assistant has to know this is a missing deposit, who to route it to, and whether the member is angry or only tired. We invented that message, and it is exactly the sort of sentence nobody has benchmarked. Whether it costs two tokens more than its English twin matters less than whether the answer is right.
What this means in Jamaica and the wider Caribbean
Jamaica is the clearest case because it has both the demand and an active build effort. NJIT reports that the Jamaica Artificial Intelligence Association and the University of the West Indies are developing IRIS, described as the Caribbean's first sovereign large language model. Jamaican banks, insurers and the BPO sector in Kingston and Montego Bay already serve Patois speakers by phone and message, so the cheap English-first models will reach them before any local model does. Our inference: those firms will adopt a rented model first and discover its Patois accuracy in production, because no public Jamaican test exists to check it beforehand.
Elsewhere the picture is thinner. Haitian Creole has the most material on the Hub, mostly speech and translation. Guyanese Creole has one text fine-tune published on 15 April 2025, and Papiamento, spoken across Aruba, Bonaire and Curaçao, has a single text model. Barbados has two Whisper large-v3 speech models labelled "bajan", created on 11 June 2026, with no model card to say what data trained them. Our searches for Trinidad and Tobago ("trinidad", "trinidadian") and Saint Lucia ("saint lucia", "st lucia", "kweyol") found no dialect models or datasets, only unrelated items and national law collections, though that may reflect how people label things. The regional pattern: a few volunteers have built speech tools, and almost no one has published an evaluation set that tells a buyer whether a frontier model understands the language.
The lab's view, stated plainly: cost no longer justifies building a Caribbean language model, and the three remaining reasons are accuracy, custody and continuity. Our earlier piece on telling a local model from a rented one covers how to check custody, and the case on Caribbean Creole as an AI blind spot covers the language side. Accuracy is the one a buyer can test this month, which is why the steps below start there.
What to do this week
- List 20 real messages from your own customers, with consent and names removed, and write a Standard English version of each beside it.
- Measure the token count of every pair in the vendor's own token counter, divide Patois by English, and record the median ratio by Friday 16 October 2026. A test we suggest: if the median is above 1.5, redo your budget at that multiple.
- Build a 50-question test set from your own tickets, have two fluent staff members write the correct answers, and score two models on it by Friday 23 October. Set your launch bar, for example 90% correct, before you see any score.
- Ask each vendor in writing, by Friday 9 October, for its notice period before a price change and whether prompts are kept or used for training.
- Check the Hugging Face Hub for your own language, and write down how many models and datasets exist and how many downloads they have.
The first two steps cost an afternoon and no money. A pilot lead at a bank or ministry can run them without a procurement process.
Where this can go wrong
A bank or ministry sees a 50% price cut, reads an English benchmark, and moves its Patois-speaking line to a cheap model. The model misreads a payment dispute, a pension query or a symptom described in dialect, and the person who pays is the customer, usually one with the least recourse. Cheaper tokens also make a wrong answer cheaper to repeat at scale. The guardrail is the test set from step three, a launch bar agreed in advance, a human route for anything the model flags as uncertain, and a log of which answers were wrong. A second risk is lock-in. Budgets set at today's price are exposed if a vendor reprices, which is why step four asks for notice periods in writing.
Where this analysis is weakest
We do not know the Patois token ratio, and the 15 times figure comes from other languages. We do not know how well Opus 5.5, Sol or Luna understand Patois, because no public test exists. The Hub counts depend on how people name their uploads, and they leave out private work such as IRIS. The arithmetic assumes 200 tokens in and 300 out, which is our illustration and not a measured average. If the median ratio from your own 20 messages comes back near 1.0, the token argument disappears and the accuracy test is the only one left.