Small AI models in October 2026: Haiku 5.5, GPT-6 Luna, Gemini 3.8 Flash — dozens of times cheaper, and three traps in the price list

A shop receives about 30,000 customer messages a month through Zalo, Facebook and its website. The owner wants AI to do two things: put each message in the right group (price question, return, complaint, order) and draft a reply for staff to approve. The first question: which model?

The natural instinct is to pick the strongest model available. But October 2026 is a good moment to ask the opposite: what is the smallest model that can do this job? In about five weeks, all three big labs released a new generation of small models, and the gap between them and the flagships has narrowed considerably — while the price gap has not.

Three new small models

• GPT-6 Luna (OpenAI), released 22 September. $0.10 per million input tokens and $0.50 per million output tokens — about half the previous generation. Context window of about 1 million tokens.
• Claude Haiku 5.5 (Anthropic), released 7 October. $0.10/$0.50 for requests under 100,000 tokens; above that, $0.50/$2.50. Anthropic says it is about 90% cheaper than Haiku 4.5 on short requests. The first Haiku-class model with a reasoning-effort setting.
• Gemini 3.8 Flash (Google), released 2 September. Introductory price of $0.75/$3.75 until the end of 2026, rising to $1.50/$7.50 from 1 January 2027.

For comparison: the mid-tier models (Claude Sonnet 5.5, GPT-6.1 Sol) cost $2/$10, which is 20 times Luna and Haiku 5.5 per token.

How strong is small

On the Artificial Analysis Intelligence Index (version 4.3) — an independent index combining around ten tests — the numbers are as follows. Each model's score changes with the reasoning level chosen, so each line states the level:

• Claude Haiku 5.5: 43 at its highest level, 41 at xhigh, about 34 at the default. Cost about $0.12 per task at xhigh, $0.21 at the highest level.
• GPT-6 Luna: 37–38 at its highest level (two pages from the same benchmarker differ by a point), 22 at low. Cost about $0.07 per task at the highest level.
• Gemini 3.8 Flash: about 41 at high.
• For reference: GPT-6 Astra scores 53 at about $3.26 per task; Claude Opus 5.5 leads with 58.

In other words, Haiku 5.5 at xhigh reaches about 77% of Astra's score at roughly one twenty-seventh of the cost per task. Luna reaches about 72% of the score at roughly one forty-seventh of the cost.

Between the two cheapest small models: Haiku 5.5 scores a few points above Luna on the independent index, while Luna is cheaper per task. Anthropic also published a table comparing Haiku 5.5 with Luna, in which Haiku leads on every test both have results for — for example, computer use (OSWorld) 72.4% versus 48.9%. Those are the vendor's own figures and should be read as a claim to verify; the direction matches the independent index, but the size of the gap has not been confirmed by a third party.

These numbers come from hard tests — multi-step reasoning, coding, specialist knowledge. Sorting a price-question message is far easier. A 15-point gap on a composite index does not mean the small model makes 15% more mistakes on your task — it may make almost no extra mistakes, or more than that. The only way to know is to test, and the end of this article shows how.

Costing the shop's 30,000 messages

Assume each message needs about 2,000 input tokens (instructions, the list of groups, a few examples, the message itself) and 400 output tokens (the chosen group and a drafted reply), at a low reasoning level. At list prices:

• Haiku 5.5 or Luna ($0.10/$0.50): about $12 a month
• Gemini 3.8 Flash: $90 a month at the promotional price, $180 at the 2027 price
• A mid-tier model ($2/$10): $240 a month
• Split: 85% of messages handled by the small model, the harder 15% escalated to a mid-tier model: about $46 a month

Two things stand out. First, the small model is 20 times cheaper — exactly as the price list says. Second, and more important: at this scale, every option costs less than a few days of one employee's salary. The $228 monthly difference is real, but it is not the deciding factor. The deciding factor is which model makes the fewest mistakes, because every misfiled message or reply that has to be rewritten costs staff time — which is far more expensive than tokens.

The calculation flips when volume is hundreds of times larger — processing millions of documents, or an automated task running continuously. Then a 20-fold gap becomes serious money, and moving most of the volume to a small model is the most valuable decision you can make.

Trap one: the price cliff at 100,000 tokens

Haiku 5.5 has two price levels based on request length. According to analyses citing Anthropic's documentation, when a request exceeds 100,000 tokens, the entire request — not just the part above the line — is billed at the price five times higher, including tokens read from cache.

A worked example with 1,000 output tokens:

• A 90,000-token request: 90,000 × $0.10 / 1 million + 1,000 × $0.50 / 1 million ≈ $0.0095
• A 120,000-token request: 120,000 × $0.50 / 1 million + 1,000 × $2.50 / 1 million ≈ $0.0625

A request one third longer costs 6.6 times as much. Each request is priced on its own, so one long request does not push others into the higher price — but if your system often sends long requests, the difference is large.

Two situations where the line is easily crossed without anyone noticing:

• Long conversations: every turn resends the whole history. A long customer-service conversation can cross the line at turn twenty-something without anyone changing a line of code.
• A new tokenizer: Anthropic says Haiku 5.5 uses a new tokenizer that produces slightly more tokens for the same content. One tester found a request measuring 80,000 tokens on Haiku 4.5 came to about 104,000 on Haiku 5.5. Measure length with the model you actually use, not with old numbers.

Many other models also have separate pricing for very long requests, at different thresholds — read the full pricing page for the exact model you use, not just the first line.

Trap two: promotional prices and creeping token counts

Gemini 3.8 Flash's $0.75/$3.75 is a promotional price until the end of 2026. From 1 January 2027 it doubles. On top of that, Artificial Analysis found that 3.8 Flash's cost per task rose about 40% over 3.7 Flash with no change in the per-token price, because the new model writes about 30% more tokens per task and needs more turns on multi-step work.

The general lesson: a new version of the same model line can cost more per task even when the price list stays the same. With each upgrade, rerun your tests and compare token counts, not just scores.

Trap three: the reasoning setting

All three models let you choose a reasoning level — how much the model "thinks" before answering. Thinking tokens are billed as output. According to Artificial Analysis's measurements, Luna's cost per task is about $0.0045 at the lowest level and about $0.07 at the highest — roughly fifteen times apart on the same model and the same price list.

Higher levels score better on hard tests. But for sorting a message into four groups, a low level is usually enough. The rule: start at the lowest level and only raise it when your tests show you need to. Many teams do the opposite — leave the default or a high level "to be safe" — and pay several times over for thinking they did not need.

How to test whether a small model is enough

One — take 200 real requests. Pick 200 recent messages, documents or requests from exactly the task you plan to automate. Run the same prompt on a small model and a mid-tier model, at the lowest reasoning level. Store the results and the token counts.

Two — measure according to the kind of task. For tasks with a clear answer (classification, extracting a phone number, extracting an order code), count how often the small model matches the large model and the person who handled it before. For writing tasks (drafting replies), take 30 random pairs, hide the model names, and ask two staff members to pick the better one or mark them equal.

Three — look closely at the mistakes. The small model's first ten errors tell you more than its accuracy rate. If they cluster in one kind of message (say, messages that both complain and ask for an exchange), that is the kind to escalate to the large model. If they are spread evenly, the small model may not be enough for this task.

Four — set an escalation rule. Let the small model go first. Escalate to the large model when the output has the wrong format (for example, JSON missing a required field), when the small model chooses an "unsure" group you included in the list, or when the request exceeds a length you set. Track the escalation rate weekly — if it creeps up, the content of the messages may have changed.

Five — budget at standard prices and measured tokens. Do not multiply answer length by the price list — that ignores thinking tokens and repeated calls. Use the actual token totals from step one.

Conclusion

October 2026's small models are not weaker versions of the large ones. They are tools for a different kind of work: tasks repeated many times, with fairly clear right and wrong answers, where speed and cost per run matter more than deep reasoning.

For small and medium businesses, the question is usually not "how much can we save on tokens" — at a few tens of thousands of requests a month, every option is cheap. The right question is which model makes the fewest mistakes on your exact task, and whether the system knows to send hard cases to the right place. All three traps in this article — the 100,000-token cliff, promotional prices, the reasoning setting — show up in a single test run of 200 requests.

The previous article in this series compared three mid-tier models at the same $2/$10 price. If you want to set up testing and split work between small and large models on your own data, contact us or see KonexForge's AI & ML capability.