The Cheapest LLM in October 2026: Claude Haiku 5.5 vs GPT-6 Luna (and the Two 'Cheap' Models That Cost More)

Two models now share the bottom of the frontier price list. Claude Haiku 5.5 and GPT-6 Luna both list at $0.10 per million input tokens and $0.50 per million output. The models people usually compare them with, Gemini 3.8 Flash and DeepSeek Flash, list higher.

We ran all four through the same three script-graded tasks, five times each. All four passed all 15 calls, so accuracy does not separate them on this work. Cost and latency do:

• GPT-6 Luna was the cheapest, at $0.00022 per battery pass.
• Haiku 5.5 cost 36% more than Luna and was the fastest of the four on every task.
• DeepSeek Flash cost 4.6× Luna at its peak-hour rate.
• Gemini 3.8 Flash cost 6.7× Luna as Promptster displayed it at the time, and an estimated 31× once its hidden thinking tokens are counted.

Prices

From shared/pricing.ts, per million tokens:

On list price alone, Haiku and Luna are tied, DeepSeek Flash is 2.4–3× more, and Gemini 3.8 Flash is 7.5×.

Quality: no separation

All four extracted the right order id from a ticket that mentions two, picked the new email over the old one, returned 89.5 as a number, flagged the chargeback threat as "high", and did it in bare JSON with no code fences. All four got 580 on the arithmetic problem. All four wrote exactly three short bullets with no preamble.

What each one cost

Cost per battery pass (one call of each task), averaged over five passes:

For three of the four models, the two cost columns are the same. For Gemini they are not, and the difference is the point of this post.

Why Gemini has two numbers

Gemini reports thinking separately. Its usage object has candidatesTokenCount (the visible answer) and thoughtsTokenCount (reasoning), and totalTokenCount includes both. Until 2026-10-22, Promptster's displayed cost priced the visible answer only, so for Gemini it was too low. It now counts the thoughts and matches the billed estimate. The billed estimate re-prices output as total minus input, which counts the thoughts as output.

The thoughts are not small. On the JSON task, Gemini 3.8 Flash returned 46–47 tokens of JSON after 357–502 thought tokens per call. On the bullets task it thought for 527–830 tokens to write about 40. Across the battery the billed estimate is 4.7× the displayed cost.

The 10-word token probe ("Reply with only the word OK…") makes it plain:

Gemini used the fewest prompt tokens and one output token, then 86–118 tokens of thinking to say "OK".

OpenAI, Anthropic and DeepSeek all count reasoning inside their output figure, so their displayed costs already include it. Luna reasoned on every call (29–39 tokens on JSON, 75–125 on bullets), DeepSeek Flash reasoned more (177–264 on bullets), and Haiku thought only on the arithmetic task (126–145 tokens). All of that is in the numbers above. We covered how reasoning tokens get billed in May.

The "billed estimate" is our reconstruction from the usage object, not an invoice. It assumes thought tokens bill at the output rate and applies no discounts.

After the promo

Gemini 3.8 Flash's $0.75 / $3.75 rate ends 2026-12-31. At $1.50 / $7.50 both columns double: $0.002922 displayed and $0.013752 billed per pass, or about 63× Luna on the billed estimate.

DeepSeek off-peak

We priced DeepSeek Flash at its peak rate. shared/pricing.ts notes off-peak is half, which would put it near $0.0005 per pass. That is still above Haiku and Luna, but closer.

Latency

Median wall-clock time per call:

Haiku 5.5 was the fastest on every task, and its slowest JSON call (1.24 s) beat Luna's fastest (1.44 s). Gemini 3.8 Flash was the slowest on every task, which fits the hundreds of thought tokens it generates before answering.

Haiku 5.5 or Luna?

Same list price, same graded results. The trade is simple:

• Luna is 27% cheaper per pass. On this battery it was the cheapest model of any tier.
• Haiku 5.5 is roughly twice as fast on JSON and bullets, and 0.9 s faster on arithmetic.

Haiku's extra cost comes from two places: Anthropic counts the same prompt as more input tokens (209 against 141 on the JSON task), and Haiku's arithmetic answers ran to 353–429 output tokens, including its thinking, against Luna's 141–193.

If you are running batch jobs where nobody waits on the answer, Luna. If a user is waiting, Haiku. At a million passes the difference is $79.

Where the "cheap" alternatives fit

On this battery, neither Gemini 3.8 Flash nor DeepSeek Flash is the cheap option any more. Both passed everything, but both cost more and ran slower than the two $0.10 models. DeepSeek Flash is closer than it looks if your traffic falls off-peak. Gemini 3.8 Flash is the one to check most carefully: if your cost dashboard reads candidatesTokenCount, it is under-reporting by about 4.7× on prompts like these.

In August, the cheapest frontier configurations clustered around $0.0033–$0.0039 on a harder battery. Different tasks, so not a direct comparison, but the floor has moved.

Method

Three script-graded tasks: strict JSON extraction from a messy support ticket (four keys, exact match, code fences fail), a four-step arithmetic word problem with answer 580, and an instruction-adherence task asking for exactly three bullets of 12 words or fewer with no preamble. Five runs per task per model, temperature 0.7 and max_tokens 1,500, run 2026-10-19 through Promptster's own provider handlers. GPT-6 Luna and Haiku 5.5 reject a custom temperature, so none was sent to them; Luna went out with reasoning_effort: medium. Costs are from shared/pricing.ts at list rate with no cache, batch or off-peak discounts. Latency is wall-clock from one machine, not a provider SLA. The token probe ran three times per model.

For how these compare with open-weight models on the same work, see Benchmarking open-source vs. closed-source models.

Try it yourself

• Run the three-bullet instruction task — the task where Gemini spent the most hidden tokens, no account needed

To run Haiku 5.5, Luna and the Flash models on your own prompts and see each provider's token counts side by side, that's Promptster.

All figures from content/blog/data/q4-2026-battery/ and content/blog/data/q4-2026-token-overhead/ (summaries plus one raw file per call, including each provider's usage object verbatim), run 2026-10-19. Displayed cost uses Promptster's normalised token counts; billed estimate re-prices output as total minus input where the provider reports a larger total. Per-pass costs are the 15-call total divided by five. Ranges are min–max across runs.