Same Sticker, Different Bill: Haiku 5.5 vs Luna in 12 Runs

Same sticker, different bill. Anthropic introduced Haiku 5.5 on October 7, 2026, cutting unit prices 10x for work under the 100k-token line: $0.10 input and $0.50 output per million. Cross the line and the rate jumps 5x ($0.50 and $2.50), with Anthropic promising 75% lower average running cost. According to Anthropic, nine out of ten previous Haiku requests fit under that line.

On the OpenAI side GPT-6 Luna lists $0.10 and $0.50 for short context, matching Haiku to the cent, while long context rises gently to $0.20 and $0.75. According to OpenAI, the short-versus-long columns in the current price table confirm this split. But the ArtificialAnalysis team adds the decisive detail: Haiku burns about 162k output tokens per task against 50k for Luna, reaching similar intelligence scores with roughly 3x the tokens . According to ArtificialAnalysis, the token counts behind the Intelligence Index confirm the gap.

Both models stay rough in the room test

In the first task the creator hands both models four photos of his room from different angles and asks for an explorable Three.js room. Haiku spends 355k tokens in 32 minutes at $1.48, Luna finishes the same job around $0.18. Both outputs stay rough; Luna mixes up chair and table orientation while Haiku arranges a slightly more flexible layout. Why single-prompt 3D worlds are hard to grade is also the subject of the WorldBench study published on arXiv. According to arXiv, only a judge that both runs the world and reads the program gives a fair score.

Anthropic's own scoreboard puts Haiku ahead of Luna on every row: 1620 vs 1437 on knowledge work, 72.4% vs 48.9% on computer use, 39.2% vs 16.4% on terminal coding, 46.4% vs 29.1% on visual reasoning. On the independent side ArtificialAnalysis scores Haiku 43 against 38 for Luna on its Intelligence Index, placing Haiku at the top of the small class. The DataCamp review adds the backstory: the expected Haiku 5 never shipped, 5.5 arrived after a full year, two weeks after Opus 5.5 and nine days after Sonnet 5.5. According to DataCamp, the release calendar confirms this sequence.

Old buildings never fall in the civilization run

The second task renders a civilization story from the Stone Age to the future with one prompt. Haiku produces a pleasant front end in 35 minutes for $2.68, flowing from pyramids to the medieval era, yet never thinks to demolish old structures as new ones rise; Luna trips on the same point, keeping old buildings inside while new ones ring the outside. Where big models test and repair their own output, the small ones show no such self-correction. The creator's verdict for this class of work is blunt: high-volume daily jobs like mail, summaries, reports and spreadsheet reading are these models' home, not 3D world building.

In the arena the robot programs written by each model face off; after more than a hundred bouts Haiku finishes clearly ahead at a reported 156 to 45 . Haiku's edge in reasoning and strategy looks genuine. Yet Opus 5.5 beats both even at its lowest effort setting, and whether Sonnet at low effort could have made it a close duel stays untested. The picture matches the Firecrawl comparison: the Anthropic coding helper goes deep in the terminal loop while the OpenAI product spreads across more surfaces, so the choice follows the task. According to Firecrawl, the surface and sandbox comparison supports exactly that reading.

The third task asks for a landing page for an imaginary smart mug. Haiku delivers a decent page in warm tones with its brand colors, pricing, FAQ and a preorder panel, but section sizes balloon at random and an oversized phone visual with ragged spacing costs points. Luna builds a tidier page in brand typefaces that holds together at every resolution; even the naming is telling, Cor from Haiku against Cora from Luna. Time reads 13 minutes against 4 minutes , cost around $0.88 against $0.04 in Luna's favor. The bonus run hands the same prompt to Sonnet 5.5, which finishes in 30 minutes at $4.76 with a livelier, animated page; small models would need 10-15 prompts to get near that level, if ever.

The 12-run invoice ends the theory

The final invoice closes the gap between theory and practice: across 12 runs Haiku lands around $12.5 while Luna stops at $0.81, with time at 287 minutes against 93. Two engines drive the gap: Haiku crosses the 100k-token line into the 5x tariff, and it consumes about 3x Luna's tokens per task. According to Claude, the pricing page's above-threshold rows confirm the first engine.

Who should pick which? For everyday work on a $20-a-month plan both models are more than enough; small models protect the budget on mail, summaries, reporting and file tidy-ups. For large projects the Sonnet and Sol class stays more consistent, and the creator's rule is simple: buy the best model your money reaches, don't chase cheap models down a 10-15 prompt detour.

The last word is that list price and true cost are different things. Haiku 5.5 at low-to-medium effort is a quiet workhorse for volume jobs; a 1M-token context window with adaptive thinking lifts it into a different league from Haiku 4.5. Luna stays the budget friend on long runs with its frugal token habits. Both sip pennies under 100k tokens; above the line Luna's calm tariff pulls ahead.