Claude Haiku 5.5 vs GPT-6 Luna: Same Price, Different Models (2026)

The short answer: Claude Haiku 5.5 and GPT-6 Luna list at exactly the same price — $0.10 per million input tokens, $0.50 per million output tokens — and the official benchmark record is 2–2. Luna leads the two knowledge-and-work composites; Haiku 5.5 leads computer use, coding, and raw speed. The real pricing difference is structural: Haiku 5.5 pays 5× rates on any prompt over 100K tokens, while Luna stays flat. Under 100K, test both on your own data and let the bill decide; over 100K, Luna is the cheaper pick by default.

Anthropic released Haiku 5.5 on October 7, 2026 and benchmarked it directly against Luna (out since September 22, 2026), so this is one of the few head-to-heads where both sides' numbers come from a single official source.

This guide compares them on price, official benchmarks, speed, and long-context costs, plus one workload most comparisons skip: generating motion graphics as code. If you are actually comparing video generation models, start with the model index, the tools index, or the head-to-head Arena.

Quick verdict

• Coding and computer use: Haiku 5.5. It scores higher on both official coding (FrontierCode 46.4% vs 42.4%) and computer-use (OSWorld 72.4% vs 48.9%) benchmarks — Anthropic's own numbers.
• General knowledge and work tasks: GPT-6 Luna. It leads both composite work benchmarks Anthropic published: GDPval-AA 1840 vs 1620, AA-Briefcase 1824 vs 1578.
• Prompts under 100K tokens: a tie on price. Identical input/output rates, and cached input runs about 90% off on both.
• Prompts over 100K tokens: GPT-6 Luna. Haiku 5.5 is the only current Claude that pays a premium above 100K prompt tokens — 5× on both input and output.
• Iteration speed: Haiku 5.5. Third-party measurement clocks it at 237 output tokens per second against Luna's 137, though it also writes more tokens to finish the same task.

Pricing: identical rates, one big exception

Both models list the same per-token rates, with one structural difference: Haiku 5.5 is billed by prompt length, Luna is flat across its whole 1M-token window.

The over-100K premium applies to the whole request, not just the part past 100K, and it covers the cache rates too. Every other current Claude bills a 900K-token prompt at the same per-token rate as a 9K one. Anthropic also notes prompts up to 100K tokens made up about 90% of requests to the previous Haiku — for most existing workloads, the premium tier never kicks in.

What that means in practice, using the published rates:

• A 90K-token prompt with 8K of output: 0.9¢ + 0.4¢ on either model. Identical bills.
• A 150K-token prompt with 8K of output: 7.5¢ + 2¢ = 9.5¢ on Haiku 5.5, versus 1.5¢ + 0.4¢ = 1.9¢ on Luna. Five times the cost for the same request.

Two more price notes:

• Batch API: Anthropic bills Haiku 5.5 batch requests at half price — $0.05/$0.25, and $0.25/$1.25 above 100K.
• Family context: Haiku 5.5 replaces Haiku 4.5 ($1/$5) at about 75% lower average cost, and 90% lower for prompts under 100K. Claude Sonnet 5.5 ($2/$10) is the next step up when a task outgrows the small models.

Official benchmarks: 2–2

Anthropic's Haiku 5.5 announcement published results for both models; the last two rows come from Artificial Analysis, which measures both in the same lab.

Luna holds an edge on composite work and knowledge scoring; Haiku 5.5 on operating real software, writing code, and raw throughput. Neither model sweeps the table, and Anthropic left several Luna cells blank in the same comparison (Humanity's Last Exam, Terminal-Bench 4.0), so close scores should be read as close.

Cost per task makes the verbosity gap concrete: on Artificial Analysis' measure of what a typical benchmark task actually bills, Haiku 5.5 lands at $0.21 against Luna's $0.07 — a three-times gap on identical list prices. Token appetite is the cause: Haiku 5.5 writes about 4.4× the median output tokens on the same test suite, while Luna runs near typical.

Verbosity is why identical list prices can still produce different bills. One lever helps: Haiku 5.5 is the first Haiku-class model with an adjustable effort setting (Low to Max), so you can trade intelligence for a smaller bill. Early community reports on these two models split for exactly this reason: some find Luna cheaper end-to-end, others find Haiku's speed and benchmark edge worth the extra tokens.

Long context: same 1M window, different bills

• Both models accept 1 million input tokens. Haiku 5.5 outputs up to 128K tokens.
• Luna keeps flat per-token pricing across the whole window.
• Haiku 5.5 switches to its 5× tier once a prompt passes 100K tokens.

For workloads that live above 100K prompt tokens — whole-codebase reviews, long transcripts, large generated files — Luna is the cheaper model by a wide margin. Under 100K, price alone decides nothing.

Motion graphics: what these models can actually do

Both models produce motion graphics the way current LLMs do: as code. Animated SVG and CSS, GSAP timelines, Lottie JSON, Remotion compositions for video-shaped output, Manim for explainer-style math animation. Neither model renders video footage itself — that is a job for a video generation model from our tools index. For UI graphics and explainers, a code-writing LLM is often the faster pipeline.

Why there is no video-length cost comparison here: these models bill by token, and there is no fixed mapping from tokens to seconds of finished animation — a motion-graphics render is generated as code whose length depends on scene complexity, not on playback duration. Per-second pricing only exists where a model actually outputs footage, which is what the video model price database tracks.

What the verified numbers say about that workload:

• Code generation: close. FrontierCode puts them four points apart (46.4% vs 42.4%, Haiku 5.5 ahead).
• Driving design tools and browsers: Haiku 5.5, by a lot. OSWorld measures operating real software, and 72.4% vs 48.9% is the widest gap in the official table. Browser use is one of Anthropic's stated Haiku workloads.
• Fast iteration loops: Haiku 5.5. Regenerating an animation a dozen times in a session favors the 237-token-per-second model.
• Large project files: Luna. A full Remotion project or a long Lottie JSON can push prompts past 100K tokens, where Luna's flat pricing wins.
• Cost per revision: nominally identical, but Haiku's verbosity shows up here as more generated tokens per iteration, blunted when repeated context hits the cache.

A practical split: use Haiku 5.5 for interactive iteration and browser-driven workflows; use Luna when prompts are huge or you are sweeping many long-context revisions on a budget. And when the deliverable is footage rather than vector or UI animation, switch to the video side of this site: model-vs-model pricing comparisons, the Arena, or free AI video generators.

Early demos worth watching:

• Motion design on a budget — WorldofAI's Haiku 5.5 review has a dedicated motion-graphics chapter: a design job that took the bigger models over 40 minutes finished in about 15 on Haiku 5.5, plus an animation head-to-head against GPT-6 Luna the reviewer scored clearly for Haiku.
• Official interactive UI — OpenAI's own Intelligent UI demo shows the GPT-6 generation answering inside ChatGPT with fully interactive, rendered interfaces — the same code-driven rendering these animation workflows ride on.
• Driving After Effects — a wave of tutorials, led by Higgsfield's official demo, wires GPT-6 Astra — the family's bigger sibling above Luna — into After Effects directly: editable layers and keyframes, kinetic type, full explainers, professional motion tooling driven from a chat box.

Which should you pick?

• Pick Haiku 5.5 for coding, agentic and computer-use work, latency-sensitive products, and fast iteration loops.
• Pick GPT-6 Luna for knowledge-heavy work, any job whose prompts reliably pass 100K tokens, and cost-sensitive batches where output volume runs high.
• Pick neither alone if your tasks are hard: both vendors position these as the small tier under heavier models (Sonnet 5.5 at $2/$10 on Anthropic's side). Route the failures up a tier instead of fighting the small model.

FAQ

Is Claude Haiku 5.5 better than GPT-6 Luna?

The official head-to-head is 2–2: Luna wins the two composite work benchmarks (GDPval-AA, AA-Briefcase), Haiku 5.5 wins computer use (OSWorld) and coding (FrontierCode). The better pick depends on which of those your workload resembles. Third-party testing puts Haiku 5.5 higher on general intelligence (43 vs 38) and well ahead on speed, with the caveat that it writes far more tokens.

Which is cheaper, Haiku 5.5 or GPT-6 Luna?

At list price they are identical: $0.10 input / $0.50 output per million tokens. Costs diverge above 100K-token prompts, where Haiku 5.5 pays 5× rates and Luna stays flat, and in practice, because Haiku 5.5 tends to write more tokens per task than Luna.

Why does Haiku 5.5 cost more above 100K tokens?

Anthropic prices Haiku 5.5 by prompt length. Once a request's prompt passes 100,000 tokens, the whole request bills at $0.50/$2.50 per million tokens instead of $0.10/$0.50 — input, output, and cache rates all step up. It is currently the only Claude model priced this way; every other model keeps one rate across the full 1M-token window.

Can I use Haiku 5.5 or GPT-6 Luna for motion graphics?

Yes, as code generators: both can write animated SVG/CSS, GSAP timelines, Lottie JSON, Remotion compositions, and Manim scenes. Haiku 5.5 fits fast interactive iteration (faster output, stronger software-driving scores); Luna fits very large project files, where its flat pricing beats Haiku 5.5's 100K+ tier. Neither generates video footage — for that, use a dedicated video model (free options compared here).

When were Haiku 5.5 and GPT-6 Luna released?

GPT-6 Luna on September 22, 2026; Claude Haiku 5.5 on October 7, 2026.

What is GPT-6 Luna?

OpenAI's fast, low-cost model in the GPT-6 generation: $0.10 input / $0.50 output per million tokens, a 1M-token context window, text and image input, released September 22, 2026. Per OpenAI, it powers ChatGPT's Free and Go tiers.

Related comparisons on vidirect

The same head-to-head idea, applied to video models:

• MiniMax H3 vs Wan 2.2 — official and reseller per-second pricing side by side for two current video models.
• Kling 2.1 Master vs Veo 3.1 — real output clips from both models playing back to back.
• Veo 3 vs Veo 3.1 — generation over generation within one family.
• Video model price database — every reseller's per-second price in one filterable table, with per-channel pages like fal's rates.
• Kling pricing — per-model rates with a cost calculator.