Claude Haiku 5.5 vs GPT-6 Luna: Same $0.10/$0.50, Different Breakpoints

Short answer: Anthropic launched Claude Haiku 5.5 on October 7 at $0.10 input / $0.50 output per 1M tokens for prompts up to 100,000 tokens — exactly matching GPT-6 Luna on the sticker. Past that line the two models diverge hard: Haiku 5.5 steps to $0.50/$2.50 for the whole request, while Luna keeps its base rates until 272,000 input tokens. If your prompts sit under 100K, the rate cards tie. If they don't, Luna is 5× cheaper on that workload until you cross 272K, and still cheaper after. I ran both through the same token counts below so you can see where the tie breaks.

Published October 8, 2026 · Updated October 10, 2026 · Pricing verified October 10, 2026

What changed on October 7

Haiku 5.5 is Anthropic's new small model for high-volume work — classification, extraction, routing, subagents. It replaces Haiku 4.5 at $1.00/$5.00 per 1M, so the headline cut is 90% under 100K tokens. The model carries a 1M-token context window and 128K max output, supports adaptive thinking with an effort control, and is live under the ID claude-haiku-5-5.

Two fine-print items matter more than the launch-day coverage suggests. First, Haiku 5.5 uses the newer Claude tokenizer, which Anthropic says counts roughly 30% more tokens for the same text than Haiku 4.5 did. Your per-token rate fell, but your token count for identical copy rises, so recount before you budget. Second, the same announcement cut Claude Sonnet 5.5 cache reads by 50%, from $0.20 to $0.10 per 1M — now level with GPT-6.1 Sol on cached input. If you run Sonnet 5.5 agents that resend a stable prefix, that line just halved. Prices verified October 8, 2026 — always confirm on the official Anthropic pricing page and OpenAI's pricing page before you commit.

Rate card, side by side

Read the table the way your invoice will. Under 100K, every line ties: $0.10 in, $0.01 cached, $0.125 write, $0.50 out. Between 100K and 272K, Haiku is already on its higher row while Luna is still on base — a 5× gap on input, cache, and output for the exact same token mix. Above 272K, Luna finally steps up too, to $0.20/$0.75, which narrows the gap but does not close it: Haiku at $0.50/$2.50 still costs 2.5× on input and about 3.3× on output. Both models discount Batch by 50% on input and output, so Batch does not change the ranking — it halves both bills.

Where the tie breaks: three illustrative workloads

These are my scenarios, not industry averages — same token counts pushed through both rate cards at Standard pricing. Luna numbers come from this site's Luna cost calculator; Haiku numbers use Anthropic's published rates above.

The middle row is the one that will surprise people who only read the launch headline. A 150K-token prompt — a medium codebase dump, a long support thread, a fat retrieval bundle — is unremarkable on Luna and already in Haiku's penalty tier. At 100K requests a month that gap is $7,000 vs $1,750. The first row shows the opposite: short, cache-friendly traffic — exactly the classification and routing work Haiku targets — bills identically. Neither model is "cheaper." Your prompt-length distribution decides.

Caching, Batch, and the tokenizer — the lines that move monthly spend

• Caching ties under 100K. Both read cache at $0.01/M and write at $0.125/M (5-minute TTL on the Anthropic side). Your cache hit rate, not the rate card, decides what you save. Past 100K, Haiku's cache lines step up 5× with everything else; Luna's double only past 272K.
• Batch halves both. Anthropic's Message Batches API and OpenAI's Batch both take 50% off input and output. On the 50K scenario above, Batch drops the per-request cost to $0.00195 on either model. If your work can wait, run it there — the tie survives, just smaller.
• Count your own tokens. Haiku 5.5's newer tokenizer inflates counts by roughly 30% versus Haiku 4.5 for the same text, per Anthropic's migration notes. Tokenizers differ across providers, so a "150K-token prompt" on one API is not the same text as 150K on the other. Measure with each provider's counter before you compare monthly totals — my methodology page covers how this calculator handles that.
• Capability is your eval, not the rate card. Anthropic publishes benchmark gains for Haiku 5.5 over 4.5; OpenAI positions Luna for the same high-volume tier. This site measures pricing. Run your own classification or extraction set on both before you move production traffic.

Run your prompt lengths through the calculator

Update, October 10, 2026: Haiku 5.5 is now in the widget above — the calculator applies the 100,000-token breakpoint automatically, repricing the whole request past that line, so you can run both models against your own prompt-length mix directly instead of approximating. Under 100K the two bills still tie at the same token mix; over 100K and under 272K, watch the Haiku row jump while Luna holds base rates (then recount your tokens — see above). The full provider comparison, including Sonnet 5.5 at its new $0.10 cache-read rate, lives on the GPT-6 vs Claude cost page; tier-selection tactics are in my GPT-6 API bill guide.

Quick answers

• Is Claude Haiku 5.5 cheaper than GPT-6 Luna? At or under 100K input tokens per request, the published rates are identical. Above 100K, Haiku costs 5× Luna until Luna's own step-up at 272K, and stays more expensive after.
• What does 1M tokens cost on each? Input-only at base rates: $0.10 on both. A 1M-input request itself bills at long-context rates on both — $0.50 input on Haiku, $0.20 input on Luna — because the whole request moves to the higher row.
• Did anything else get cheaper on October 7? Yes — Sonnet 5.5 cache reads dropped from $0.20 to $0.10 per 1M tokens, matching GPT-6.1 Sol's cached input.
• Should I switch? Only your prompt-length histogram knows. If 90% of your requests sit under 100K, decide on quality and latency. If a fat tail runs past 100K, price that tail on Luna before you move.

Official sources, verified October 10, 2026: Claude Haiku 5.5 model documentation, Anthropic pricing, Anthropic newsroom — Haiku 5.5 announcement (Oct 7, 2026), GPT-6 Luna model documentation, and OpenAI API pricing.