When two models list the exact same price, the price is the one number that can't help you choose. Claude Haiku 5.5 (shipped October 7) and GPT-6 Luna (September 22) both bill $0.10 per million input tokens and $0.50 per million output, both carry roughly 1M-token context windows, and both cap output at 128K. So here's the actual decision, in one screen:
- Prompts routinely over 100K tokens, or you live in Codex → GPT-6 Luna. Its cheap rate holds out to 272K tokens; Haiku's jumps 5x at 100K.
- You need computer-use/terminal-agent muscle, live in Claude Code, or need Bedrock/Vertex/Azure → Claude Haiku 5.5. Just keep prompts under the 100K line.
- **Whichever you lean toward, decide on cost-per-completed-task over your own text — not the matching $0.10** — because the two models don't even count tokens the same way.
That last point is the one the rate card hides, so let's start there.
1. The tokenizer makes "$0.10 = $0.10" a lie#
A per-token price is only comparable when a token means the same thing on both sides. It doesn't.
Anthropic states plainly that Haiku 5.5's new tokenizer turns the same input text into about 30% more tokens than Haiku 4.5 did; Simon Willison measured roughly 1.25x on one long prompt, and chat-style text has tested higher. OpenAI's tokenizer splits the same text differently again. So feed byte-for-byte identical input to both models and you get two different token counts — which means two different bills at the same $0.10 rate.
A matching per-token price with mismatched tokenizers is two different prices wearing the same sticker.
This is also why the "Haiku 5.5 is 90% cheaper" line needs an asterisk. The per-token rate did fall 90% (from $1/$5 to $0.10/$0.50), but once you count ~30% more tokens for the same work, the effective cut is smaller — which is exactly why Anthropic's own headline is "~75% cheaper on average." The rule generalizes: when a model ships a lower price and a new tokenizer in the same release, net the two before you re-forecast anything.
What to do: take a representative sample of your real prompts and completions, run it through both models' token counters, and compare total tokens, not rates. (Here's how to count Claude's tokens before you send them.) That five-minute measurement beats any launch-day pricing table, including this one.
2. The cliff is in a different place#
Both models have a cheap tier and a cliff where it ends — but the cliffs sit far apart.
- Haiku 5.5: $0.10/$0.50 up to 100K tokens, then a 5x jump to $0.50/$2.50 above it. Anthropic notes ~90% of Haiku 4.5 traffic sat under 100K, so most workloads never hit it — but a long-running agent that re-sends a growing context, or a RAG app stuffing big documents, can.
- GPT-6 Luna: the short-context rate holds all the way to 272K tokens before OpenAI's long-context pricing applies — the same cliff dynamic we pulled apart for GPT-5.5's 272K long-context pricing.
If your prompts live comfortably under 100K, this doesn't matter and you're choosing on §1 and §3. If they regularly push past it — long documents, big tool catalogs, fat agent histories — Luna keeps the cheap rate 2.7x further, and that gap can dwarf any tokenizer difference. Know your prompt-size distribution before you pick; the median matters less than the 90th percentile.
3. Capability and harness: the tie actually breaks here#
On raw agentic capability, Haiku 5.5 is the stronger story on Anthropic's own numbers — OSWorld 2.1 offline subset 72.4% (37.1% if you demand every checkpoint) vs Luna's 48.9%, and Terminal-Bench 4.0 39.2% vs 16.4%. Two honest caveats: those are vendor-reported, and they aren't like-for-like — Anthropic ran Haiku in Claude Code and Luna in Codex, so the harness is part of what's being measured. Artificial Analysis's independent Terminal-Bench number for Haiku came in lower, ~33%.
Which is the useful insight, not a disqualifier: the harness moves your results as much as the model does. Haiku 5.5 is built to shine in Claude Code and across Bedrock/Vertex/Azure (the data-residency play); Luna is built to shine in Codex and the OpenAI ecosystem. If your team already works in one of those, that gravity will move your velocity more than a few benchmark points. This is the same logic that decides the tier above, where GPT-6 Sol and Claude Sonnet 5.5 also landed on an identical price — the harness, not the rate, is the switch.
One migration note if you choose Haiku: computer use now requires the computer_toolset_20260801 toolset (the old computer_20250124 is out), and sending a non-default temperature, top_p, or top_k returns a 400. Budget an afternoon for the cutover, not a commit.
The decision, restated#
The sticker is a tie, so ignore it. Pick Luna for large prompts and the OpenAI/Codex world; pick Haiku 5.5 for computer-use work, Claude Code, and the big-cloud residency story — and in both cases run the cost-per-completed-task test on your own repo before you commit. The matching $0.10 is where this comparison starts, not where it ends — and if you only remember one thing, remember that a per-token price lies the moment the tokenizers differ. For the week's other moves in the cheap-and-local tier, see the morning Wire.
