Last updated: October 9, 2026
LLMFly AI is a multi-model AI API platform for developers
comparing the real cost of leading models. In this case, the
useful question is not which price card looks cheaper. It is whether
Claude Haiku 5.5 or GPT-6 Luna costs less for a result your application
can actually accept.
Anthropic has finally refreshed its smallest model.
Claude Haiku 5.5 arrives with a surprisingly aggressive headline
price: $0.10 per million input tokens and $0.50 per million
output tokens for prompts up to 100,000 tokens.
Those numbers should look familiar. They are exactly what OpenAI
charges for GPT-6 Luna at standard short-context pricing.
So the obvious comparison is:
If Claude Haiku 5.5 and GPT-6 Luna cost the same, which one
should you actually use?
The first benchmark results make Haiku look like the stronger model.
Artificial Analysis currently scores Claude Haiku 5.5 at
43 on its Intelligence Index at max effort, versus
38 for GPT-6 Luna at max. Haiku is also substantially
faster while generating output in that test.
But there is a catch.
The two models may have the same price per token, but they do not
necessarily use the same number of tokens to finish the same job. And
once your prompts get long, they stop costing the same at all.
That makes this less of a benchmark story and more of an LLM
economics story.
Quick answer
For short-context workloads under standard API processing, Claude
Haiku 5.5 has the stronger capability ceiling at the same headline token
price as GPT-6 Luna.
But GPT-6 Luna can still be cheaper per completed
task, especially at lower reasoning levels and in workloads
where Haiku generates substantially more reasoning and output
tokens.
The difference becomes much larger with long context. Haiku’s pricing
jumps fivefold once a prompt exceeds 100K tokens. GPT-6 Luna keeps its
base pricing until 272K input tokens, after which its input price only
doubles and output price rises by 50%.
So the practical answer is:
Haiku 5.5 looks more attractive when you want the most
intelligence you can get from a very cheap model. Luna looks more
attractive when predictable cost, token efficiency, or long-context
economics matter more.
Claude
Haiku 5.5 and GPT-6 Luna start at exactly the same price
At standard pricing and short context, the two models are almost
perfectly aligned:
| Price per 1M tokens | Claude Haiku 5.5 | GPT-6 Luna |
|---|---|---|
| Input | $0.10 | $0.10 |
| Cached input / cache read | $0.01 | $0.01 |
| Cache write | $0.125 | $0.125 |
| Output | $0.50 | $0.50 |
For Haiku 5.5, those prices apply to prompts of up to 100K tokens.
For Luna, standard short-context pricing applies up to 272K input
tokens.
This is unusually direct competitive pricing.
Simon Willison pointed this out immediately after the Haiku launch:
Anthropic has effectively brought Haiku down from a model that
previously cost ten times as much as Luna to a direct competitor at the
same short-context price.
Latent Space framed the release similarly, calling Haiku 5.5 a model
that is effectively competing with GPT-6 Luna at the same pricing
tier.
But $0.10/$0.50 is only the sticker price.
On benchmarks,
Haiku 5.5 has the higher ceiling
The independent numbers from Artificial Analysis are the strongest
argument for Haiku right now.
At maximum reasoning effort, Claude Haiku 5.5 scores
43 on the Artificial Analysis Intelligence Index. GPT-6
Luna at max scores 38.
Haiku also reaches around 240 output tokens per
second in the evaluation, compared with roughly 126
tokens per second for Luna Max.
That is an impressive result for a model priced at ten cents per
million input tokens.
It also explains why Haiku 5.5 is attracting so much attention. This
is no longer just a tiny model you use for classification because it is
cheap. Anthropic explicitly positions it for high-volume work, real-time
applications, and as a subagent alongside stronger Claude models.
But the benchmark breakdown also gives us a reason not to declare an
easy winner.
On Artificial Analysis’ Terminal-Bench 4.0, Haiku 5.5 Max scores
32.8%, versus 12.6% for Luna Max. On
AutomationBench-AA, however, the relationship reverses: Luna Max scores
53.2%, while Haiku Max scores
35.4%.
In other words, “more intelligent” does not mean “better at every
workload.”
And once cost enters the equation, the picture gets even more
interesting.
Same
token price does not mean the same cost per task
Artificial Analysis currently estimates its average Intelligence
Index task at approximately $0.21 for Haiku 5.5 Max and
$0.07 for GPT-6 Luna Max.
That’s a threefold difference despite identical headline input and
output prices.
Why?
Token consumption.
Latent Space highlighted this as perhaps the biggest caveat in the
launch data. In the Artificial Analysis testing it reviewed, Haiku 5.5
at max effort generated roughly 162K output tokens per Index
task, compared with about 50K for GPT-6 Luna
Max. Moving Haiku from xhigh to max delivered only a small
Intelligence Index gain while consuming considerably more tokens.
There’s an even cleaner comparison further down the curve.
Haiku 5.5 at high effort and GPT-6 Luna at
max both score 38 on the Artificial
Analysis Intelligence Index. Their estimated costs per task are
approximately $0.08 and $0.07, respectively.
That is a much more useful comparison than simply saying:
Both APIs cost $0.10 in and $0.50 out.
For production applications, the number that matters is closer
to:
cost per accepted task = total model + tool + retry cost /
successful results
A model that is cheaper per token can lose if it retries more. A more
capable model can lose if it spends three times as many reasoning tokens
solving routine problems.
You need both sides of the equation.
Simon
Willison spotted another hidden cost: tokenization
Simon Willison found one more wrinkle.
Using his Claude Token Counter, he observed that the same long prompt
produced roughly 1.25× as many tokens with Haiku 5.5 as with
Haiku 4.5. Anthropic itself notes that Haiku 5.5 uses an
updated tokenizer and can use somewhat more tokens per task than its
predecessor.
This does not mean Haiku 5.5 is secretly expensive. Its headline
price is dramatically lower than Haiku 4.5.
It does mean that comparing LLMs solely on “dollars per million
tokens” is increasingly inadequate.
Different tokenizers, different reasoning budgets, different output
lengths, retries, tool calls, caching behavior and context pricing can
all change the final bill.
Long
context is where Luna’s pricing advantage becomes obvious
The most important pricing difference appears after 100K
tokens.
Claude Haiku 5.5 pricing jumps from:
$0.10 → $0.50 per million input tokens
and:
$0.50 → $2.50 per million output tokens
when the prompt exceeds 100K tokens.
Cache reads similarly rise from $0.01 to $0.05, while cache writes
rise from $0.125 to $0.625.
GPT-6 Luna has a much later pricing threshold.
OpenAI charges its normal $0.10/$0.50 rates until a prompt exceeds
272K input tokens. Beyond that point, input and cache
prices double, while output rises by 50%, putting long-context Luna at
$0.20 input and $0.75 output per million tokens.
That creates a particularly important range between 100K and
272K input tokens.
In that range, Haiku input costs five times its base rate while Luna
is still at $0.10 per million tokens. Haiku output costs $2.50 per
million while Luna remains at $0.50.
For long documents, large repositories, long agent histories, or
systems that routinely resend large context windows, this can dominate
every other pricing consideration.
Simon Willison’s conclusion is hard to argue with: if your workload
stays below 100K, Haiku looks very competitive; once you move past that
threshold, Luna becomes much more attractive economically.
So is Haiku 5.5
actually better than GPT-6 Luna?
Sometimes.
If your definition of “better” means the highest intelligence
available from a tiny, inexpensive model, Haiku 5.5 currently
has the stronger case. Its maximum Artificial Analysis score is higher,
its output generation is faster in those tests, and its improvement over
the previous Haiku generation is enormous.
If “better” means the lowest predictable cost for high-volume
production traffic, Luna has a strong argument. Its reasoning
settings include a non-reasoning mode, Artificial Analysis measures
significantly lower costs at the low end, and its long-context pricing
is much more forgiving.
For coding subagents, short classification jobs, summarization,
routing or interactive experiences, Haiku deserves serious testing.
For large-scale extraction, repetitive automation, workloads with
100K+ prompts, or tasks where Luna already clears your quality
threshold, paying for Haiku’s higher capability ceiling may not improve
the economics.
The important phrase there is your quality
threshold.
Benchmarks can tell you where to start. They cannot tell you how many
of your own support tickets, code patches, document extractions or agent
runs will actually pass review.
The
small-model price war is becoming more interesting than the flagship
race
The biggest takeaway from Haiku 5.5 may not be that Anthropic beat
Luna on one leaderboard.
It’s that the $0.10 input / $0.50 output tier has suddenly
become extremely capable.
OpenAI launched GPT-6 Luna as its efficient model for focused,
high-volume tasks. Anthropic has now priced Haiku directly into the same
category while pushing its maximum benchmark performance considerably
higher.
For developers, this changes model architecture decisions.
You do not necessarily need to send every step of an agent workflow
to an expensive frontier model. A stronger model can plan the difficult
parts while cheap models handle retrieval, classification,
summarization, validation, context compaction or parallel sub-tasks.
At that point, shaving fractions of a cent from each subagent call
starts to matter.
And so does being able to switch models quickly.
Testing
Haiku, Luna and other models through LLMFly AI
This is also where a multi-model API becomes useful.
LLMFly AI lets developers access supported GPT, Claude,
Gemini and other leading models from one multi-model platform.
Supported clients can use the appropriate API format for each model
family, while teams manage model access and application keys in one
place. That makes A/B testing and model fallback considerably simpler
than maintaining a separate account and integration workflow for every
provider.
There is a cost advantage as well. LLMFly’s public pricing page
currently lists a 0.3× billing multiplier for its ChatGPT/Codex
group — 70% below the referenced official rates — and a 0.5× multiplier
for its Claude Code group. Exact models, routes and current
rates vary, so the live Model Plaza should always be checked before
calculating a production budget.
That distinction matters here. Provider list pricing tells you how
Anthropic and OpenAI position Haiku and Luna. Your actual production
decision should use the price of the route you are actually going to
call.
With LLMFly, you can check the currently available route and model ID
in Model Plaza, run a representative workload, inspect the usage record,
and compare the result with another model without rebuilding the entire
application around a second provider.
For a fair Haiku-versus-Luna test, keep the prompt, tool schema,
acceptance criteria and workload constant. Then record the reasoning
level, input tokens, output and reasoning tokens, latency, retries,
successful outputs and total cost.
Don’t optimize for the cheapest million tokens.
Optimize for the cheapest result you are actually willing to
ship.
Bottom line
Claude Haiku 5.5 is one of the most interesting small-model releases
of 2026.
At first glance, the story is simple: it matches GPT-6 Luna at
$0.10 input and $0.50 output per million tokens, while
reaching a higher maximum score on Artificial Analysis.
But the real comparison is more nuanced.
Haiku 5.5 has the higher capability ceiling. GPT-6 Luna often
has the lower cost per task.
Haiku is particularly compelling below 100K tokens when you want more
intelligence from a very inexpensive model. Luna becomes increasingly
attractive when token efficiency and long-context costs dominate the
workload.
That is why the best answer is not to choose a winner from a pricing
table.
Test both against the work your application actually performs,
measure cost per accepted result, and keep your model layer flexible
enough to change when the next price cut arrives.
In the current small-model price war, it probably won’t take
long.
Frequently asked questions
Is Claude Haiku 5.5
cheaper than GPT-6 Luna?
Not automatically. Their standard short-context list prices match at
$0.10 per million input tokens and $0.50 per million output tokens, but
GPT-6 Luna used fewer output tokens in the current Artificial Analysis
max-effort test. Haiku 5.5 also moves to a higher price tier once the
prompt exceeds 100K tokens.
Is Claude
Haiku 5.5 better than GPT-6 Luna for coding?
Haiku 5.5 leads Luna on Terminal-Bench 4.0 in the current Artificial
Analysis results, but one benchmark does not establish production
reliability. Test both on repository navigation, patch acceptance, tool
calls, test pass rates, latency and cost per accepted change.
Which model is
better for long-context workloads?
GPT-6 Luna currently has the stronger pricing case between 100K and
272K input tokens. Haiku 5.5 moves to $0.50 input and $2.50 output per
million tokens above 100K, while Luna keeps its $0.10/$0.50 standard
rate until the request crosses 272K input tokens.
Does the same
token price mean the same API bill?
No. Different tokenizers, reasoning budgets, output lengths, retries,
tools and cache-hit rates can produce different bills even when the
price per million tokens is identical.
How should
developers compare Haiku 5.5 and Luna?
Run the same production-shaped evaluation set through both models.
Record acceptance rate, input tokens, reasoning and output tokens, cache
usage, latency, retries, tool failures and total cost. Choose the model
with the lowest cost per accepted result, not the lowest sticker
price.
Leave a Reply