⚡ Quick verdict
- Best value → Meta Muse Ember. It ties Claude Haiku 5.5 at 36 on the Artificial Analysis Intelligence Index for about a third of the cost per task ($0.14 vs $0.41) — the cheapest frontier-level intelligence of November 2026.
- Best for automation → Muse Ember again. On AutomationBench-AA it scores 37.8, against 13.9 for Haiku 5.5 and 20.2 for GPT-6 Luna. It is not close.
- Best for coding → Claude Haiku 5.5. The honest counterweight: Haiku still wins FrontierCode 2.0, 34.9% against Ember's 27.2%. For code, Haiku keeps its crown.
- GPT-6 Luna → third place. It trails on the headline index (28 vs 36) and on automation (20.2 vs 37.8). Pick it only if you are already deep in OpenAI's ecosystem.
- Our pick for most people → Muse Ember. Haiku-level intelligence at a third of the cost, with the best automation scores of the three.
We may earn a commission if you buy through our links, at no extra cost to you.
On 9 November 2026, Meta launched Muse Ember — its smallest Muse model — and quietly set off the most interesting price war in AI this year. Ember is reported at $0.10 per million input tokens and $0.50 per million output tokens, with a 1-million-token context window and a 512,000-token output limit. Those numbers alone would make it cheap. What makes it remarkable is what it does at that price: it ties Claude Haiku 5.5 at 36 on the Artificial Analysis Intelligence Index, for roughly a third of Haiku's cost per task — $0.14 against $0.41 — and it beats GPT-6 Luna, OpenAI's small model, everywhere that matters.
Small models are the workhorses of the AI economy. They are what runs behind customer support chats, document processing, summarisation pipelines, coding assistants, and the millions of agentic workflows that call a model thousands of times a day. A flagship model gets the headlines; a small model gets the bill. When a new small model matches the previous leader's intelligence at a third of the cost, every team running volume has to rethink its stack. This comparison is about that rethink: where Ember genuinely wins, where Haiku still holds on, and where Luna fits — or does not.
This is a researched comparison, not a lab test. We have not run all three models through our own benchmark suite. Instead, our analysis draws on the published Artificial Analysis figures and the independent 3-way comparison coverage published on 9–10 November 2026 — the same sources you would check yourself — organised into honest verdicts. All benchmark figures below are dated to the November 2026 leaderboard updates; model pricing and availability change often, so check the official sites before you commit. For the flagship end of the same leaderboard, see our Claude Fable 5.2 vs GPT-6 Astra comparison, and our best AI model November 2026 roundup puts all six models in one place.
At a glance: Ember vs Haiku 5.5 vs Luna
#1 · Meta Muse Ember · Best value & automation
- AA Intelligence Index
- 36 — tied with Haiku 5.5
- Reported price
- $0.10 / $0.50 per million tokens
- Context window
- 1 million tokens; 512K output
- Cost per AA task
- $0.14
- Best for
- Automation, agents, high-volume work
#2 · Claude Haiku 5.5 · Best small-model coding
- AA Intelligence Index
- 36 — tied with Ember
- Reported price
- Check the official site
- Context window
- Check the official site for the current figure
- Cost per AA task
- $0.41
- Best for
- Coding, careful reasoning, Claude users
#3 · GPT-6 Luna · Best if you live in OpenAI
- AA Intelligence Index
- 28
- Reported price
- Check the official site
- Context window
- Check the official site for the current figure
- Cost per AA task
- Check the official site
- Best for
- Existing OpenAI integrations
The price-per-task table: the whole comparison in one place
Small models are bought on cost, so we lead with cost. Artificial Analysis publishes a "price per task" figure: what a model actually costs to complete one of its benchmark tasks. It is the closest thing the industry has to a real-world bill, because it captures what token prices alone hide — how many tokens a model burns to get to an answer.
| Metric | Muse Ember | Claude Haiku 5.5 | GPT-6 Luna |
|---|---|---|---|
| Artificial Analysis Intelligence Index | 36 | 36 | 28 |
| Cost per task | $0.14 | $0.41 | Check the official site |
| AutomationBench-AA | 37.8 | 13.9 | 20.2 |
| FrontierCode 2.0 | 27.2% | 34.9% | Check the official site |
| Reported token price | $0.10 / $0.50 per million | Check the official site | Check the official site |
| Context window | 1 million tokens | Check the official site | Check the official site |
| Max output | 512,000 tokens | Check the official site | Check the official site |
Read the table this way. Ember and Haiku 5.5 are tied on headline intelligence at 36. Ember costs about a third as much per task. Ember wins automation by a landslide — 37.8 against 13.9 and 20.2. Haiku wins the one coding benchmark we have, 34.9% against 27.2%. Luna is third on every number we have. The table does not crown one model; it draws the map. Ember is the value and automation pick. Haiku is the coding pick. Luna is the pick you make for ecosystem reasons, not benchmark reasons.
One caveat on the cost figures: the per-task numbers are Artificial Analysis figures at their standard effort settings, and the Luna per-task cost was not in the coverage we reviewed. The $0.14 vs $0.41 comparison between Ember and Haiku is the solid one; Luna's absence from that row is a gap in the published data, not a judgement. Check the official pricing pages for all three before you commit, because small-model prices move fast — this launch is itself a price cut in disguise, and competitors respond within weeks.
What Meta Muse Ember actually is
Muse Ember is the smallest model in Meta's Muse family, launched on 9 November 2026 alongside the flagship news of the same week. Meta framed the launch exactly as this comparison frames it: a three-way shootout against Claude Haiku 5.5 and GPT-6 Luna, the two small models it is designed to replace in high-volume workloads. The reported price — $0.10 per million input tokens and $0.50 per million output tokens — makes it one of the cheapest models in its class, and the 1-million-token context window with a 512,000-token output limit is generous for a small model: long documents in, long answers out.
The headline result is the tie. Ember matches Haiku 5.5 at 36 on the Artificial Analysis Intelligence Index, a score that was the small-model benchmark to beat. But the headline understates the story, because the Intelligence Index is a composite that averages many tasks. On AutomationBench-AA — the benchmark that measures what small models actually do all day, automating multi-step tasks — Ember does not tie Haiku. It more than doubles it: 37.8 against 13.9, with Luna in the middle at 20.2. That is the number that should decide this comparison for anyone building agents or automation, because it measures the job, not the average.
The honest counterweight is coding. Ember scores 27.2% on FrontierCode 2.0, against Haiku's 34.9%. Neither number is high — small models are not flagship coders — but the gap is real, and it is the one place where Haiku keeps a clear lead. If your small model writes code all day, that matters more than the automation score. We come back to this below, because it is the single most important nuance in the comparison.
One thing worth noting: unlike the Claude Fable 5.2 launch the same week, Ember was a trade-press story, not a social one. Social coverage in the 8–10 November window showed effectively no discussion of Ember, while Fable 5.2 drew 108 mentions. That does not make Ember less important — it makes it less hyped, which is often where the real value hides. The teams that will save the most money from this launch are not the ones posting about it; they are the ones quietly swapping the model behind their API calls.
What Claude Haiku 5.5 actually is
Claude Haiku 5.5 is Anthropic's small model — the fast, cheap member of the Claude family, and until 9 November the small model to beat. It ties Ember at 36 on the Intelligence Index, and it is the model that Claude users already know: same product, same API, same safety posture, same ecosystem as the flagship Fable 5.2. For teams already built on Claude, Haiku 5.5 is the default, and defaults have real value — no migration, no re-testing, no new vendor.
Haiku's case in this comparison rests on two things. First, coding: 34.9% on FrontierCode 2.0 against Ember's 27.2% is a genuine lead, and for a coding assistant that gap shows up every day as the difference between a fix that works first time and one that needs a second attempt. Second, trust: Anthropic's small model has been in production use longer, with more published evaluations and more real-world mileage. Ember is days old. Haiku is proven.
But Haiku's problem is the price. At $0.41 per Artificial Analysis task against Ember's $0.14, Haiku costs roughly three times as much for the same headline intelligence. For a team running a million tasks a month, that is the difference between about $410,000 and about $140,000 — a $270,000 gap that buys a lot of engineering time, a lot of second attempts, or a lot of human review. Haiku has to earn that premium on every task, and on most tasks it cannot, because the tasks are not hard enough for its coding edge to matter.
What GPT-6 Luna actually is
GPT-6 Luna is OpenAI's small model — the fast, cheap member of the GPT-6 family, designed for the same high-volume workloads as Ember and Haiku. It scores 28 on the Artificial Analysis Intelligence Index, eight points behind both rivals, and 20.2 on AutomationBench-AA, ahead of Haiku's 13.9 but far behind Ember's 37.8. On the numbers we have, Luna is third.
Third place on benchmarks does not mean a bad model. A score of 28 on the Intelligence Index is still a capable small model, and for simple tasks — classification, extraction, short summaries, straightforward Q&A — the difference between 28 and 36 may be invisible. Luna's real case is the ecosystem: if your team already lives in OpenAI's products, if your prompts and evaluations are tuned for OpenAI models, if your compliance and procurement are already set up with OpenAI, then switching to Meta or Anthropic for a benchmark gain has a cost that no table captures. For a team starting fresh, though, the numbers point elsewhere — and that is the honest verdict.
We do not have Luna's per-task cost or its FrontierCode 2.0 score from the coverage we reviewed, so we will not invent them. Both are worth checking on the official pages if Luna is on your shortlist, because the per-task figure in particular could change the value judgement. What we can say is that Luna would need a dramatically lower price to compete with Ember on value, and the published numbers give no reason to expect one.
Benchmarks: where each model wins, and by how much
The headline index ties Ember and Haiku at 36, with Luna at 28. The detailed benchmarks tell a more interesting story — one where each model wins somewhere, and the winner depends entirely on the job.
Artificial Analysis Intelligence Index: 36 vs 36 vs 28
The Intelligence Index is a composite of many tasks, and it is the single most-cited independent score in the industry. Ember and Haiku both score 36; Luna scores 28. The tie between Ember and Haiku is the story of this launch: Meta's cheapest model matches Anthropic's established small model on the headline number. The 8-point gap to Luna is large by the standards of this index — small models usually move by one or two points per generation. For context, the flagship models score far higher — Claude Fable 5.2 leads the field at 65 — but nobody buys a small model to beat a flagship. They buy it to do routine work cheaply, and on that measure 36 is a strong score. The composite hides as much as it reveals, though: it averages away the tasks where the models differ most. The next two benchmarks are where the real differences live.
AutomationBench-AA: 37.8 vs 13.9 vs 20.2
AutomationBench-AA measures what small models actually do all day: multi-step automation tasks — following instructions across tools, handling edge cases, completing workflows without human rescue. Ember scores 37.8. Haiku scores 13.9. Luna scores 20.2. Ember does not just win this benchmark; it laps the field, scoring nearly three times Haiku and nearly twice Luna.
This is the most important number in the comparison for anyone building agents or automation, because it measures the job rather than the average. A small model that scores 37.8 on automation will complete workflows that a 13.9 model abandons halfway, and every abandoned workflow is either a human rescue — expensive — or a silent failure — worse. If you are building customer support automation, document processing pipelines, or any agentic workflow where the model runs unsupervised, the AutomationBench gap is worth more than the headline tie. It is also the number that makes Ember's launch a genuine event rather than a price cut: Meta did not just make a cheaper Haiku, it made a better automation model.
One note of caution: AutomationBench-AA is a new benchmark, and new benchmarks deserve scrutiny. The gap is so large that it is unlikely to be noise, but it is worth watching whether independent testers reproduce it on real-world workflows over the coming weeks. If the gap holds — and the 9–10 November coverage from five separate outlets all points the same way — then Ember is the default small model for automation until someone beats it.
FrontierCode 2.0: 34.9 vs 27.2 — Haiku's counterweight
Here is the honest counterweight, and it matters. On FrontierCode 2.0, the hard coding benchmark, Haiku 5.5 scores 34.9% against Ember's 27.2%. That is a 7.7-point lead for Haiku, and it is the one place where Haiku keeps a clear advantage. Neither number is high — small models are not flagship coders, and both are far behind Claude Fable 5.2's 49.1% — but the gap between them is real, and it shows up in daily work as the difference between a fix that works first time and one that needs a second attempt.
If your small model writes code all day — a coding assistant, a code-review bot, an agent that fixes bugs — Haiku 5.5 is still the better pick, and the AutomationBench gap does not outweigh it. Code is the one workload where Haiku earns its premium. For every other workload, Ember's automation lead and its third-of-the-cost price make it the better choice. This is not a contradiction; it is a segmentation. Buy Haiku for code, Ember for everything else, and do not let a single benchmark decide a purchase that spans many jobs.
Automation and agentic work: Ember's home turf
Small models are the engines of the agentic economy. They are what runs behind the scenes when a support ticket gets triaged, a document gets processed, a workflow gets executed, or an agent chains ten tool calls together to finish a job. The model that wins automation wins the volume — and volume is where the money is.
Ember's 37.8 on AutomationBench-AA is the strongest automation score of any small model in the November 2026 leaderboard, and it is not close. For teams building agents, the practical implications are straightforward: fewer failed workflows, fewer human rescues, and longer unsupervised runs. An agent that completes 37.8% of its benchmark automation tasks, against 13.9% for the previous leader, is an agent you can trust with more work — and trust is the constraint on automation, not capability. Most teams automate only the workflows they are confident the model will finish. A model that finishes more workflows expands the set of things you can automate at all.
The cost side compounds this. An agentic workflow may call the model dozens of times, and at $0.14 per task against $0.41, Ember lets you run three times as many agent steps for the same money — or the same steps for a third of the money. For a team running a million agent tasks a month, that is the difference between about $140,000 and about $410,000. The savings fund the migration, the re-testing, and the human review, with money left over. Our best AI agents 2026 guide covers the tools that make this kind of multi-model agent setup practical.
Coding: Haiku keeps its crown
If you write code with a small model, Haiku 5.5 is still the pick. The FrontierCode 2.0 gap — 34.9% against 27.2% — is the largest single-model gap in this comparison after automation, and it points the other way. For a coding assistant, a code-review bot, or an agent that fixes bugs, that gap is the difference between useful and frustrating.
That said, keep the numbers in perspective. Neither small model is a flagship coder: Fable 5.2 scores 49.1% on the same benchmark, and even that is below half. Small-model coding is for autocomplete, small fixes, boilerplate, and review — not for architecting systems. Within that scope, Haiku's lead is real but narrow in absolute terms: both models fail most hard coding tasks, and both succeed on the easy ones. The practical test is your own codebase: run both models on your hardest recent tickets for a week and count first-try fixes. The benchmark says Haiku will win; your tickets will tell you by how much, and whether the win is worth three times the price per task.
One more coding consideration: context. Ember's 1-million-token context window is a genuine advantage for code work — you can load an entire repository into one conversation. Haiku's context figure was not in the coverage we reviewed, so check the official documentation. If Haiku's window is smaller, Ember's context advantage may partly offset its coding-score disadvantage on large-repo work, where holding the whole codebase in memory matters as much as raw coding skill.
Context window and output: Ember's generous limits
Ember's 1-million-token context window and 512,000-token output limit are generous for a small model — flagship-grade context at a small-model price. A million tokens is an entire codebase, a shelf of documents, or months of chat history in a single conversation. A 512,000-token output limit means the model can write very long answers without being cut off: long reports, long code files, long-form content in one pass.
For the workloads small models actually run — summarising long documents, processing batches of tickets, working across a codebase — these limits remove a real constraint. Many small models force you to chunk documents or truncate conversations, and every chunk is a chance to lose context. Ember lets you skip the chunking. The honest caveat is cost: long contexts are expensive, because every token in the window costs money whether the model needs it or not. A million-token context is a powerful tool and an easy way to burn budget. Use it when the task genuinely needs it, not by default.
The cost question, worked out properly
Small models are bought on cost, so let us make the cost concrete. The per-task figures from Artificial Analysis are the honest unit: $0.14 for Ember, $0.41 for Haiku 5.5. Luna's per-task figure was not in the coverage we reviewed.
- 1,000 tasks a day: Ember costs about $140 a day; Haiku costs about $410. Monthly (22 working days): about $3,080 vs about $9,020.
- 10,000 tasks a day: Ember costs about $1,400 a day; Haiku costs about $4,100. Monthly: about $30,800 vs about $90,200.
- 100,000 tasks a day: Ember costs about $14,000 a day; Haiku costs about $41,000. Monthly: about $308,000 vs about $902,000.
- 1 million tasks a month: Ember costs about $140,000; Haiku costs about $410,000 — a gap of roughly $270,000 a month.
At 10,000 tasks a day, the monthly gap is roughly $59,400. That buys a lot of engineering time, a lot of second attempts with the cheaper model, or a lot of human review. The break-even question is simple: does Haiku's coding edge save you more than $59,400 a month in rework and errors? For a team whose small model writes code all day, possibly. For a team whose small model triages tickets, summarises documents, and runs automations, almost certainly not — and Ember's automation lead means it will also fail less often on those jobs.
There is also the token-price view. Ember is reported at $0.10 per million input tokens and $0.50 per million output tokens. Those are among the lowest prices in the industry for a model at this intelligence level, and they are the reason the per-task figure is so low. But token prices alone mislead: a model that burns three times as many tokens to reach the same answer costs three times as much, whatever its sticker price. The per-task figure captures that, which is why we lead with it. Check the official pricing pages for all three models, because small-model prices move fast — and this launch is itself a price cut that competitors will answer within weeks.
Who should pick which
- Pick Muse Ember if: you run automation or agentic workflows at volume; you want Haiku-level intelligence at a third of the cost; you need a 1-million-token context window for long documents or whole codebases; or you are building a new stack and can choose freely.
- Pick Claude Haiku 5.5 if: your small model writes code all day and the FrontierCode lead is worth the premium; your team already lives in Claude and switching costs are real; or you need a proven model with more production mileage behind it.
- Pick GPT-6 Luna if: your team already lives in OpenAI's ecosystem — prompts, evaluations, compliance, and procurement all set up there — and the cost of switching outweighs an 8-point intelligence gap. For a fresh build, the numbers point elsewhere.
- Pick Ember and Haiku together if: you run a mixed workload — Ember for automation and bulk work, Haiku for coding. Most agent frameworks let you assign different models to different steps, and that split captures both models' strengths at a blended cost well below Haiku-only.
- Pick neither yet if: your current small model handles your work fine. A new launch does not make last month's model bad, and the price war this launch starts will likely make all three models cheaper within months.
Everyday use: chat, writing, and support
Most small-model work is not benchmark work. It is chat support, summarisation, extraction, classification, and drafting — the routine jobs that run thousands of times a day behind products people use. For those jobs, the difference between Ember and Haiku is small, and the difference between either of them and Luna is modest.
For customer support, the deciding factors are reliability and cost, not headline intelligence. A support model needs to answer correctly, stay on script, and escalate when it is unsure — and it needs to do it thousands of times a day without the bill becoming the story. Ember's automation lead suggests it will handle multi-step support flows better, and its price means you can afford to run it on every ticket rather than reserving AI for the easy ones. Haiku's longer production history is a genuine comfort for support teams that cannot afford a bad week. Either is defensible; the tie-breaker is your volume. Above a few thousand tickets a day, Ember's price wins. Below it, Haiku's track record may matter more.
For summarisation and extraction — turning long documents into short answers, pulling structured data from messy text — all three models are capable, and the differences are hard to measure without testing on your own documents. Ember's 1-million-token window is the practical advantage here: feed it the whole document instead of chunking it, and the summary is better because the model saw everything at once. For short documents, any of the three will do, and the cheapest one wins.
For writing and drafting at small-model scale — first drafts, rewrites, short-form content — the gap between 36 and 28 on the index is unlikely to be visible to a reader. A small model is not where you go for your best prose; it is where you go for volume. Pick the cheapest model that clears your quality bar, test it on your own content, and spend the savings on human editing, which is where the quality actually comes from.
What to watch next
This comparison is a snapshot of 11 November 2026, and it will age fast. Three things are worth watching. First, the response: Anthropic and OpenAI both have small models to defend, and a launch this aggressive usually draws a price cut or a new model within weeks. Haiku's successor — or a repriced Haiku — could close the value gap from the other direction. Second, independent testing: the AutomationBench gap is so large that it deserves verification on real-world workflows, and the coming weeks will show whether it holds outside the benchmark. Third, the price war: Ember's $0.10/$0.50 pricing is a gauntlet thrown down, and competitive pressure tends to push small-model prices down across the board — good news for everyone running volume, whatever they choose.
None of that changes today's recommendation. Ember is the best small-model value of November 2026, Haiku is the best small-model coder, Luna is the ecosystem pick, and most buyers should run Ember by default and reach for Haiku when the task is code. But if you are reading this in December or later, check the current leaderboard before you decide — the November 2026 numbers may already be history.
How we compared them
Our method is simple and we state it plainly. We used the published Artificial Analysis Intelligence Index, price-per-task, AutomationBench-AA, and FrontierCode 2.0 figures as updated on 10 November 2026, and the reported pricing and specifications from the independent 3-way comparison coverage published on 9–10 November 2026. We did not run our own benchmark suite, and we did not test any of the three models hands-on for this comparison. Benchmarks are a useful guide, not a verdict on your work: the model that wins on AutomationBench may not win on your workflows, and the model that is cheapest per Artificial Analysis task may not be cheapest on your prompts. Where a figure comes from a single source — notably the $0.10/$0.50 pricing and the 1-million-token context and 512,000-token output figures — we have said so, and you should confirm it on the official site. Where a figure was not in the coverage we reviewed — Luna's per-task cost and FrontierCode score, Haiku's context window — we have said that too, rather than inventing a number. Model capabilities, prices, and availability change fast; this comparison is a snapshot of November 2026, and it will age.
The verdict
Meta Muse Ember is the most important small-model launch of 2026 — not because it is the smartest small model, but because it is the cheapest smart one. Tying Claude Haiku 5.5 at 36 on the Intelligence Index for a third of the cost per task would be enough on its own. Winning AutomationBench-AA by a landslide — 37.8 against 13.9 — makes it the default small model for the automation and agentic work that small models actually do.
But the honest verdict has a counterweight, and it is Haiku's coding lead: 34.9% on FrontierCode 2.0 against Ember's 27.2%. If your small model writes code all day, Haiku 5.5 is still the better pick, and its longer production history is worth something too. GPT-6 Luna, meanwhile, is third on every number we have — a capable model whose case is the OpenAI ecosystem, not the benchmarks.
Our recommendation: if you are building a new stack, start with Ember. If you run automation at volume, switch to Ember — the savings fund the migration. If your small model writes code, keep Haiku 5.5, or split the work: Ember for automation and bulk, Haiku for code. And if you are choosing today and can only pick one, pick Ember — then revisit in a few months, when the price war this launch has started will have changed the maths again.
Frequently asked questions
Is Muse Ember better than Claude Haiku 5.5?
It depends on the job. Ember ties Haiku 5.5 at 36 on the Artificial Analysis Intelligence Index, and it wins AutomationBench-AA by a landslide — 37.8 against 13.9 — at about a third of the cost per task ($0.14 vs $0.41). But Haiku still wins on coding: 34.9% on FrontierCode 2.0 against Ember's 27.2%. For automation and bulk work, Ember is better. For code, Haiku is still better.
How much does Muse Ember cost per task?
Artificial Analysis reports $0.14 per task for Muse Ember, against $0.41 for Claude Haiku 5.5 — roughly a third of the cost for the same headline intelligence score. The token price is reported at $0.10 per million input tokens and $0.50 per million output tokens. Both figures are snapshots from the November 2026 launch coverage, so check the official pricing pages before you commit.
What is AutomationBench-AA?
It is a benchmark that measures multi-step automation tasks — following instructions across tools, handling edge cases, and completing workflows without human rescue. As of November 2026, Muse Ember scores 37.8 on it, against 20.2 for GPT-6 Luna and 13.9 for Claude Haiku 5.5. It matters because automation is the job small models actually do all day, and a model that finishes more workflows is a model you can trust with more work.
Is GPT-6 Luna still worth using?
Only if you are already deep in OpenAI's ecosystem. Luna scores 28 on the Artificial Analysis Intelligence Index, eight points behind Ember and Haiku at 36, and 20.2 on AutomationBench-AA against Ember's 37.8. It is still a capable small model for simple tasks, and if your prompts, evaluations, and procurement are all set up with OpenAI, switching has a real cost. For a fresh build, the numbers point to Ember or Haiku instead.
Which small model is best for coding?
Claude Haiku 5.5. It scores 34.9% on FrontierCode 2.0, against 27.2% for Muse Ember — a genuine lead that shows up as more first-try fixes in daily coding work. Neither small model is a flagship coder, though: Claude Fable 5.2 scores 49.1% on the same benchmark. For the hardest coding work, use a flagship model; for everyday coding assistance, Haiku is the best small model of the three.
Which small model should I choose in November 2026?
Choose Muse Ember if you run automation or high-volume work and want Haiku-level intelligence at a third of the cost. Choose Claude Haiku 5.5 if your small model writes code all day. Choose GPT-6 Luna only if you already live in OpenAI's ecosystem. If you run a mixed workload, use both Ember and Haiku: Ember for automation and bulk work, Haiku for code.