TL;DR
- GPT-6 Luna (OpenAI, Sept 22) costs $0.10/$0.50 per million input/output tokens. Muse Spark 1.3 Contributor (Meta, Sept 2) costs $0.10/$0.20 — 60% cheaper on outputs.
- The Contributor tier trades that discount for permission to train future Meta models on your prompts and completions. Meta publishes no retention periods, no human review policy, and no consent revocation mechanism.
- On the Artificial Analysis Intelligence Index (independent), Muse Spark 1.3 scores 61 vs Luna’s 37 — a real quality gap that holds across DeepSWE and OSWorld too.
- For a team running 10,000 PR code reviews per month: Contributor costs $6, Luna costs $9, Muse Spark Standard costs $92.50.
Who should care: Teams building agentic pipelines that need cheap tokens at scale and are deciding whether the Contributor data deal is worth $3/month in savings.
Verdict: Use Luna for most production workloads — the price gap over Contributor is small and you keep your prompts. Use Contributor only for throwaway or public-domain code where no proprietary content flows through.
Two Sub-Cent Output Models From Rival Labs
Two cheap, long-context models now sit at the bottom of the agentic token market. Meta’s Muse Spark 1.3 Contributor landed on September 2, 2026 at $0.10/$0.20 per million input/output tokens. OpenAI’s GPT-6 Luna followed on September 22 at $0.10/$0.50. Both offer context windows just over one million tokens. Both handle multimodal inputs. The pricing at first glance looks similar — until you read what each deal actually requires.
Muse Spark 1.3 Contributor is not a separate model. It is the same model as the $1.25/$4.25 Standard tier, accessed under a different contract: “Heavily discounted token pricing in exchange for permission to use your prompts and completions to train future Meta models.” The Standard tier explicitly opts out of training use. The Contributor tier opts in, permanently, on every request.
Luna carries no training clause. OpenAI’s terms for the direct API specify that prompt data is not used to improve models unless you explicitly opt into that programme. The two models cost the same per input token. On output, Contributor is 60% cheaper: $0.20/M vs $0.50/M.
Benchmark Scores — A 24-Point Quality Gap
Muse Spark 1.3 is the stronger model. According to Artificial Analysis’s independent Intelligence Index, the customer-available xhigh tier scores 61. GPT-6 Luna at max reasoning scores 37. Artificial Analysis also puts the cost per Intelligence task at $0.55 for Muse Spark Standard and $0.07 for Luna — a gap that reflects both the quality difference and the price difference.
The vendor-run benchmarks point the same direction. On DeepSWE v1.1 (a software-engineering evaluation), OpenAI reports Luna at 66.6% (max reasoning). Meta reports Muse Spark 1.3 Max at 75.4% — but that Max tier is not yet generally available; the xhigh customer tier is lower. On OSWorld 2.0 (computer-use agent tasks), benchlm.ai aggregates Luna at 52.7% and Muse Spark 1.3 at 66.9%, both vendor-reported.
One independent data point for Luna: ARC Prize ran ARC-AGI-2 evaluations and published Luna at 59.3% — a number OpenAI did not run. For Muse Spark, Artificial Analysis is the only source of non-vendor scores currently available.
| Model | AA Intel. Index | DeepSWE v1.1 | ARC-AGI-2 | AA cost/task | Data safe? |
|---|---|---|---|---|---|
| Muse Spark 1.3 Standard | 61 (independent) | 75.4% (vendor, Max tier) | n/a | $0.55 | ✓ No training |
| Muse Spark 1.3 Contributor | 61 (same model) | 75.4% (vendor, Max tier) | n/a | ~$0.03* | ✗ Trains Meta |
| GPT-6 Luna | 37 (independent) | 66.6% (vendor) | 59.3% (ARC Prize) | $0.07 | ✓ No training |
*Contributor cost per AA task derived from Standard’s $0.55 scaled by the ~18x blended price ratio ($0.15/M blended Contributor vs $2.75/M blended Standard). AA = Artificial Analysis Intelligence Index v4.1.1. DeepSWE and OSWorld scores are vendor-reported; AA scores are independent.
What the Contributor Tier Actually Requires
Meta’s API documentation states: “Prompts and outputs may be used to improve Meta’s products.” The pricing page is clearer: “Heavily discounted token pricing in exchange for permission to use your prompts and completions to train future Meta models.”
What Meta does not publish: retention periods for Contributor prompts, whether human reviewers access completions, whether attachments and multimodal inputs are included in training data, or any consent revocation mechanism after submission. A review by Codersera summarises the gap: Meta does not specify retention periods, human review practices, deletion rights, or GDPR/EU compliance details.
This matters for agentic pipelines more than for simple chat. An agent calling the Contributor endpoint with access to a codebase, a customer database, or internal documentation sends proprietary context with every tool call. The Standard tier costs $4.25/M output to avoid that. Luna costs $0.50/M and does not require the deal at all.
Cost Per Task — The Math
The task: a team’s CI pipeline runs automated code review on every PR using an agentic call. Each review sends 4,000 input tokens (system prompt plus diff) and receives 1,000 output tokens (structured comments). At 10,000 reviews per month — roughly 50 engineers merging eight PRs each — the numbers break down as follows. Monthly token volume: 40 million input tokens, 10 million output tokens.
| Model & tier | Input cost | Output cost | Monthly total | Data policy |
|---|---|---|---|---|
| Muse Spark 1.3 Contributor | 40 × $0.10 = $4.00 | 10 × $0.20 = $2.00 | $6.00 | Trains Meta |
| GPT-6 Luna (OpenAI direct) | 40 × $0.10 = $4.00 | 10 × $0.50 = $5.00 | $9.00 | No training |
| GPT-6 Luna (Azure EU westeurope) | 40 × $0.12 = $4.80 | 10 × $0.60 = $6.00 | $10.80 | No training, EU residency |
| Muse Spark 1.3 Standard | 40 × $1.25 = $50.00 | 10 × $4.25 = $42.50 | $92.50 | No training |
Prices from digitalapplied.com (Luna), codersera.com (Muse Spark), and requesty.ai (Azure EU). Prices in USD per million tokens. Calculation: (tokens / 1,000,000) × rate.
At this scale, choosing Contributor over Luna saves $3/month — a 33% reduction in model cost. Whether that trade is worth it depends entirely on what flows through the pipeline. If your PR diffs contain internal business logic, credentials in environment variables, or unreleased product code, the saving is not worth it. If your pipeline reviews only open-source repositories, the math is different.
For Swiss & EU teams
GPT-6 Luna is available via Microsoft Azure in the West Europe region at $0.12/$0.60 per million tokens — a 20% premium over the direct API. Requesty’s listing of that deployment specifies zero data retention and no training use, governed by the Microsoft Privacy Statement. Prompts do not leave the EU datacenter.
On Amazon Bedrock, both Sol and Luna are available across EU regions including eu-central-1 (Frankfurt), but only via Global CRIS — cross-region inference that may route compute outside the EU. There is no EU-only processing option. Teams with strict data localisation requirements should confirm with their DPO before using Bedrock in EU regions.
Muse Spark 1.3 Contributor is effectively off the table for EU teams. Meta publishes no GDPR compliance details, no data processing agreement for the Contributor tier, and no EU data residency guarantee. Under GDPR Article 28, sending personal data to a sub-processor for model training without documented safeguards is not compliant. Standard tier avoids the training clause but still has no documented EU processing guarantee. For context on the broader regulatory implications, see EU AI Act’s Agentic AI Gap: What Engineering Teams Must Do.
Verdict
Use Muse Spark 1.3 Contributor if your pipeline processes only open-source or public-domain content, no proprietary code or customer data flows through, and you need the quality headroom the 24-point AA advantage delivers. Document the data-sharing scope before you deploy.
Use GPT-6 Luna (OpenAI direct) for most production agentic workloads. Sub-cent output tokens, no data deal, 59.3% on ARC-AGI-2 (independent). The 24-point quality gap matters for complex multi-step reasoning; for classification, summarisation, or structured extraction, Luna is sufficient.
Use GPT-6 Luna (Azure EU) if you are EU- or Swiss-based and need verifiable EU data residency. The extra $1.80/month over Luna direct for 10,000 PR reviews is a fair price for a documented compliance position.
Skip Muse Spark 1.3 Standard at agentic scale. $92.50/month for the same model the Contributor tier delivers at $6 is hard to justify unless procurement policy blocks the Contributor tier and Luna is also off the table.
Artificial Analysis puts Luna’s cost per Intelligence task at $0.07. The Contributor tier’s implied figure is roughly $0.03. A $0.04/task difference only becomes compelling once you are running millions of tasks per month. Most teams are not there yet. Start with Luna; revisit when the bill is large enough to warrant a data trade-off conversation with legal. For a framework to make that switch decision, see the vortx.ch guide How to Evaluate a New Frontier Model Before Switching.
Further Reading
- GPT-6 Sol and Luna: API Prices, Benchmarks and Trade-offs — Digital Applied — the most complete pricing and benchmark summary at GPT-6 Luna’s September 22 launch.
- Muse Spark 1.3: Pricing, Specs, and the Contributor Tier — Codersera — full Artificial Analysis breakdown and the documented gaps in Meta’s data policy.
- GPT-6 Luna West Europe via Requesty — exact Azure EU pricing, zero-retention terms, and routing guarantee for EU-based deployments.
Your turn: Has your team run any workloads through the Muse Spark 1.3 Contributor tier, and did the data-sharing clause require a conversation with legal or procurement? Reply to our newsletter or send us a note — we feature the best answers in the Friday Scorecard.
How this article was made: AI researched and wrote this article under the Adrian persona, using the sources linked above, and it was published automatically without a human edit. Editorial guidelines are set by Adi. Spotted an error? Tell us and we will correct it.