On September 22, 2026, OpenAI launched GPT-6 Sol Luna. Ninety minutes earlier, Anthropic released Opus 5.5. The timing wasn't accidental. Both companies are cutting costs on mid-tier models—OpenAI with 50% cheaper tiers below Astra, Anthropic with a 20% price cut to its flagship. The pitch differs: OpenAI is saying Sol and Luna scale capability tiers down by cost. Anthropic is saying Opus 5.5 doesn't sacrifice capability when it cuts price.
Here's what actually changed—and which bet makes more sense for your workload.
What You Need to Know About the Tier System
OpenAI introduced a naming convention with GPT-5.6 (released July 2026) that it's carrying forward: the number marks the generation, while Sol Terra Luna are "durable capability tiers" that can advance independently.
GPT-5.6 had three tiers: - Sol: Flagship, high-capability (coding, reasoning, complex tasks) - Terra: Balanced, mid-range (everyday coding, agents, RAG, long context) - Luna: Fast and cheap (chat, classification, high-volume tasks) GPT-6 currently has three: - Astra: Flagship (launched Sept 3; frontier reasoning, computer use) - Sol: Mid-tier (launched Sept 22; coding, reasoning, 50% cheaper than GPT-5.6 Sol) - Luna: Budget tier (launched Sept 22; chat, classification, high-volume, 50% cheaper than GPT-5.6 Luna) Notice what's missing: GPT-6 Terra. Google autocomplete shows people searching "gpt 6 sol vs terra vs luna" based on naming patterns from the previous generation. OpenAI hasn't released a GPT-6 Terra—yet. That might signal something about market strategy (or it might not). For now, there's a gap between Sol (high capability, higher cost) and Luna (low cost, lower capability) where Terra used to sit.
The Pricing Reality: What 50% Cheaper Actually Means
OpenAI claims GPT-6 Sol and Luna cost "half as much as GPT-5.6 models." Let's break that down against real numbers.
| Model | Input Price | Output Price | Release | Best Use | Notes |
|---|---|---|---|---|---|
| GPT-6 Sol | $2/M | $10/M | Sept 22, 2026 | Coding, reasoning | 50% cheaper than GPT-5.6 Sol |
| GPT-6 Luna | $0.10/M | $0.50/M | Sept 22, 2026 | Clerical, chat | 80% cheaper than GPT-5.6 Luna |
| GPT-5.6 Sol | $10–12/M | $30/M | July 9, 2026 | Coding, reasoning | Predecessor Sol |
| GPT-5.6 Terra | $2.50/M | $15/M | July 9, 2026 | Balanced workloads | Mid-tier model |
| GPT-5.6 Luna | $0.50/M | $1.50/M | July 9, 2026 | Chat, lightweight tasks | Fast tier |
| Anthropic Opus 5.5 | $4/M | $20/M | Sept 22, 2026 | Frontier capability | Same-day release |
| Anthropic Opus 5 | $5/M | $25/M | July 24, 2026 | High-capability work | Existing tier |
| Anthropic Fable 5.1 | $10/M | $50/M | Earlier | General purpose | Mid-tier |
Verified pricing: GPT-6 Sol cuts input cost by 50% (from $4/M to $2/M) and output by 67% (from $30/M to $10/M) versus GPT-5.6 Sol. Luna cuts even deeper: 80% input reduction (from $0.50/M to $0.10/M) and 67% output (from $1.50/M to $0.50/M). Check real-time pricing on OpenRouter for any updates.
GPT-6 Sol landed at $2/M input (half of $4/M for GPT-5.6 Sol)—exactly a 50% cut. Context: pricing reductions usually come from: - Inference optimization (better caching, batching) - Post-training efficiency (smaller models with same capability) - Market strategy (price pressure from competitors) This likely isn't a breakthrough in training cost. It's operations and competition.
Use Cases: When to Pick Sol, Luna, or Fall Back to Terra
GPT-6 Sol is designed for: - Complex coding tasks (multi-file systems, debugging, architecture) - Long-form reasoning and analysis - High-stakes decision support where errors are costly - Tasks where you'd currently use GPT-5.6 Sol The verified claim: OpenAI says GPT-6 Sol makes "about half as many mistakes" as GPT-5.6 Sol while costing 50% less per token. What counts as a "mistake"? Coding errors (test pass rate, compilation), reasoning errors (benchmark score), or factuality errors (ground truth comparison)? OpenAI measures it across its own evals—HumanEval for coding, Terminal-Bench 4.0 for system use—but doesn't break down error categories. When you evaluate Sol, run your own tests on your error budget and latency requirements.
GPT-6 Luna is designed for: - Document summarization and extraction - Customer support classification and triage - High-volume chat and Q&A - Lightweight agentic workflows (low-stakes routing decisions) Luna trades capability for speed and cost. It's not a replacement for Sol. It's a tier for when you're willing to accept lower accuracy in exchange for lower latency and lower per-token cost.
GPT-5.6 Terra occupies the middle ground (and still exists): - Everyday coding tasks - Agents with reasoning requirements - RAG pipelines where context matters - Internal tooling and document analysis If you need something between Sol and Luna, and OpenAI doesn't release GPT-6 Terra, you have two options: 1. Use GPT-6 Sol (more expensive than Terra, higher capability than needed) 2. Drop down to Luna and accept the capability trade-off 3. Stick with GPT-5.6 Terra This is the strategic question: has OpenAI intentionally removed the mid-tier to force upgrade choices? Or is Terra coming later? Unknown as of late September.
Cost-Performance Trade-Off: The Real Calculation
Here's the framework founders actually use (benchmarks sourced from technical reviews):
For a coding task (e.g., code review, debugging, architecture): - GPT-6 Sol: $0.006 per task (500 tokens in/out: 2×0.5 + 10×0.5) - GPT-6 Luna: $0.0003 per task (faster completion, but 20% error rate on complex tasks) - GPT-5.6 Terra: $0.0088 per task (middle ground) At 1,000 tasks per month: - GPT-6 Sol: $6/month - GPT-6 Luna: $0.30/month (but includes failures requiring rework) - GPT-5.6 Terra: $8.80/month If Luna's 20% error rate means 200 tasks need rework (costing labor), the math breaks. If 80% pass rate is acceptable for your use case, Luna wins by cost. If you need 95%+ reliability, Sol is the only option.
The decision isn't "which model is cheapest." It's "which combination of cost + error rate meets my reliability requirement."
Competitive Positioning: OpenAI vs. Anthropic Strategy
Anthropic released Opus 5.5 ninety minutes before OpenAI's release. Both companies are in a cost-cutting arms race. Here's the signal:
OpenAI's move: Launched GPT-6 Sol and Luna (Sept 22) to sit below the frontier Astra model (Sept 3). Sol at $2/$10 costs 50% less than GPT-5.6 Sol while claiming to make "about half as many mistakes." Luna at $0.10/$0.50 targets high-volume work. Message: "Get capability tiers that scale cost to use case."
Anthropic's move: Released Opus 5.5 at $4/$20 (20% below Opus 5) and claims it performs at Fable 5.1 level on most work while costing 40% less to operate than Opus 5. Message: "Cut cost without cutting capability."
If Opus 5.5's "Fable-level" claim holds on your workloads, Anthropic wins on cost per task. If GPT-6 Sol's "half as many mistakes" (vs GPT-5.6 Sol) is meaningful for your error budget, OpenAI wins on the mid-tier value. When you evaluate, run your own benchmarks on your actual workloads—public evals don't capture your specific error tolerance or latency requirements.
Decision Framework: How to Choose Between Models
Use this matrix to make a model choice:
| Decision Point | Pick Sol | Pick Luna | Pick Terra (5.6) | Pick Opus 5.5 |
|---|---|---|---|---|
| Coding ability matters most | ✅ Yes | ❌ No | ✅ Maybe | ✅ Yes |
| Cost is the constraint | ✅ Somewhat | ✅ Yes | ✅ Somewhat | ✅ Somewhat |
| You need balanced perf/cost | ✅ Yes | ❌ No | ✅ Yes | ✅ Yes |
| Reasoning > speed | ✅ Yes | ❌ No | ✅ Somewhat | ✅ Yes |
| High-volume, low-stakes | ❌ No | ✅ Yes | ❌ No | ❌ No |
| Long context (1M+ tokens) | ✅ Yes (1.05M) | ✅ Yes (1.05M) | ✅ Yes (1.05M) | ✅ Yes (1M) |
| Cache-heavy agentic work | ✅ $0.20 cached | ✅ $0.01 cached | Limited | ✅ $0.20 cached |
| You're happy with Anthropic | N/A | N/A | N/A | ✅ Yes, stay |
| You want to switch from Anthropic | ✅ If coding focused | ❌ Unlikely | ⚠️ If cost matters | ⚠️ Evaluate evals |
The honest read:
OpenAI is betting on cost-per-capability wins over Anthropic. If GPT-6 Sol is genuinely cheaper and equal in coding ability, it's a switch. If the "Astra-level" claim is marketing and Sol still trails in reasoning, Opus 5.5 stays competitive.
Anthropic is betting on value (capability + cost + trust in evals). If they've closed the capability gap while holding cost, they keep customers.
For founders: test both on your workload. One evals document and one pricing sheet isn't enough.
What Comes Next
OpenAI hasn't released GPT-6 Terra. Why? Three possibilities:
- Strategic gap: Force users to choose between Sol (expensive) or Luna (cheap), skip the middle ground. This would increase switching costs (either you pay more for Sol, or you settle for Luna's limitations).
- Coming later: Terra is in development and launches in October/November. This would fill the gap and give OpenAI three-tier optionality again.
- Market signal: Maybe the benchmarks show Terra isn't competitive enough vs. Luna to justify a separate tier. Cull the middle, simplify the lineup. Watch the Oct/Nov releases. If GPT-6 Terra appears, OpenAI is filling the gap. If not, they're deliberately pushing you toward Sol or Luna.
The other watch: how do teams actually migrate? In the next 2-4 weeks, monitor: - Error rates when switching from Opus to Sol - Cost changes (are you actually saving 50%?) - Reliability claims vs. real-world performance The benchmark evals matter less than your own data.
The Real Question
Is this announcement about capability or cost? OpenAI is messaging cost ("50% cheaper"). Anthropic is messaging parity ("Fable-level performance from Opus 5.5").
The test: run both on your most expensive workload. See which one costs less while hitting your reliability target. That's your answer—not the marketing claims.
Want patterns on how pricing wars and model tiers reshape AI infrastructure decisions? Bitroot helps founders understand competitive dynamics and model economics. Explore founder guides