GPT-5.6 Sol vs Terra vs Luna: Which Tier Do You Need?

Most people will pick Sol because it's the flagship. That's a mistake for 80% of use cases. Luna scores 75 on the Coding Agent Index and handles the majority of real work. After the July 30 price cut, Luna costs just $0.20/$1.20 per million tokens. That's an 80% drop from launch pricing.

Here's how all three stack up, what the benchmarks mean in practice, and which setup saves the most money without losing quality.

Quick verdict: Use Luna as your default. Switch to Sol for hard problems. Skip Terra.

Update (July 30, 2026): OpenAI announced major price cuts. Luna dropped 80% to $0.20/$1.20. Terra dropped 20% to $2/$12. Sol gets a new Fast mode at 2.5x speed for 2x the price. All pricing below reflects the new rates.

The Three Tiers at a Glance

OpenAI launched GPT-5.6 on July 9, 2026 with these three tiers sharing the same architecture but differing in size, speed, and price. All three have 1.05M context windows and 128K max output. On July 30, OpenAI slashed pricing: Luna dropped 80% and Terra dropped 20%. Sol also got a Fast mode (2.5x speed, 2x price).

Luna to Sol is 8 points on the Intelligence Index. That gap matters for frontier problems. For emails, simple coding, summarization, and content drafts, Luna handles those fine. At $0.20/$1.20, Luna is now one of the cheapest frontier-capable models available.

An important finding from the Artificial Analysis Intelligence Index v4.1: Luna and Sol always sit on the cost-performance Pareto frontier, ahead of Terra. After the price cuts, Luna's position on the Pareto frontier is even stronger. It achieves the highest intelligence score of any model at its price point.

Max and Ultra Reasoning

GPT-5.6 adds two reasoning effort levels beyond the existing low/medium/high/xhigh.

Max works like existing levels but goes deeper. More thinking tokens, longer processing, better output on hard problems. All benchmark scores above were measured at max. Every tier supports it.

Ultra is something different. It coordinates four agents in parallel, each working on separate aspects of the problem. This trades higher token usage for stronger results and faster time-to-result on demanding tasks. Built-in agentic parallelism without needing an external orchestrator.

OpenAI's docs say ultra defaults to four agents and can scale up to sixteen. Anthropic takes a different approach with dynamic workflows that can scale to hundreds of agents because coordination runs as a deterministic program rather than inside the model's context. Different architecture, similar idea.

In testing, ultra shines on tasks that benefit from parallel exploration: large codebase analysis, multi-file refactoring, research synthesis. For simple sequential tasks, max or even high effort gives you better value.

Sol Leads the Coding Agent Index

The Artificial Analysis Coding Agent Index v1.1 pairs models with their native coding tools and evaluates them on three benchmarks: DeepSWE, Terminal-Bench v2, and SWE-Atlas-QnA. Unlike lab-reported benchmarks, this tests models inside the actual coding environment they ship with.

Sol (max) in OpenAI's Codex scored 80, leading every evaluation. It tied with Grok 4.5 in Grok Build on SWE-Atlas-QnA, but won the other two.

Sol's cost per coding task is about 40% cheaper than Fable 5 in Claude Code. Luna at 75 now costs a fraction of a dollar per task after the price cut.

Luna's Coding Agent Index score of 75 puts it ahead of GPT-5.5 and close to Fable 5. At $0.20/$1.20, it's by far the cheapest path to near-frontier coding performance. See the full best AI model for coding rankings.

GPT-5.6 Sol: When You Need the Best

Sol sits third on the Intelligence Index at 59, behind Claude Opus 5 (61) and Claude Fable 5 (60). At $5/$30, it costs about a third of what Fable charges ($10/$50) and matches Opus 5's input pricing. Sol is the fastest of the three at 85 tok/s. On July 30, OpenAI added a Fast mode to Sol's API: 2.5x the speed for 2x the price ($10/$60). Same intelligence, faster responses.

Sol supports all reasoning levels from none through ultra. At none or low, it behaves like a fast production model. At max or ultra, it spends significant time and tokens on complex problems. One model serving multiple roles.

Sol also has the highest Presentation Elo score in the new AA-Briefcase benchmark, which tests models on realistic knowledge work tasks. PowerPoint slides, Excel models, document formatting. The outputs look better than any other model tested.

On token efficiency, Sol uses about 15K output tokens per Intelligence Index task. That's fewer than Claude Opus 4.8 or Gemini 3.5 Flash at comparable intelligence levels. Fewer tokens means lower real-world costs than the headline pricing suggests.

Use Sol for: Architecture decisions. Complex multi-file debugging. Security-sensitive code review. Problems that stump cheaper models. Knowledge work that needs polished output.

GPT-5.6 Terra: The Middle Ground

Terra scores 55 on the Intelligence Index at $2/$12 (down from $2.50/$15 after the July 30 cut). It replaces GPT-5.4 and is now cheaper than it was at launch. The 75 tok/s speed is comfortable for interactive use.

Here's the catch. The Artificial Analysis Intelligence Index v4.1 data shows Terra never sits on the Pareto frontier. For any given Terra effort level, you can find a Luna or Sol effort level that gives you either better intelligence at the same price, or the same intelligence cheaper. In our own testing, we couldn't find a use case where Terra was clearly the best choice. Luna handled the easy work. Sol handled the hard work. Terra sat in between, good at both but the best pick for neither.

Use Terra for: When you want one model for everything and don't want to manage routing logic. Daily coding. Feature implementation. Data analysis.

GPT-5.6 Luna: The One Most People Should Default To

Luna is the real story of this release, and the July 30 price cut made it even more so. At $0.20/$1.20 per million tokens (down from $1/$6), it scores 51 on the Intelligence Index and 75 on the Coding Agent Index. That Coding Agent score beats GPT-5.5 and sits near Fable 5 territory. At these prices, Luna costs less than most open-source model hosting.

Speed matters here too. Luna runs at 150 tok/s. For agent sub-tasks where you need quick responses to many small calls, that speed compounds into meaningful time savings.

The 1.05M context window at $0.20/$1.20 is remarkable. No other model with a million-token context and reasoning capabilities comes close on price. Luna is now the obvious default for any workflow where you don't need frontier intelligence.

Use Luna for: Agent sub-tasks. Email drafts. Summarization. Code explanation. Simple edits. High-volume workflows where speed and cost matter more than peak quality.

Cache-Write Pricing

GPT-5.6 introduces cache-write pricing to OpenAI for the first time, matching what Anthropic already does with Claude.

• Cache writes: 1.25x the input token price (tokens being committed to memory)
• Cache reads: 90% discount (same as before)

The first time you send a long system prompt, you pay a 25% premium on those input tokens. Every subsequent request that hits the cache pays only 10% of the input price. For production workflows with consistent system prompts, this is a net positive. You pay slightly more upfront, then save heavily on every repeat call.

Real-World Performance

Benchmarks tell part of the story. Real usage tells the rest. Here's what we've seen and what early testers report.

When we ran Sol and Luna through our benchmark dashboard, Sol consistently placed first on coding tasks across all effort levels. Luna surprised us. At a fifth of Sol's cost, it handled code explanation and simple refactoring well enough that we switched our internal agent sub-tasks from Terra to Luna within a day. The 5-point Intelligence gap didn't matter for those tasks.

Early access tester IndyDevDan ran 67 projects across $180,000-$240,000 of inference. His key finding: Sol grabs onto tasks and doesn't let go until finished. With GPT-5.5, the model would get lost in long sessions or break when bad context crept in, requiring a new thread. Sol holds context across 20+ hour sessions without degradation.

Sol's computer use capabilities are a clear step up. Browser automation, simulator interaction, setting up multi-machine fleets over SSH. Where 5.5 would get confused by vague instructions, Sol infers what you mean and acts on it.

There's an honest critique worth making here. OpenAI led with the Coding Agent Index and AA-Briefcase, where Sol tops the charts. They didn't report SWE-Bench Verified or SWE-Bench Pro in their launch materials, benchmarks where Anthropic's models have historically been stronger. When a lab leads on a benchmark, they cite it. When they trail, they question its methodology. Both things can be true.

In head-to-head testing across 10 tasks (3D renders, interactive maps, first-person shooters, presentations, creative writing), Sol consistently placed first or second against Fable 5, Opus 4.8, and GPT-5.5. Sol won on tasks that reward decisiveness and speed. Fable 5 won on the hardest problems where extended reasoning matters more.

Full native app rewrites (React Native to Swift UI) completed in 2-4 hours with working results. A from-scratch Rust rewrite of a TypeScript project produced 195,000 lines across 29 crates with a functioning transpiler. Not production-ready, but far beyond what 5.5 could produce.

How They Compare to Everything Else

Sol vs Opus 5: Opus 5 leads on intelligence (61 vs 59), novel reasoning (30.2% vs 7.8% on ARC-AGI-3), and costs less on output ($25 vs $30). Sol leads the Coding Agent Index (80 in Codex), is 63% faster (85 tok/s), and wins on DeepSWE and presentation quality. For coding in an IDE, Sol. For everything else, Opus 5 is the stronger pick.

Sol vs Fable 5: Sol costs a third of what Fable charges and runs faster. Fable leads by 1 point on intelligence and wins on deep grinding tasks like SWE-bench. For most work, Sol is the better value.

Luna vs everything at its price: After the 80% price cut, Luna at $0.20/$1.20 has no real competitor at this price point. It scores 51 on the Intelligence Index, which is higher than any other model under $2/M blended cost. It's the obvious choice for agent sub-tasks, batch processing, and any high-volume workflow.

Terra vs Sonnet 5: Terra at $2/$12 now costs less than Sonnet 5's intro pricing ($2/$10 on output). Terra scores higher on intelligence (55 vs 53). After August 31, when Sonnet 5 moves to $3/$15, Terra is clearly cheaper.

Real-World Costs

Estimate your monthly spend with the AI Cost Calculator. Here are rough numbers for 500 tasks at 20,000 tokens average:

Best Setup for Late July 2026

A two-model setup using the GPT-5.6 family:

Primary model: GPT-5.6 Sol ($5/$30) at medium or high effort. Handles complex tasks, code review, analysis, writing. Consider Claude Opus 5 ($5/$25) as an alternative primary if you value novel reasoning and self-correction over speed.

Sub-agent model: GPT-5.6 Luna ($0.20/$1.20). At this price, you can run hundreds of agent calls per dollar. Luna handles delegated tasks, simple edits, lookups, and summarization.

Estimated monthly cost: $40-80 depending on volume (significantly lower than before the price cuts).

For tighter budgets, use Luna as both primary and sub-agent. At $0.20/$1.20 with score 51, it handles most work. Monthly cost drops to $5-15. Or mix providers: Claude Opus 5 at high effort for reasoning-heavy work, Luna for everything else.

FAQ

Is GPT-5.6 Sol better than GPT-5.5?

Yes. Sol scores 59 on the Intelligence Index vs. GPT-5.5's 55. Same price ($5/$30). Sol also leads the Coding Agent Index at 80 points. There's no reason to use GPT-5.5 now.

Which GPT-5.6 model should I use for coding?

Start with Luna ($0.20/$1.20 after the July 30 cut). Its Coding Agent Index score of 75 handles most tasks. Use Sol for complex debugging, architecture work, or when you need max/ultra reasoning. Terra at $2/$12 works if you want a single middle-ground model.

How does GPT-5.6 compare to Claude Opus 5 and Fable 5?

Claude Opus 5 (score 61) now leads the Intelligence Index, ahead of Fable 5 (60) and Sol (59). Opus 5 costs $5/$25 (cheaper output than Sol's $30), but runs slower at 52 tok/s vs Sol's 85. Sol leads DeepSWE and presentation quality. Fable 5 at $10/$50 is the most expensive and hardest to justify unless you need peak medical/legal performance.

What are max and ultra reasoning?

Max is deeper thinking within a single model, similar to existing effort levels but more thorough. Ultra coordinates four or more agents in parallel to tackle complex problems. Ultra trades higher token usage for faster, stronger results on demanding tasks.

Is GPT-5.6 Luna good enough for production?

For many use cases, yes. Luna scores 51 on the Intelligence Index and 75 on the Coding Agent Index with full reasoning capabilities. After the 80% price cut to $0.20/$1.20 per million tokens, it's cheaper than most self-hosted open models. It handles email, summarization, simple coding, and agent sub-tasks. Step up to Sol for complex analysis or frontier-quality output.

Is GPT-5.6 available on OpenRouter?

GPT-5.6 Sol, Terra, and Luna are available on OpenRouter. Check their model page for current pricing and availability.

How does Grok 4.5 compare to GPT-5.6?

Grok 4.5 scores 54 on the Intelligence Index at $2/$6. It sits between Terra (55) and Luna (51) on quality. After the Luna price cut, Grok 4.5's output is 5x more expensive than Luna ($6 vs $1.20). Grok 4.5 is still exceptionally token-efficient, using 60% fewer tokens per task. But for raw cost-per-task, Luna at $0.20/$1.20 is now substantially cheaper.

Benchmark scores from Artificial Analysis Intelligence Index v4.1 and Coding Agent Index v1.1. Pricing from OpenAI's official API documentation and Sam Altman's July 30, 2026 announcement. Real-world observations from IndyDevDan's early access testing and our own benchmark dashboard usage. Updated July 30, 2026. For Claude Opus 5 comparison, see our Opus 5 benchmarks guide. For full coding model rankings, see Best AI Models for Coding. For the overall model comparison beyond GPT-5.6, see Best LLM in 2026 and the LLM Leaderboard. Compare all models on our Benchmark Dashboard or get personalized picks with the Model Selector.