GPT-5.6 Sol, Terra, Luna: Full Benchmark Analysis and Which Tier to Actually Use

TL;DR: OpenAI shipped GPT-5.6 to general availability on July 9, 2026, as a three-tier family: Sol (flagship, $5/$30 per M tokens), Terra (balanced, $2.50/$15), and Luna (efficient, $1/$6). Sol leads on agentic benchmarks — Terminal-Bench 2.1 (88.8%), BrowseComp (92.2%), and the Artificial Analysis Coding Agent Index (80) — while coming within 1 point of Claude Fable 5 on the AA Intelligence Index at roughly one-third the cost. On pure code (SWE-Bench Pro), Fable 5 still leads 80% to Sol’s 64.6%. The practical question is routing: Terra is the sensible default for most work, Sol for the hardest agentic tasks, Luna for high-volume pipelines, and ultra mode only when parallelism justifies 3× the cost for ~3 extra points.

Introduction

Four weeks after OpenAI previewed GPT-5.6 Sol under controlled access, the full family launched into general availability on July 9, 2026 — across ChatGPT, Codex, and the API (Source: OpenAI — GPT-5.6: Frontier intelligence that scales with your ambition). The release marks two structural shifts. First, OpenAI replaced the traditional single-model launch with three “durable capability tiers” — Sol, Terra, and Luna — that can evolve on independent cadences. Second, they introduced two new capability levers on top of the tier system: max reasoning effort and ultra mode, which coordinates multiple agents in parallel.

The naming is deliberate. The number (5.6) identifies the generation. The names identify persistent product lines: Sol as the flagship, Terra as the balanced everyday model, and Luna as the cost-efficient option. The bare gpt-5.6 API alias routes to gpt-5.6-sol, but explicit model IDs are gpt-5.6-sol, gpt-5.6-terra, and gpt-5.6-luna (Source: Vellum — GPT-5.6 Sol vs Terra vs Luna).

This is not just a model release. It is OpenAI’s bid to own the cost-performance frontier at every price point — from Luna beating Claude Opus 4.8 at one-quarter the estimated cost, to Sol tying Fable 5 on aggregate intelligence while running 61% faster. The practical question for engineering teams is not whether GPT-5.6 is good. It is. The question is which tier to route to which workload, because the price gap between Sol and Luna is 5×, and the benchmark gap is often much narrower.

The Tier System and Pricing

Three Tiers, One Generation

Tier	Input / 1M tokens	Output / 1M tokens	Role
Sol	$5.00	$30.00	Flagship — hardest agentic tasks, max reasoning, ultra mode
Terra	$2.50	$15.00	Balanced — everyday work, scoped coding, first-pass review
Luna	$1.00	$6.00	Cost champion — high-volume pipelines, classification, first-pass

For comparison, Claude Fable 5 costs $10/$50 and Claude Opus 4.8 costs $5/$25 per million tokens input/output (Source: CodeRabbit — GPT-5.6 Sol and Terra: Where they fit). Sol is priced below Fable 5 and at parity with Opus 4.8 on input, while Terra and Luna create substantially cheaper routing lanes.

Prompt caching got more predictable. Cache writes cost 1.25× the uncached input rate — a first for OpenAI, matching Anthropic’s approach. Cached reads get a 90% discount, with a minimum cache life of 30 minutes and support for explicit cache breakpoints. For high-volume workloads hitting repeated prefixes, effective input costs can drop dramatically.

Why This Pricing Matters

OpenAI’s headline: Terra and Luna outperform Claude Fable 5 on Agents’ Last Exam at approximately one-sixteenth the estimated cost, in roughly one-third the time, with about half as many output tokens (Source: OpenAI — GPT-5.6). Even if those exact ratios don’t hold outside OpenAI’s own benchmarks, the pricing strategy is clear: OpenAI is competing on cost-performance, not just raw capability. Fable 5 costs 2× what Sol costs per token and 4× what Terra costs.

On DeepSWE, the cost-performance story is sharper still. GPT-5.6 Luna delivers roughly 24 benchmark points per estimated API dollar, compared with 4.5 for Claude Opus 4.8 and 3.2 for Claude Fable 5 — a 5× to 7.5× efficiency advantage (Source: Trilogy AI — GPT-5.6 Gains a Powerful Edge). Luna doesn’t just cost less; it does more work per dollar.

Benchmarks: Where Each Tier Wins

Long-Horizon Agentic Work: Agents’ Last Exam

Agents’ Last Exam evaluates long-running professional workflows across 55 fields. It is the benchmark OpenAI led with, and for good reason: all three GPT-5.6 tiers beat every non-OpenAI model listed.

GPT-5.6 Sol with max reasoning sets a new state-of-the-art at 53.6, eclipsing Claude Fable 5 by 13.1 points. At medium reasoning effort, Sol still beats Fable 5 by 11.4 points at roughly one-quarter the estimated cost (Source: OpenAI — GPT-5.6).

The spread within the family is telling. The gap from Sol to Luna is approximately 3.3 points on this benchmark, but Luna costs one-fifth what Sol costs per token. Terra sits in between: you pay 2× Luna’s price for roughly 3.2 additional points over Luna. For workloads that don’t need the absolute ceiling, Terra and Luna offer dramatically better cost-performance.

Terminal Work: Terminal-Bench 2.1

Terminal-Bench 2.1 tests command-line workflows requiring planning, iteration, and tool coordination. Sol sets a new state-of-the-art:

Model	Score
GPT-5.6 Sol (single-agent)	88.8%
GPT-5.6 Sol Ultra (4 agents)	91.9%
Claude Fable 5	~86.0%
GPT-5.6 Terra	~84-85%
GPT-5.6 Luna	~80-83%

(Source: Trilogy AI, Vellum)

The Ultra configuration reaches 91.9% but costs roughly 3× what single-agent Sol costs: approximately $5 in estimated API spend versus $1.70 for single-agent Sol at 88.8%. That is a steep price for 3.1 extra points.

Fable 5’s terminal score places it between Terra and Luna. OpenAI’s headline — Sol beats Fable 5 by 2.8 points while using less than half the output tokens and roughly one-third the estimated cost — is accurate, but Fable 5’s terminal score is Anthropic-reported. Independent evaluators place it slightly lower, between 83.4% and 84.3%.

Engineering in Real Codebases: DeepSWE

DeepSWE tests long-horizon engineering in real codebases — navigating unfamiliar repositories, making surgical changes, and passing behavioral checks. This is where the tier separation matters most.

On absolute scores, Sol leads the family (72.7% in Trilogy AI’s reporting), with Terra at 69.6% — virtually tied with Fable 5 at 69.7%. But the cost-performance chart is where the story gets interesting:

Model	Score	Est. API $/Benchmark Point
GPT-5.6 Luna @Max	67.2%	~24 points/dollar
GPT-5.6 Terra @Max	69.6%	very high efficiency
GPT-5.6 Sol @Max	72.7%	moderate efficiency
Claude Opus 4.8 @Max	59.0%	~4.5 points/dollar
Claude Fable 5 @Max	69.7%	~3.2 points/dollar

(Source: Trilogy AI)

Luna’s efficiency on DeepSWE is the single strongest argument for routing high-volume coding work to the cheapest tier first. It beats Opus 4.8 on absolute score while costing a fraction of the estimated inference budget. The routing recommendation from Trilogy AI is straightforward: use Luna as the default for routine coding, Terra with max reasoning for escalations, and Sol for genuinely hard problems where marginal reliability matters more than cost.

Coding Agent Index: Sol Leads, but the Gap Is Narrow

The Artificial Analysis Coding Agent Index pairs models with agentic harnesses and evaluates them across three frontier coding evaluations: DeepSWE, Terminal-Bench v2, and SWE-Atlas-QnA (Source: Artificial Analysis — GPT-5.6 has landed).

Model	Index Score	Relative Cost/Task vs Sol
GPT-5.6 Sol (max) in Codex	80	baseline
GPT-5.6 Terra (max)	77.4	~60% cheaper
Claude Fable 5 (max) in Claude Code	77.2	~40% more expensive
GPT-5.6 Luna (max)	74.6	~80% cheaper
Claude Opus 4.8 (max) in Claude Code	72.5	~10% more expensive

Sol sets the state of the art at 80, 2.8 points above Fable 5. Terra effectively ties Fable 5 at 77.4, and Luna outperforms Opus 4.8 at 74.6. Each GPT-5.6 tier beats its Anthropic price-equivalent on this index. But the spread from Sol to Luna is only 5.4 points — an 80% cost reduction for a 7% capability gap. That is a strong argument for defaulting to Terra or Luna on most coding tasks.

The SWE-Bench Pro Elephant

OpenAI did not report SWE-Bench Verified in their GPT-5.6 benchmark suite. They did report SWE-Bench Pro, where Sol scores 64.6% compared to Claude Fable 5’s 80%. That is a 15.4-point gap — large enough that OpenAI published a separate article arguing approximately 30% of SWE-Bench Pro tasks are broken and advising model developers to “carefully examine results” (Source: Vellum).

Maybe OpenAI is right about the benchmark. Maybe Fable 5 genuinely leads on repo-level code generation. Both can be true. What matters for engineering teams: if your primary workload is open-ended code generation in unfamiliar repositories and you care about SWE-Bench Pro as a proxy, Fable 5 still holds the lead. GPT-5.6’s coding strength is agentic — terminal workflows, multi-step tool coordination, and long-horizon engineering — not single-shot code generation.

CodeRabbit’s independent testing confirms this split. In their long-horizon coding run with 100+ tasks across TypeScript, Go, Python, JavaScript, and Rust, Sol passed 63.7% of tasks with an average of 20,968 output tokens per task. Terra passed 40.7% but used 55,594 output tokens per task — cheaper per token but more expensive per solved task (Source: CodeRabbit). For long coding jobs, measure cost per solved task, not cost per token.

Browsing and Computer Use: BrowseComp and OSWorld 2.0

Sol sets a new state-of-the-art on BrowseComp at 92.2% and OSWorld 2.0 at 62.6% (Source: OpenAI — GPT-5.6). On OSWorld, Sol surpasses Claude Opus 4.8 while using 85% fewer output tokens — a dramatic efficiency gain.

The performance-per-dollar gains extend across the family. Luna nearly matches GPT-5.5’s peak performance at less than half the estimated cost, while Terra surpasses it at a lower cost. This is Sol’s cleanest story: browsing and computer use are agentic tasks where planning, iteration, and recovery from failure matter more than single-shot reasoning — exactly what GPT-5.6 was trained to do.

OpenAI also highlights improved design judgment and computer-use capabilities. GPT-5.6 can inspect and refine rendered interfaces, not just generate the underlying code, catching visual and functional issues before handing work back. The latest Artificial Analysis benchmarks confirm Sol has the highest Presentation Elo of any model tested in AA-Briefcase — its outputs across PowerPoint and Excel are the most visually polished available (Source: Artificial Analysis).

Intelligence Index: The 1-Point Gap

On the Artificial Analysis Intelligence Index v4.1, a broad measure spanning agentic work, coding, scientific reasoning, and general capabilities, GPT-5.6 Sol with max reasoning scores 59, just 1 point behind Claude Fable 5 at 60 — while completing tasks 61% faster at roughly half the estimated cost (Source: Artificial Analysis).

The Pareto frontier across reasoning effort levels is equally instructive. GPT-5.6 Terra and Luna push past GPT-5.5 at every reasoning level. Notably, Luna and Sol are always on the Pareto frontier ahead of Terra — meaning there is always a Luna or Sol effort level that is more intelligent at no extra cost, or equally intelligent at lower cost, than any Terra configuration.

Cybersecurity: ExploitBench and ExploitGym

On ExploitBench 1, Sol scores 73.5% compared to GPT-5.5’s 47.9% at a comparable output-token budget — a 25.6-point leap. This is one of the largest generation-over-generation improvements in the benchmark suite, and one reason the models spent 12 days behind a government gate. The Trump administration’s June AI cybersecurity executive order required review of powerful models before public release, and the Commerce Department’s Center for AI Standards and Innovation cleared GPT-5.6 after OpenAI sent technical experts to Washington (Source: OpenAI — GPT-5.6).

Anthropic reports 78.0% on ExploitBench, but that score belongs to Claude Mythos 5 — the unrestricted variant gated to Project Glasswing partners. Fable 5 uses the same weights with safety classifiers layered on top, and Anthropic has not published a Fable 5-specific ExploitBench number (Source: Vellum).

Long-Context Recall: The Luna Cliff

Multi-hop Retrieval over Contextual Reasoning (MRCR) tests long-context recall. The results show genuine tier separation:

Model	MRCR Score
GPT-5.6 Sol	91.5%
GPT-5.6 Terra	89.6%
GPT-5.6 Luna	41.3%

(Source: Vellum)

Luna drops off a cliff. If your workload involves document analysis, large codebase reasoning, or multi-document synthesis, Luna is the wrong tool. GPT-5.5 (the prior generation) posted 74.0% on OpenAI’s MRCR v2 at 512K-1M tokens, placing Luna substantially behind its predecessor on recall tasks. This is the one category where Luna’s cost advantage does not compensate for its capability gap.

Max Reasoning and Ultra Mode

GPT-5.6 introduces two new capability levers on top of the tier system.

Max Reasoning

max is a new reasoning effort level above xhigh. It gives Sol more time to reason deeply, run checks, explore alternatives, and revise its approach. You can set reasoning.mode: "pro" on any GPT-5.6 tier, and reasoning effort is configurable per request from "none" to "max". Reasoning also persists across multi-turn conversations — a quality-of-life improvement for long agent sessions.

On the AA Intelligence Index, the jump from default to max reasoning adds approximately 2-4 points across tiers, with diminishing returns at the highest effort levels.

Ultra Mode: Parallel Agents

ultra coordinates four agents in parallel by default, trading higher token use for stronger results and faster time-to-result on demanding tasks. The gains are real but narrow:

Benchmark	Single-Agent Sol	Sol Ultra (4 agents)	Delta	Cost Multiplier
Terminal-Bench 2.1	88.8%	91.9%	+3.1	~3×
BrowseComp	~90.4%	92.2%	+1.8	~3×
SEC-Bench Pro	—	—	+3.1	~3×

(Source: OpenAI — GPT-5.6, Trilogy AI)

The cost math is brutal. On Terminal-Bench 2.1, single-agent Sol reaches 88.8% at roughly $1.70 in estimated API cost, while Ultra reaches 91.9% at a little over $5. That’s approximately 3× the cost for a 3.1-point gain. Ultra is a sensible option when a task can be split into independent workstreams and finishing sooner has real value. It is not the sensible default. For API users, developers can build ultra-like experiences using the multi-agent beta in the Responses API.

What OpenAI Did Not Report

OpenAI’s benchmark suite for GPT-5.6 is deliberately agentic. They lead with Terminal-Bench, Agents’ Last Exam, BrowseComp, OSWorld, and the AA Coding Agent Index — benchmarks that resemble real work: browse, inspect, coordinate tools, compare evidence, validate results, and recover when the first attempt is wrong (Source: Vellum).

What is conspicuously absent: no SWE-bench Verified, no GPQA Diamond, no AIME, no MMLU, no ARC-AGI-2, no FrontierMath. These are the traditional academic benchmarks that have defined frontier model comparisons for two years. OpenAI also changed their reporting format: instead of publishing a single benchmark score, they now show performance as a curve across reasoning-effort levels.

The charitable read: single-shot academic benchmarks don’t capture how models are actually used, and agentic benchmarks are a better proxy for real work. The cynical read: OpenAI leads on the agentic benchmarks and trails on several academic ones, so they foregrounded the former and backgrounded the latter.

Both can be true. Claude Fable 5 leads on SWE-Bench Pro, GPQA Diamond, FrontierMath, HealthBench Professional, and the AA Intelligence Index (by 1 point). GPT-5.6 Sol leads on Terminal-Bench, BrowseComp, OSWorld, Agents’ Last Exam, the AA Coding Agent Index, and cybersecurity. The frontier is split. Anyone telling you one lab swept the board is selling something.

The Routing Decision: Which Tier for What

Based on the full benchmark data, here is the practical guidance:

Default to Terra

Terra matches or comes within 2-4 points of Sol on most benchmarks at half the price: 77.4 vs 80 on the Coding Agent Index, 69.6% vs 72.7% on DeepSWE, and 89.6% vs 91.5% on MRCR. On Agents’ Last Exam, the Sol-Terra gap is approximately 3.2 points — you pay 2× for 3.2 extra points. Terra also surpasses GPT-5.5’s peak performance on OSWorld and BrowseComp at a lower cost, making it a generational upgrade for anyone currently on GPT-5.5.

CodeRabbit’s recommendation: use Terra for first-pass implementation, review triage, and scoped fixes with escalation available to Sol (Source: CodeRabbit).

Escalate to Sol for the Hardest Agentic Tasks

Sol’s leads on BrowseComp (92.2%), Terminal-Bench (88.8% single-agent), and OSWorld (62.6%) are real and significant. Reserve Sol for:

Long-horizon multi-step agentic workflows
Complex terminal operations and debugging
Cybersecurity research
Computer use and browser automation at the frontier
Tasks where marginal reliability matters more than inference cost: critical migrations, security review, architecture decisions
When you need max reasoning or ultra mode

CodeRabbit’s engineering assessment: “Sol is best pictured as a persistent engineer: plain-spoken and stubborn about the list.” It follows through on long checklists, stays oriented across large repos, and does the unglamorous work that makes an agent useful. For architectural discussion, product tradeoffs, or open-ended planning, Fable 5 remains the stronger pick (Source: CodeRabbit).

Use Luna for Volume, Not Depth

Luna is the cost champion. On DeepSWE, it delivers 24 benchmark points per estimated API dollar. For high-volume pipelines, classification, summarization, PR summaries, lightweight review prechecks, and tasks where 85% of Sol’s quality at one-fifth the price is a good trade, Luna is the right pick.

But Luna has sharp edges. Long-context recall (41.3% on MRCR) is a cliff. CodeRabbit’s long coding run found that while Luna is great for top-of-funnel agent work, “the key is low reasoning. If the task is mostly summarizing, labeling, extracting, or drafting a scaffold, Luna is the place to start.” For anything deeper, escalate (Source: CodeRabbit).

Use Claude Fable 5 for Pure Code Generation

If your primary workload is repo-level code generation and you care about SWE-Bench Pro, Fable 5’s 80% still leads Sol’s 64.6%. OpenAI’s argument that the benchmark is flawed may be valid, but until a replacement emerges, the number stands. For agentic coding (terminal workflows, multi-step coordination, long-horizon engineering), Sol is the better pick.

The Multi-Model Stack

CodeRabbit’s testing points to a practical multi-model routing pattern for software development:

SDLC Stage	First Model to Try	Why
Planning	Fable 5 for architecture, Sol for execution plans	Fable better for open-ended tradeoffs; Sol better when the plan becomes a checklist
Research	Sol	Strong at reading existing code, tracing dependencies, staying oriented
Execution	Sol	Persistent, follows long task lists, works through boring implementation
Testing	Sol	Runs suites, interprets failures, adds targeted tests, loops until fixed
Review (recall)	Sol (with filtering)	Finds more issues; needs product filtering around output
Review (comment quality)	Claude Sonnet 5	Higher precision, cleaner comments, better developer attention ratio

(Source: CodeRabbit)

FAQ

Q: What’s the difference between max reasoning and ultra mode?

max reasoning gives a single model more time to think — it explores alternatives, runs checks, and revises its approach within one agent. ultra spawns four parallel agents by default, each working on different angles of the same problem. Ultra costs roughly 3× as much as single-agent Sol and adds approximately 1.8-3.1 points on benchmarks. Use max for deeper single-agent reasoning. Use ultra only when a task can be parallelized and finishing faster has real value.

Q: Why does Luna score so poorly on long-context recall (41.3% vs 91.5% for Sol)?

The tier separation is most visible in retrieval-intensive tasks. Luna sacrifices context-window efficiency and retrieval accuracy for speed and cost. If your workload involves analyzing documents, navigating large codebases, or synthesizing across multiple sources, use Terra or Sol instead. Luna is designed for bounded, low-reasoning tasks.

Q: Is GPT-5.6 better than Claude Fable 5?

It depends on the task. On agentic benchmarks (Terminal-Bench, BrowseComp, OSWorld, Agents’ Last Exam, Coding Agent Index), Sol leads. On pure code generation (SWE-Bench Pro), academic benchmarks (GPQA Diamond, MMLU), and aggregate intelligence (AA Intelligence Index), Fable 5 leads — but by just 1 point on the aggregate index, while costing roughly 3× more. The two models are complementary: Fable 5 for architectural judgment and open-ended reasoning, Sol for execution and follow-through.

Q: I’m currently on GPT-5.5. Should I switch everything to Terra?

Terra surpasses GPT-5.5’s peak performance on OSWorld and BrowseComp at a lower cost, and matches or exceeds GPT-5.5 on coding benchmarks while using fewer tokens. For most workloads, migrating from GPT-5.5 to Terra is a no-brainer upgrade — better quality at lower cost. For the hardest agentic tasks, add Sol as a specialized escalation lane.

Q: Why didn’t OpenAI report SWE-Bench Verified or MMLU scores?

OpenAI’s benchmark strategy for GPT-5.6 is deliberately agentic — they chose evaluations that resemble real work over traditional academic benchmarks. The practical consequence is that we don’t have head-to-head comparisons on several benchmarks where Anthropic leads. OpenAI argues these benchmarks don’t capture how models are actually used. Whether that’s a principled position or a tactical omission is a matter of interpretation.