Quick Overview
OpenAI shipped GPT-5.6 to general availability on July 9, 2026, across ChatGPT, Codex, and the API. GPT-5.6 ships as three models: Sol, Terra, and Luna. The number (5.6) identifies the generation. The names identify what OpenAI calls "durable capability tiers that can advance on their own cadence."
The practical question is not whether GPT-5.6 is good. It is. The question is which tier to use for what, because the price gap between Sol and Luna is 5x on input and 5x on output, and the benchmark gap is often much narrower than that.
This guide walks through every benchmark OpenAI published for each tier, the pricing math, the new reasoning modes, and the benchmarks OpenAI chose not to report. By the end you should know whether to default to Terra (probably), when to escalate to Sol, and where Luna breaks.
The Naming Change
GPT-5.6 replaces OpenAI's previous model naming with a tier system. Sol is the flagship. Terra is the balanced everyday model. Luna is the fast, affordable tier. The bare gpt-5.6 API alias routes to gpt-5.6-sol, but the explicit model IDs are gpt-5.6-sol, gpt-5.6-terra, and gpt-5.6-luna (Digital Applied).
The naming is deliberate. OpenAI wants you to think about tiers as ongoing product lines, not one-off releases. Terra can get smarter without becoming Sol. Luna can get faster without becoming Terra. Whether this holds depends on whether OpenAI actually updates tiers independently or just ships GPT-5.7 and starts over.
Pricing: The Three-Tier Math
API pricing per 1M tokens, confirmed flat from preview to GA:
- Sol: $5 input / $30 output
- Terra: $2.50 input / $15 output
- Luna: $1 input / $6 output
- Claude Fable 5: $10 input / $50 output (Anthropic)
Prompt caching got more predictable. Cache writes cost 1.25x the uncached input rate. Cached reads get a 90% discount. Cache life is a minimum of 30 minutes with support for explicit cache breakpoints (TechJack Solutions). For high-volume workloads hitting repeated prefixes, caching can cut effective input costs dramatically.
The headline number: OpenAI says Terra and Luna outperform Claude Fable 5 on Agents' Last Exam at approximately one-quarter the estimated cost, in roughly one-third the time, with about half as many output tokens (OpenAI). Fable 5 costs 2x what Sol costs per token, and 4x what Terra costs. Whether that holds outside OpenAI's own benchmark is a separate question. But the pricing strategy is clear: OpenAI is competing on cost-performance, not just raw capability.
1. Long-Horizon Agentic Work: Agents' Last Exam
Agents' Last Exam evaluates long-running professional workflows across 55 fields. OpenAI led with this benchmark for a reason: all three tiers beat every non-OpenAI model listed.
Sol beats Fable 5 by 13.1 points. But look at the spread within the family: Sol to Luna is 3.3 points. Luna costs one-fifth what Sol costs. If your workload is long-horizon agentic work and you are deciding between Terra and Sol, the benchmark says you are paying 2x for 3.2 points.
2. Terminal and Engineering: Terminal-Bench 2.1 and DeepSWE
Terminal-Bench 2.1 tests command-line workflows that require planning, iteration, and tool coordination. Sol sets a new state-of-the-art here (Neowin).
Sol Ultra (four parallel agents) reaches 91.9% but costs roughly 3x what single-agent Sol costs: about $5 in estimated API spend versus $1.70 for single-agent Sol at 88.8% (Trilogy AI). That is a steep price for 3.1 points.
Fable 5 scores 86.0% on Terminal-Bench 2.1, per Anthropic's own reporting. That puts it between Terra and Luna, not at the top. 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 (MarkTechPost). The catch: Fable 5's terminal score is Anthropic-reported, and independent evaluators place it slightly lower, between 83.4% and 84.3% (Yellow). Either way, Sol leads on terminal work. Fable 5 leads on repo-level coding. Different task, different winner.
DeepSWE, which tests long-horizon engineering in real codebases, is where the cost-performance story is sharpest. 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 (Trilogy AI). Luna wins on work done per dollar, and by a lot.
3. Coding: The Artificial Analysis Coding Agent Index
The Artificial Analysis Coding Agent Index scores Sol at 80, Terra at 77.4, and Luna at 74.6. Fable 5 scores 77.2, placing it between Terra and Luna (MarkTechPost). Sol beats Fable 5 by 2.8 points while using less than half the output tokens.
But the Coding Agent Index measures agentic coding: terminal workflows, tool coordination, and real codebase navigation. That is GPT-5.6's home turf. The SWE-Bench Pro elephant is a different room.
OpenAI did not report SWE-Bench Verified. They did report SWE-Bench Pro, where Sol scores 64.6% compared to Claude Fable 5's 80% (Simon Willison). That is a 15.4-point gap. OpenAI's response was to publish a separate article estimating that approximately 30% of SWE-Bench Pro tasks are broken and advising model developers to "carefully examine results."
Maybe OpenAI is right. Maybe SWE-Bench Pro has methodology problems. But when a lab leads on a benchmark, they cite it. When they lose on it, they question its methodology. Fable 5 still leads on SWE-Bench Pro, and OpenAI's own table shows Claude Mythos 5 also ahead of Sol there (Kingy.ai). For pure code generation and repo-level coding, Fable 5 is still the strongest publicly available model. GPT-5.6's coding strength is agentic: terminal workflows, multi-step tool coordination, and long-horizon engineering. Different task, different winner.
4. Computer Use and Browsing: OSWorld 2.0 and BrowseComp
Sol sets a new state-of-the-art on BrowseComp at 92.2% and OSWorld 2.0 at 62.6% (OpenAI). On OSWorld, Sol surpasses Claude Opus 4.8 while using 85% fewer output tokens. Terra surpasses GPT-5.5's peak at a lower cost. Luna nearly matches GPT-5.5 at less than half the estimated cost.
OpenAI did not publish Fable 5 scores on BrowseComp or OSWorld 2.0 in the GPT-5.6 comparison table. Anthropic separately reports Fable 5 at 85.0% on OSWorld-Verified, a different variant of the benchmark with human-verified task completions (Anthropic, MorphLLM). Because OSWorld 2.0 and OSWorld-Verified use different task sets, the numbers are not directly comparable. What we can say: Sol's 62.6% on OSWorld 2.0 is the state-of-the-art on that specific benchmark, and Fable 5's 85.0% is the state-of-the-art on OSWorld-Verified. Both lead, on different tests.
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. That is exactly what GPT-5.6 was trained to do.
5. Cybersecurity: ExploitBench
On ExploitBench 1, Sol scores 73.5% compared to 47.9% for GPT-5.5 at a comparable output-token budget (Neowin). The Fable 5 comparison is murky. 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 (LLM Stats, MorphLLM). If you see Fable 5 quoted at 78% on ExploitBench, that is the Mythos 5 number.
On ExploitGym, a benchmark created by UC Berkeley with OpenAI and other frontier labs, all three GPT-5.6 tiers show strong improvements in cyber capabilities as reasoning increases (OpenAI).
This is also why the models spent 12 days behind a government gate. Trump's June AI cybersecurity order required review of powerful models before public release. The Commerce Department's Center for AI Standards and Innovation cleared GPT-5.6 after OpenAI sent technical experts to Washington (Engadget, TechTimes). The review was legally voluntary. Practically, it functioned like preclearance.
6. The Luna Problem: Long-Context Recall
Here is where the tier separation actually shows up.
Sol and Terra are nearly tied at 91.5% and 89.6%. Luna drops to 41.3%. That is a cliff. If your workload involves long-context recall (document analysis, large codebase reasoning, multi-document synthesis), Luna is the wrong tool. OpenAI did not publish Fable 5 MRCR scores in the GPT-5.6 comparison table. Anthropic's Fable 5 context window is unpublished, and GPT-5.5 (the prior generation) posted 74.0% on OpenAI's MRCR v2 at 512K-1M tokens (PromptsRush). Without a head-to-head on the same benchmark, long-context recall is one category where we genuinely cannot compare the two.
What OpenAI Did Not Report
OpenAI's benchmark suite for GPT-5.6 is deliberately agentic. It leads with Terminal-Bench, Agents' Last Exam, BrowseComp, and OSWorld. These are benchmarks that resemble work: browse, inspect, coordinate tools, write code, compare evidence, validate a result, and recover when the first attempt is wrong (Kingy.ai).
What is missing: no SWE-bench Verified, no GPQA Diamond, no AIME, no MMLU, no ARC-AGI-2, no FrontierMath (o-mega). 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 (Codersera).
The charitable read: single-shot academic benchmarks do not 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, GDPval-AA v2, FrontierMath, HealthBench Professional, and the Artificial Analysis Intelligence Index v4.1 (Kingy.ai). On that last one: Sol with max reasoning scores 59 on the Artificial Analysis Intelligence Index, just one point behind Fable 5 at 60, while completing tasks 61% faster at roughly half the estimated cost (The Decoder). GPT-5.6 Sol leads on Terminal-Bench, BrowseComp, OSWorld, Agents' Last Exam, and cybersecurity. The frontier is split. Anyone telling you one lab swept the board is selling something.
Max Reasoning and Ultra Mode
GPT-5.6 introduces two new capability levers on top of the tier system.
Max reasoning effort is a new level above "high." It gives Sol more time to reason deeply on difficult tasks. You can set reasoning.mode: "pro" on any GPT-5.6 tier, and reasoning effort is configurable per request from "none" to "max" (Digital Applied). Reasoning also persists across turns now, so multi-turn conversations do not restart from zero each time.
Ultra mode goes beyond a single agent. It spawns parallel subagents (four by default) to attack a problem from multiple angles. The API equivalent is the multi-agent beta in the Responses API. Ultra mode improves results on BrowseComp (+1.8 points), SEC-Bench Pro (+3.1), and Terminal-Bench (+3.1) (Trilogy AI). The cost is steep: Sol Ultra on Terminal-Bench costs roughly 3x single-agent Sol for those 3.1 extra points.
The takeaway: max reasoning and ultra mode are quality-first tools for the hardest tasks. For routine work, they waste compute.
The Routing Decision
Here is the practical guidance, based on the numbers above.
Default to Terra. It matches or comes within 2 to 3 points of Sol on most benchmarks at half the price. On Agents' Last Exam, Terminal-Bench, and the Coding Agent Index, the gap is small enough that the cost difference dominates. Terra also surpasses GPT-5.5's peak performance on OSWorld and BrowseComp at a lower cost, which makes it a generational upgrade for anyone currently on GPT-5.5.
Escalate to Sol for the hardest agentic tasks. Long-horizon multi-step workflows, complex terminal operations, cybersecurity research, and computer use at the frontier. Sol's leads on BrowseComp (92.2%), OSWorld (62.6%), and Terminal-Bench (88.8% single-agent, 91.9% Ultra) are real. If the task requires max reasoning or ultra mode, Sol is the only tier that supports them fully.
Use Luna for volume, not for depth. Luna is the cost champion. On DeepSWE, it delivers 24 benchmark points per estimated API dollar versus 3.2 for Fable 5. For high-volume pipelines, classification, summarization, and tasks where 85% of Sol's quality at one-fifth the price is a good trade, Luna is the right pick. But do not use Luna for long-context recall (41.3% on MRCR) or for the hardest reasoning tasks where the gap to Sol widens.
Use Claude Fable 5 for pure code. If your primary workload is repo-level coding and you care about SWE-Bench Pro, Fable 5's 80% still leads Sol's 64.6%. OpenAI's argument that the benchmark is broken may be valid, but until a replacement emerges, the number stands.
Takeaways
- Terra is the practical center of the lineup. GPT-5.5-class performance at half the price on most benchmarks. This should be the default for most teams.
- Sol owns agentic work, not coding. Terminal-Bench, BrowseComp, OSWorld, and Agents' Last Exam are Sol's clean wins. SWE-Bench Pro and FrontierMath are not.
- Luna wins on cost-per-unit-of-work, not on capability. The MRCR cliff (41.3% vs Sol's 91.5%) means Luna has a hard ceiling on long-context tasks.
- Ultra mode is expensive. Three times the cost of single-agent Sol for 3 extra points on Terminal-Bench. Use it when correctness matters more than budget.
- The benchmark table is curated. OpenAI led with agentic benchmarks where they lead and omitted traditional academic benchmarks where they trail. That does not make the agentic wins less real. It means the full picture requires reading both tables.
- GPT-5.4 retires July 23. GPT-5.5 stays available. If you are on GPT-5.4, move to Terra or Luna now, not Sol, unless you specifically need the flagship tier (9to5Mac).
