Capability ceiling
58.9intelligenceSol
For architecture, difficult debugging, security-sensitive reasoning and high-cost mistakes.
- Fast effective speed
- 51.8 tok/s
- API output list price
- $30 / 1M
- Highest local mode
- Ultra
GPT‑5.6 DECISION ATLASTR editionYGT LABS AI / RESEARCH EDITION · JULY 10, 2026
Three models.
One clear
route.
Choose GPT‑5.6 Sol, Terra or Luna by the real bottleneck: capability, daily delivery or high-volume speed. This English edition keeps the benchmark evidence, local measurements and operating limits readable without hiding the trade-offs.
- 40
- official benchmark rows
- 5
- independent sources
- 34
- isolated local runs
- 17
- Standard / Fast pairs
MODEL ROLES
Capability, balance
and throughput.
These are decision roles, not claims that one model is universally best. The intelligence index is a max-reference comparative index; local speed comes from a single fixed-task run per cell.
Daily driver
55intelligenceTerra
For daily coding, research and controlled agent delivery with a strong price-to-capability balance.
- Fast effective speed
- 60.7 tok/s
- API output list price
- $15 / 1M
- Highest local mode
- Ultra
Speed and scale
51.2intelligenceLuna
For summaries, classification, first drafts and high-volume helper tasks after quality has been measured.
- Fast effective speed
- 68.4 tok/s
- API output list price
- $6 / 1M
- Highest local mode
- Max
REASONING LEVELS
More thinking changes
the resource budget.
Effort can improve how much work a model invests in a task. It is not a universal IQ dial, and published effort curves do not exist for every downstream capability.
- 01Sol · Terra · Luna
Low
Short planning budget for quick, bounded work.
- 02Sol · Terra · Luna
Medium
Balanced daily problem-solving budget.
- 03Sol · Terra · Luna
High
Longer chains and more checks for non-trivial work.
- 04Sol · Terra · Luna
XHigh
Expanded intermediate reasoning for complex work.
- 05Sol · Terra · Luna
Max
Highest standard single-agent budget for difficult problems.
- 06Sol · Terra
Ultra
An orchestration mode, not a directly comparable single-agent effort level.
YGT LABS AI / APPLIED AI
Do not just select
a model. Build a system.
YGT Labs AI joins product engineering, AI, IoT and integrated automation. The atlas is the model-routing layer: it makes the starting capability, effort and cost boundary explicit before a workflow reaches production.
Product engineering
Architecture options, repository discovery, implementation plans and test-first changes. Use Sol when a wrong answer is expensive; start daily delivery with Terra.
Agent operations
Triage support, classify work, prepare research briefs and make human approval points explicit. A model is one part of the system, not the process itself.
Content, SEO and GEO
Create source-backed expert content, structured data and multilingual quality checks. The goal is useful, reviewable answers—not automated content volume.
Research and data quality
Turn scattered inputs into evidence matrices, decision notes and visible uncertainty. Use a stronger route for conflicts, exceptions and final review.
Vertical SaaS and automation
Connect assistants to portals, CRM, IoT or operations with business rules, data ownership and measurable outcomes intact.
EVIDENCE, NOT SLOGANS
Every claim carries
its own boundary.
Official benchmark rows, independent signals and local telemetry answer different questions. We keep their source class and limits separate.
Agents’ Last Exam
- Sol
- 52.7
- Terra
- 50.4
- Luna
- 50.3
Long-horizon professional tasks.
AA Coding Agent Index
- Sol
- 80.0
- Terra
- 77.4
- Luna
- 74.6
Composite coding-agent index.
GeneBench Pro
- Sol
- 28.7
- Terra
- 23.3
- Luna
- 10.8
Long scientific workflows.
SEC-Bench Pro
- Sol
- 71.2 · Ultra 74.3
- Terra
- 57.7
- Luna
- 48.9
Security-focused agent tasks.
MRCR 512K–1M
- Sol
- 73.8
- Terra
- 72.5
- Luna
- 41.3
Retrieval across 512K–1M context.
ARC-AGI-3
- Sol
- 7.78
- Terra
- 0.80
- Luna
- 0.18
Interactive abstract reasoning.
Local test method
- Every cell used a separate ephemeral Codex process.
- The task was a fixed weighted-interval-scheduling problem.
- Standard and Fast ran across every supported effort.
- Wall time, token classes, cache and API-equivalent cost were retained.
What this does not prove
Each local cell has n=1. Queueing, cache and sampling introduce noise. Effective token/s is not decoder throughput, and one task is not a general-intelligence test.
PRIMARY SOURCES