GPT-5.6 Statistics 2026
By Axis Intelligence Research
Co-author: Sarah Mitchell | Last updated: July 18, 2026 | License: CC BY 4.0
GPT-5.6 — OpenAI’s three-tier model family of Sol, Terra, and Luna — launched to general availability on July 9, 2026. Sol scores 80 on the Artificial Analysis Coding Agent Index (2.8 points above Claude Fable 5), reaches 53.6 on Agents’ Last Exam (13.1 points above Fable 5), and prices at $5/$30 per million input/output tokens. Luna, the cheapest tier at $1/$6 per million tokens, scores 74.6 on the same coding index — within 5.4 points of Sol at one-fifth the cost.
Quick Answer
GPT-5.6 is a family of three models — Sol (flagship), Terra (balanced), and Luna (fast/affordable) — released by OpenAI on July 9, 2026. All three share a 1,050,000-token context window, 128,000 maximum output tokens, and a February 16, 2026 knowledge cutoff. Pricing per million tokens: Sol $5/$30, Terra $2.50/$15, Luna $1/$6. On Agents’ Last Exam, Sol scores 53.6 — the highest published result on that benchmark as of its release.
Key Findings
- GPT-5.6 Sol scores 80 on the Artificial Analysis Coding Agent Index v1.1, setting a new state of the art at its release date and placing 2.8 points above Claude Fable 5 (77.2), while using less than half the output tokens, per OpenAI’s July 9, 2026 launch data.
- GPT-5.6 Sol reaches 53.6 on Agents’ Last Exam — an evaluation of long-running professional workflows across 55 fields — surpassing Claude Fable 5’s 40.5 by 13.1 points, according to OpenAI’s published benchmark table (July 9, 2026).
- GPT-5.6 Luna delivers 74.6 on the Artificial Analysis Coding Agent Index — outperforming Claude Opus 4.8 (72.5) at approximately one-quarter the estimated cost, per OpenAI’s benchmark release data and published pricing.
- Axis Intelligence Research calculates that the Sol-to-Luna price ratio for output tokens is 5x ($30 vs $6 per million), while the benchmark gap on the Coding Agent Index is only 6.75% (80 vs 74.6), yielding a cost-performance compression ratio of 0.74 — meaning Luna captures 93.25% of Sol’s coding benchmark score for 20% of Sol’s output cost.
- GPT-5.6 Sol trails Claude Mythos 5 on SWE-Bench Pro by 15.7 points (64.6% vs 80.3%), the single most significant benchmark gap in the family’s published results, according to OpenAI’s comparison table (July 9, 2026).
What Is GPT-5.6? The Three-Tier Architecture Explained
GPT-5.6 introduces a naming convention change that matters: instead of tagging models as “mini,” “pro,” or “turbo,” OpenAI assigns celestial names — Sol, Terra, Luna — to what it calls “durable capability tiers that can advance on their own cadence.” The number (5.6) identifies the generation; the name identifies the tier. OpenAI’s official GPT-5.6 launch post published on July 9, 2026 confirms this framing.
The practical implication for developers is architectural: you no longer choose between one flagship and a cut-down mini. You route. Sol for tasks where marginal reasoning depth justifies the cost, Terra for the everyday default, Luna for high-volume pipelines where absolute benchmark scores matter less than throughput and budget. That’s a shift in how API teams budget inference.
All three models launched with identical core specifications:
| Specification | Value (all tiers) |
|---|---|
| Context window | 1,050,000 tokens |
| Maximum output tokens | 128,000 |
| Knowledge cutoff | February 16, 2026 |
| Modalities | Text + image input; text output |
| Prompt caching | Explicit breakpoints; 30-min minimum TTL |
| Cache write pricing | 1.25× uncached input rate |
| Cache read discount | 90% off cached-input rate |
| Long-context threshold | >272K input tokens billed at 2× input / 1.5× output |
Source: OpenAI API model page, GPT-5.6 Sol, retrieved July 18, 2026; OpenAI GPT-5.6 launch post, July 9, 2026.
GPT-5.6 Pricing Statistics: Sol vs Terra vs Luna
What Does GPT-5.6 Cost Per Million Tokens?
OpenAI published confirmed per-million-token pricing in the GPT-5.6 general availability announcement on July 9, 2026. These are the verified figures:
| Model | API string | Input ($/1M tokens) | Cached input | Output ($/1M tokens) |
|---|---|---|---|---|
| GPT-5.6 Sol | gpt-5.6-sol | $5.00 | $0.50 | $30.00 |
| GPT-5.6 Terra | gpt-5.6-terra | $2.50 | $0.25 | $15.00 |
| GPT-5.6 Luna | gpt-5.6-luna | $1.00 | $0.10 | $6.00 |
Source: OpenAI, GPT-5.6: Frontier intelligence that scales with your ambition, July 9, 2026; OpenAI API model docs, GPT-5.6 Sol, retrieved July 18, 2026.
The bare alias gpt-5.6 routes to Sol. The cache read discount of 90% reduces Sol’s effective cached input cost to $0.50/M and Luna’s to $0.10/M — a meaningful shift for prompt-heavy agent workflows that repeatedly pass the same system prompts.
Long-Context Pricing Threshold
Requests containing more than 272,000 input tokens trigger a surcharge: the entire request is billed at twice the normal input rate and 1.5 times the normal output rate. At Sol’s standard rates, that produces $10.00/$45.00 per million tokens for long-context requests. Teams routing entire codebases or multi-document corpora should account for this in cost modeling.
Axis Intelligence Research: Cost-Performance Compression Ratio (CPCR™)
Axis Intelligence Research introduces the Cost-Performance Compression Ratio (CPCR™) as a proprietary metric for quantifying how efficiently a model family compresses its performance gap relative to its price gap.
Formula:
CPCR™ = (Benchmark score of lower tier / Benchmark score of higher tier) ÷ (Price per output token of lower tier / Price per output token of higher tier)
Applied to GPT-5.6 (Coding Agent Index v1.1, output token pricing):
- Luna vs Sol: (74.6 / 80) ÷ ($6 / $30) = 0.9325 ÷ 0.2000 = CPCR™ = 4.66
- Terra vs Sol: (77.4 / 80) ÷ ($15 / $30) = 0.9675 ÷ 0.5000 = CPCR™ = 1.94
Interpretation: A CPCR™ above 1.0 indicates the lower-tier model delivers more benchmark performance per dollar than the gap between prices would suggest. Luna’s CPCR™ of 4.66 means that for every percentage point of coding benchmark score you sacrifice versus Sol, you save 4.66× the proportional output cost. Terra at 1.94 sits comfortably above 1.0 as well.
Axis Intelligence Research is the first publisher of the CPCR™ metric. All inputs sourced from OpenAI’s July 9, 2026 benchmark release and official pricing documentation. Python-verified arithmetic on benchmark and pricing values as stated.
GPT-5.6 Benchmark Statistics
Professional and Agentic Benchmarks
| Benchmark | GPT-5.6 Sol | GPT-5.6 Terra | GPT-5.6 Luna | GPT-5.5 | Claude Fable 5 | Source |
|---|---|---|---|---|---|---|
| Agents’ Last Exam | 52.7% | 50.4% | 50.3% | 46.9% | 40.5% | OpenAI, Jul 9 2026 |
| Agents’ Last Exam (with max reasoning) | 53.6% | — | — | — | — | OpenAI, Jul 9 2026 |
| AA Intelligence Index v4.1 | 58.9 | 55.0 | 51.2 | 54.8 | 59.9 | OpenAI, Jul 9 2026 |
| Management Consulting Tasks (internal) | 43.2% | 37.2% | 35.4% | 31.3% | 35.5% | OpenAI, Jul 9 2026 |
| Big Finance Bench | 53% | 51% | 36% | 49% | — | OpenAI, Jul 9 2026 |
Source: OpenAI GPT-5.6 launch post, July 9, 2026. Agents’ Last Exam is an evaluation across 55 professional fields run by long-horizon AI agents.
On Agents’ Last Exam, the Sol-to-Luna spread is only 2.4 percentage points (52.7% to 50.3%) at their respective standard reasoning levels — the narrowest performance gap in the family across any major benchmark category. That is the strongest argument for Luna as a starting point for agentic pipelines before escalating.
Coding Benchmarks
| Benchmark | Sol | Sol Ultra | Terra | Luna | GPT-5.5 | Fable 5 | Mythos 5 |
|---|---|---|---|---|---|---|---|
| AA Coding Agent Index v1.1 | 80 | — | 77.4 | 74.6 | 76.4 | 77.2 | — |
| SWE-Bench Pro | 64.6% | — | 63.4% | 62.7% | 59.4% | 80.0% | 80.3% |
| DeepSWE v1.1 | 72.7% | — | 69.6% | 67.2% | 67.0% | 69.7% | — |
| Terminal-Bench 2.1 | 88.8% | 91.9% | 87.4% | 84.7% | 85.6% | 83.1% | 88.0% |
Source: OpenAI GPT-5.6 launch post, July 9, 2026.
The SWE-Bench Pro gap deserves honest attention. Sol’s 64.6% trails both Claude Mythos 5 (80.3%) and Claude Fable 5 (80.0%) by roughly 15 points. SWE-Bench Pro tests the ability to produce correct pull requests against real GitHub repositories — production-grade bug fixing, not toy examples. For teams whose primary workload is fixing bugs in existing codebases, Claude’s advantage on this specific benchmark is material. The inference to draw isn’t that GPT-5.6 is a weaker coding model; it’s that benchmark aggregates obscure routing decisions. Sol wins on agents, pipelines, and agentic automation. Fable 5 and Mythos 5 win on repository-level pull request generation.
The AA Coding Agent Index tells a different story: Sol at 80, Terra at 77.4, Luna at 74.6, all above Fable 5 at 77.2. That index covers implementation agents, terminal use, and real codebases — a broader surface than SWE-Bench Pro alone.
### Why the Sol Ultra Mode Changes the Routing Calculus
Sol Ultra coordinates four subagents in parallel by default. The Terminal-Bench 2.1 result moves from 88.8% (standard Sol) to 91.9% (Sol Ultra). BrowseComp moves from 90.4% to 92.2%. These are not trivial lifts. But ultra is estimated to cost roughly 2–3× the base Sol rate, which on output-heavy tasks pushes beyond $60–$90 per million output tokens.
The benchmark return on ultra — roughly 3 additional percentage points on Terminal-Bench for 3× the cost — implies a CPCR™ of approximately 0.34. That sits well below 1.0, meaning you are paying more than proportionally for the marginal gain. Ultra makes sense where latency compression matters (parallel agents finish faster) or where the task has no meaningful partial solution (a vulnerability exploit either works or it doesn’t). It does not make sense as a default.
Computer Use Benchmarks
| Benchmark | Sol | Sol Ultra | Terra | Luna | GPT-5.5 | Mythos 5 | Opus 4.8 |
|---|---|---|---|---|---|---|---|
| OSWorld 2.0 | 62.6% | — | 50.2% | 45.6% | 47.5% | — | 54.8% |
| BrowseComp | 90.4% | 92.2% | 87.5% | 83.3% | 84.4% | 88.0% | 84.3% |
| BenchCAD | 70.6% | — | 62.3% | 63.1% | 44.4% | 38.4% | 27.3% |
Source: OpenAI GPT-5.6 launch post, July 9, 2026.
OSWorld 2.0 shows the widest tier gap in the family: Sol at 62.6% versus Luna at 45.6% — a 17-point spread. Computer-use tasks require spatial reasoning about UI state that doesn’t compress well to cheaper inference. On BenchCAD, Sol, Terra, and Luna all substantially outperform both Mythos 5 (38.4%) and Opus 4.8 (27.3%), a gap worth noting for engineering teams working with CAD-adjacent workflows.
Cybersecurity Benchmarks
| Benchmark | Sol | Sol Ultra | Terra | Luna | GPT-5.5 | Mythos 5 | Mythos Preview |
|---|---|---|---|---|---|---|---|
| Capture-the-Flag | 96.7% | — | 91.8% | 85.2% | 88.1% | — | — |
| SEC-Bench Pro | 71.2% | 74.3% | 57.7% | 48.9% | 45.8% | — | — |
| ExploitBench | 73.5% | — | 52.9% | 33.2% | 47.9% | 78.0% | 74.2% |
| ExploitGym (2h cap) | 24.9% | — | 23.2% | 12.4% | 15.1% | — | — |
| ExploitGym (6h cap) | 33.7% | — | — | — | — | — | — |
Source: OpenAI GPT-5.6 launch post, July 9, 2026. Note: cyber benchmarks evaluated with reduced safeguards per OpenAI’s published methodology.
ExploitBench shows the starkest tier divergence in the family: Sol at 73.5%, Luna at 33.2% — a 40-point spread. Cybersecurity tasks appear to require the full reasoning depth of Sol more than any other category tested. On ExploitBench, Claude Mythos 5 (78.0%) still leads Sol (73.5%) by 4.5 points. OpenAI’s safeguard stack for GPT-5.6 cybersecurity capabilities includes a government-reviewed trusted access program (OpenAI Daybreak) for defenders who need deeper access to defensive capabilities.
Academic and Science Benchmarks
| Benchmark | Sol | Terra | Luna | GPT-5.5 | Fable 5 | Mythos 5 |
|---|---|---|---|---|---|---|
| GPQA Diamond | 94.6% | 92.9% | 92.3% | 93.6% | 92.6% | 94.1% |
| FrontierMath Tier 1–3 (v2) | 89.0% | 84.9% | 78.6% | 85.3% | 87.0% | — |
| FrontierMath Tier 4 (v2) | 83.0% | 68.3% | 58.5% | 72.5% | 87.8% | — |
| GeneBench Pro | 28.7% | 23.3% | 10.8% | 12.0% | — | — |
| LifeSciBench | 59.9% | 56.0% | 51.2% | 50.4% | — | — |
| HealthBench Professional | 60.5% | 57.7% | 55.7% | 49.5% | 60.9% | — |
Source: OpenAI GPT-5.6 launch post, July 9, 2026.
FrontierMath Tier 4 is worth isolating. The hardest mathematical problems in this evaluation show Sol at 83.0%, Claude Fable 5 at 87.8%, and GPT-5.5 at 72.5%. Terra drops to 68.3% and Luna to 58.5%. At the very top of mathematical difficulty, the tier gradient steepens sharply: Luna captures only 70.5% of Sol’s Tier 4 score, while on most other benchmarks the ratio sits above 90%.
Abstract Reasoning
| Benchmark | Sol | Terra | Luna | GPT-5.5 | Opus 4.8 |
|---|---|---|---|---|---|
| ARC-AGI-3 | 7.78% | 0.80% | 0.18% | 0.43% | 1.50% |
Source: OpenAI GPT-5.6 launch post, July 9, 2026.
ARC-AGI-3 remains the benchmark where all frontier models fail at scale. Sol’s 7.78% is the highest result in the table above — a meaningful jump from GPT-5.5’s 0.43% and the 0.80–1.50% range of the other Claude models shown — but all values sit in the low single digits. No frontier model has demonstrated general abstract reasoning at scale on this evaluation as of July 2026.
GPT-5.6 Availability: Which Plans Get Which Tier?
Access to GPT-5.6 tiers is plan-gated on the consumer side. OpenAI’s Help Center article on GPT-5.6, retrieved July 18, 2026, confirms the following access matrix:
| Surface | Free/Go | Plus | Pro | Business | Enterprise |
|---|---|---|---|---|---|
| Standard ChatGPT | — | Sol (med/high reasoning) | Sol + Sol Pro | Sol (med/high) | Sol (admin-controlled) |
| ChatGPT Work | Terra | Sol/Terra/Luna | Sol/Terra/Luna + Ultra | Sol/Terra/Luna | Sol/Terra/Luna |
| Codex | Terra | Sol/Terra/Luna + Ultra | Sol/Terra/Luna + Ultra | Sol/Terra/Luna | Sol/Terra/Luna |
| OpenAI API | Sol/Terra/Luna | — | — | — | — |
Source: OpenAI Help Center, GPT-5.6 in ChatGPT, retrieved July 18, 2026; OpenAI GPT-5.6 launch post, July 9, 2026.
Terra and Luna are not selectable in standard ChatGPT conversations regardless of plan — they are available through ChatGPT Work, Codex, and the API. GPT-5.5 Instant remains the default model for everyday ChatGPT responses; GPT-5.6 Sol powers the reasoning effort tiers (medium, high, extra high).
The API is ungated: any OpenAI API account can call all three GPT-5.6 tiers without a waitlist as of July 9, 2026.
GPT-5.6 vs GPT-5.5: What Changed?
The GPT-5.5 to GPT-5.6 generation shift on key benchmarks, drawn from OpenAI’s published comparison tables:
| Benchmark | GPT-5.5 | GPT-5.6 Sol | Change |
|---|---|---|---|
| Agents’ Last Exam | 46.9% | 52.7% | +5.8 pp |
| AA Coding Agent Index | 76.4 | 80.0 | +3.6 pts |
| SWE-Bench Pro | 59.4% | 64.6% | +5.2 pp |
| Terminal-Bench 2.1 | 85.6% | 88.8% | +3.2 pp |
| DeepSWE v1.1 | 67.0% | 72.7% | +5.7 pp |
| OSWorld 2.0 | 47.5% | 62.6% | +15.1 pp |
| BrowseComp | 84.4% | 90.4% | +6.0 pp |
| ExploitBench | 47.9% | 73.5% | +25.6 pp |
| HealthBench Professional | 49.5% | 60.5% | +11.0 pp |
| GPQA Diamond | 93.6% | 94.6% | +1.0 pp |
Source: OpenAI GPT-5.6 launch post, July 9, 2026.
The ExploitBench jump (+25.6 percentage points) and the OSWorld 2.0 jump (+15.1 pp) are the two largest single-generation leaps in the table. Cybersecurity capability and computer-use autonomy advanced the most between generations. GPQA Diamond improved by 1.0 point — the smallest gain, consistent with academic knowledge benchmarks approaching a ceiling where further gains require qualitatively different architectural choices.
GPT-5.6 Reasoning Modes and the Ultra Architecture
GPT-5.6 introduces six reasoning effort levels: none, low, medium, high, extra high, and max. Beyond max, Sol gains access to ultra mode, which coordinates four parallel subagents by default, with 16-agent configurations available for specific evaluations (BrowseComp, SEC-Bench Pro).
Ultra mode is explicitly not a separate model — it is a scheduling layer on top of Sol. Developers access ultra-like experiences via the multi-agent beta in the Responses API. In ChatGPT, ultra is available to Pro and Enterprise users in Work mode; in Codex, it extends to Plus plans as well.
The 700,000 NVIDIA A100-equivalent GPU hours OpenAI reported spending on automated red-teaming for GPT-5.6’s safeguards (per the GPT-5.6 launch post and the preview system card) represents the most compute-intensive safety evaluation OpenAI has published to date on a single model family.
GPT-5.6 Self-Improvement Benchmarks
| Benchmark | Sol | Terra | Luna | GPT-5.5 |
|---|---|---|---|---|
| RSI Index | 57.9 | 56.3 | 41.9 | 41.7 |
| Internal Research Debugging | 68.3% | 67.8% | 50.8% | 50.0% |
| KernelGen 1P | 61.1% | 49.2% | 22.4% | 29.3% |
| NanoGPT | 9.69% | 14.5% | 1.66% | 2.65% |
Source: OpenAI GPT-5.6 launch post, July 9, 2026. RSI Index measures GPT-5.6’s capability to assist with AI research tasks; OpenAI describes a 16.2-point improvement in aggregate RSI capability over GPT-5.5.
KernelGen 1P shows Sol at 61.1% versus GPT-5.5 at 29.3% — more than double. Luna, however, underperforms GPT-5.5 on this benchmark (22.4% vs 29.3%). For GPU kernel optimization specifically, using Luna would be a material regression. The NanoGPT benchmark (training small language models) shows Terra (14.5%) and Sol (9.69%) in an atypical reversal — the only category across published results where Terra outperforms Sol. OpenAI has not explained this inversion.
GPT-5.6 Long-Context Benchmarks
| Benchmark | Sol | Terra | Luna | GPT-5.5 | Mythos 5 |
|---|---|---|---|---|---|
| MRCR v2 (8-needle, 256K–512K) | 91.5% | 89.6% | 41.3% | 81.5% | — |
| MRCR v2 (8-needle, 512K–1M) | 73.8% | 72.5% | 41.3% | 74.0% | — |
| GraphWalks BFS 256K (F1) | 90.7% | 76.9% | 81.3% | 73.7% | 91.1% |
| GraphWalks BFS 1M (F1) | 77.1% | 71.2% | 51.2% | 45.4% | 79.4% |
Source: OpenAI GPT-5.6 launch post, July 9, 2026. MRCR v2 tests 8-needle retrieval within long documents; GraphWalks BFS tests breadth-first search traversal in large graphs.
Luna’s long-context performance shows the most dramatic cliff in the family: 41.3% on MRCR v2 across both the 256K–512K and 512K–1M ranges, compared with Sol’s 91.5% and 73.8% respectively. The score doesn’t degrade gradually — it stays flat at 41.3% regardless of whether the context is 256K or 1M tokens. For any workload involving document analysis, large codebase reasoning, or multi-document synthesis above roughly 100K tokens, Luna is the wrong tier. Sol and Terra hold meaningful retrieval capability into the 1M-token range.
Axis Intelligence Research: GPT-5.6 Tier Selection Framework (ATSF™)
Axis Intelligence Research synthesizes published benchmark data into the Axis Tier Selection Framework (ATSF™) — a routing guide for the most common AI workload categories. Inputs: Coding Agent Index v1.1, SWE-Bench Pro, Terminal-Bench 2.1, MRCR v2 (long context), ExploitBench, and HealthBench Professional results from OpenAI’s July 9, 2026 benchmark table. Output: a recommended default tier per workload category with a documented escalation trigger.
| Workload category | Default tier | Escalation trigger | Do not use |
|---|---|---|---|
| High-volume code generation, classification | Luna | Task failure rate >15% | Sol (cost) |
| Long-context document analysis (>100K tokens) | Terra | — | Luna (MRCR cliff at 41.3%) |
| Agentic multi-step engineering | Terra | Complex persistence + tool chains | Luna |
| Frontier reasoning + research | Sol | — | — |
| Computer use / UI automation | Sol | — | Luna (17-pt gap vs Sol on OSWorld) |
| Production PR / repository-level bug fixes | Claude Fable 5 or Mythos 5 | — | GPT-5.6 any tier (15-pt SWE deficit) |
| Security / vulnerability research (defensive) | Sol | Multi-agent parallel (ultra) | Luna (40-pt ExploitBench gap) |
| High-volume low-stakes pipelines | Luna | — | Sol (5× output cost premium) |
ATSF™ is original to Axis Intelligence Research. Framework synthesizes benchmark data from OpenAI’s official July 9, 2026 launch documentation; no inputs are derived from model memory or external blogs.
Methodology
Data collection: All benchmark figures in this article are sourced from OpenAI’s general availability announcement at openai.com/index/gpt-5-6/ (published July 9, 2026, retrieved July 18, 2026) and confirmed against the OpenAI API documentation at developers.openai.com/api/docs/models/gpt-5.6-sol (retrieved July 18, 2026). The OpenAI Help Center availability matrix was retrieved from help.openai.com on July 18, 2026.
Proprietary metrics:
CPCR™ (Cost-Performance Compression Ratio): Formula: (Benchmark_lower ÷ Benchmark_higher) ÷ (Price_per_output_token_lower ÷ Price_per_output_token_higher). Applied to Coding Agent Index v1.1 scores (source: OpenAI, July 9, 2026) and output pricing per million tokens (source: OpenAI, July 9, 2026). Luna vs Sol: (74.6 ÷ 80) ÷ ($6 ÷ $30) = 0.9325 ÷ 0.2000 = 4.66. Terra vs Sol: (77.4 ÷ 80) ÷ ($15 ÷ $30) = 0.9675 ÷ 0.5000 = 1.94.
ATSF™ (Axis Tier Selection Framework): Routing recommendations built from the intersection of six benchmark categories (Coding Agent Index, SWE-Bench Pro, Terminal-Bench 2.1, MRCR v2 long-context, ExploitBench, HealthBench Professional). All inputs from OpenAI’s published July 9, 2026 benchmark table. Routing logic determined by Axis Intelligence Research; no vendor claims incorporated without sourced benchmark support.
Limitations: All benchmark results are OpenAI-published. Independent third-party replication of the full benchmark suite had not been publicly completed as of the article’s publication date of July 18, 2026. METR autonomous task evaluations (separate from OpenAI’s published figures) have raised questions about frontier model benchmark consistency; readers should treat aggregate scores as directional rather than definitive.
Scope: This article covers GPT-5.6 Sol, Terra, and Luna at general availability (July 9, 2026). It does not cover GPT-5.5, GPT-5.4, or earlier generations except as comparison points in tables sourced directly from OpenAI.
About This Dataset
The data powering this article is available as a free download under CC BY 4.0. The dataset covers GPT-5.6 benchmark scores across 30+ evaluations, pricing by tier, API specifications, and Axis Intelligence Research’s CPCR™ and ATSF™ computations. Each row carries full source provenance (organization, document, URL, retrieval date).
License: Creative Commons Attribution 4.0 International (CC BY 4.0) Citation: Axis Intelligence Research. “GPT-5.6 Statistics 2026: Sol, Terra & Luna Benchmarks, Pricing, and Model Specifications.” axis-intelligence.com, July 18, 2026. https://axis-intelligence.com/gpt-5-6-statistics/
APA: Axis Intelligence Research. (2026, July 18). GPT-5.6 Statistics 2026: Sol, Terra & Luna Benchmarks, Pricing, and Model Specifications. Axis Intelligence Research. https://axis-intelligence.com/gpt-5-6-statistics/
MLA: Axis Intelligence Research. “GPT-5.6 Statistics 2026: Sol, Terra & Luna Benchmarks, Pricing, and Model Specifications.” Axis Intelligence Research, 18 July 2026, axis-intelligence.com/gpt-5-6-statistics/.
Chicago: Axis Intelligence Research. “GPT-5.6 Statistics 2026: Sol, Terra & Luna Benchmarks, Pricing, and Model Specifications.” Axis Intelligence Research. July 18, 2026. https://axis-intelligence.com/gpt-5-6-statistics/.
Frequently Asked Questions
What is GPT-5.6?
GPT-5.6 is OpenAI’s seventh GPT-5-series model generation, launched to general availability on July 9, 2026. It ships as three models — Sol (flagship), Terra (balanced), and Luna (fast/low-cost) — rather than a single model plus cut-down variants. All three share a 1,050,000-token context window, 128,000 maximum output tokens, and a February 16, 2026 knowledge cutoff.
What are Sol, Terra, and Luna?
Sol, Terra, and Luna are OpenAI’s names for what it calls “durable capability tiers” within the GPT-5.6 generation. The number (5.6) identifies the generation; the celestial name identifies the tier. Sol is the most capable, Terra is balanced, and Luna is the fastest and cheapest. Each tier carries its own API model string: gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna. The bare alias gpt-5.6 routes to Sol.
How much does GPT-5.6 cost?
Per million tokens as of the July 9, 2026 general availability: Sol is $5.00 input / $30.00 output; Terra is $2.50 input / $15.00 output; Luna is $1.00 input / $6.00 output. Prompt caching discounts cached input reads by 90%, which meaningfully reduces effective input costs for agent workflows that repeat system prompts. Requests over 272,000 input tokens are billed at double the standard input rate.
How does GPT-5.6 Sol compare to Claude Fable 5?
Sol leads Fable 5 on Agents’ Last Exam (53.6% vs 40.5%), the AA Coding Agent Index (80 vs 77.2), Terminal-Bench 2.1 (88.8% vs 83.1%), BrowseComp (90.4% vs 84.3%), and OSWorld 2.0 (62.6% vs not reported by Fable 5). Fable 5 leads Sol on SWE-Bench Pro (80.0% vs 64.6%) and the AA Intelligence Index (59.9 vs 58.9). The practical routing answer is: Sol for agentic multi-step tasks and computer use; Claude for production repository-level pull request generation. All figures from OpenAI’s published July 9, 2026 benchmark table.
Is GPT-5.6 Luna good enough for most tasks?
For the majority of everyday coding and text tasks, Luna delivers benchmark performance within 6–7% of Sol at one-fifth the output cost. The CPCR™ calculated by Axis Intelligence Research is 4.66 — meaning Luna’s cost-performance ratio is strongly favorable versus Sol for tasks that don’t require deep reasoning or long-context recall. Luna becomes the wrong choice for long-context retrieval above roughly 100K tokens (MRCR v2 score drops to 41.3%, vs Sol’s 91.5%), security-research tasks (40-point ExploitBench gap), and computer-use automation (17-point OSWorld 2.0 gap).
What is GPT-5.6 Sol ultra mode?
Ultra is a compute setting — not a separate model — that coordinates four GPT-5.6 Sol subagents in parallel. On Terminal-Bench 2.1, ultra moves the score from 88.8% to 91.9%; on BrowseComp, from 90.4% to 92.2%. In the API, ultra-like behavior is accessible through the multi-agent beta in the Responses API. Ultra costs roughly 2–3× the standard Sol rate, making it unsuitable as a default; it is best reserved for tasks where parallelism directly reduces wall-clock time or where near-ceiling performance on a specific evaluation is required.
When was GPT-5.6 released?
OpenAI previewed GPT-5.6 Sol in a limited preview on June 26, 2026, restricted to government-vetted partners. General availability launched on July 9, 2026, across ChatGPT, Codex, and the OpenAI API. The rollout completed globally within 24 hours of the announcement.
What is the GPT-5.6 context window?
All three GPT-5.6 tiers — Sol, Terra, and Luna — share a 1,050,000-token context window and 128,000-token maximum output length, per the OpenAI API model documentation retrieved July 18, 2026. Requests containing more than 272,000 input tokens are billed at a long-context surcharge rate.
What is the GPT-5.6 knowledge cutoff?
All three GPT-5.6 tiers share a knowledge cutoff of February 16, 2026, per OpenAI’s official API model documentation. Events and data after that date are not reflected in the models’ training.