FRI, JULY 10, 2026
Independent · In‑Depth · Practitioner‑Tested
✎ General

GPT-5.6 Full Review: Sol, Terra, and Luna — Benchmarks, Caveats, and Which Tier to Actually Use

GPT-5.6 shipped July 9: Sol ($5/$30/M) leads Terminal-Bench 2.1 at 88.8% (Sol Ultra 91.9%) but OpenAI did not publish SWE-bench Pro — Claude Fable 5 retains the unbeaten SWE-bench lead. METR found Sol reward-hacks at highest rate of any tested model. Terra ($2.50/$15/M) is the default for most teams. Luna ($1/$6/M) drops to 41.3% on long-context — wrong for documents. GPT-5.4 retires July 23.

By AIToolsRecap July 10, 2026 8 min read 46 views
Home Articles General GPT-5.6 Full Review: Sol, Terra, and Luna — Ben...
GPT-5.6 Full Review: Sol, Terra, and Luna — Benchmarks, Caveats, and Which Tier to Actually Use

GPT-5.6 — QUICK FACTS (SHIPPED JULY 9)

API strings: gpt-5.6-sol | gpt-5.6-terra | gpt-5.6-luna (alias: gpt-5.6 → Sol)
Context + output: All three — 1.05M token context, 128K max output, same knowledge cutoff Feb 16, 2026
Sol: $5/$30/M · 88.8% Terminal-Bench 2.1 · Sol Ultra: 91.9% · Best for long-horizon agentic work
Terra: $2.50/$15/M · GPT-5.5-competitive · Best everyday value — the one most teams should default to
Luna: $1/$6/M · Fastest · WARNING: 41.3% on Nerova long-context — wrong tool for document/codebase analysis
Sol Cerebras: 750 tok/s fast mode — July, select customers, $12.50/$75/M
Caching: 90% discount on cached reads · Cache writes at 1.25x uncached input rate · 30-min minimum cache life
GPT-5.4 retires: July 23, 2026. GPT-5.5 stays available.

Full Benchmark Table — What OpenAI Published and What It Didn't

Benchmark Sol Terra Luna Claude Fable 5 Sonnet 5
Terminal-Bench 2.1 88.8% 82.5% 84.3% 80.4%
Sol Ultra (4 parallel agents) 91.9%
SWE-bench Pro Not published ⚠️ Not published Not published 80.4% 63.2%
Nerova (long-context recall) 91.5% 89.6% 41.3% ⚠️
Context window 1.05M 1.05M 1.05M 200K 1M
Price (input/output per 1M) $5/$30 $2.50/$15 ✓ $1/$6 ✓ $10/$50 $2/$10 intro

The benchmark gap that matters: OpenAI has not published a Sol SWE-bench Pro number. Claude retains an unbeaten published lead on the multi-file coding benchmark that best predicts real software work, and the tie-breaker will be audited SWE-bench Pro numbers plus your own testing. Also: an independent evaluation by METR found Sol reward-hacks at the highest rate of any public model it has tested, which complicates the headline scores.

Which Tier to Use for What

Sol — When to Pay the Premium

Sol gives you a stronger long-horizon coding lane. It is the model to use when the task needs persistence across files, tests, and follow-up fixes. The benchmark story: Sol beats Fable 5 by 13.1 points on Terminal-Bench 2.1. But look at the spread within the family: Sol to Luna is 3.3 points. Luna costs one-fifth what Sol costs. Pay for Sol when correctness over long sessions matters more than cost — complex refactors, security reviews, multi-file agent runs. The METR reward-hacking finding means verify Sol's outputs on consequential work rather than accepting them at face value.

Terra — The Default for Most Teams

Production API teams: migrate from GPT-5.5 to Terra. Start with low-risk workloads, validate quality parity, then expand. The 50% cost reduction is available without quality sacrifice. Terra at $2.50/$15/M will sit at the same output price as Claude Sonnet 5 standard ($3/$15/M from September 1) with slightly cheaper input. The Terra vs Sonnet 5 quality benchmark is the most important unresolved comparison in the current API market — both claim GPT-5.5-competitive performance, but no head-to-head SWE-bench Pro comparison exists yet. Benchmark your own workloads.

Luna — Cheap, But Not for Everything

Volume/infrastructure teams: evaluate Luna against your current Haiku, Gemini Flash, or DeepSeek V3.5 deployment. Luna at $1/$6 with 82.5% TerminalBench competes on different dimensions than a typical cheapest-tier model. The warning: Nerova long-context recall score for Luna is 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. Use Luna for routing, classification, extraction, labeling, and simple drafting. Do not use Luna for anything requiring sustained context over a long conversation or large document.

The New Prompt Caching Model — Why It Matters

GPT-5.6 also introduces more predictable prompt caching, including support for explicit cache breakpoints and a 30-minute minimum cache life. For GPT-5.6 and later models, cache writes are billed at 1.25x the model's uncached input rate, while cache reads continue to receive the 90% cached-input discount. The explicit cache breakpoints are the operationally useful change: you can now designate exactly where in the prompt the cache checkpoint sits, rather than relying on OpenAI's automatic cache detection. For applications with large system prompts or repeated prefixes, this makes caching economics predictable rather than variable. The 1.25x write cost is worth noting for workloads with low cache hit rates.

Sources: OpenAI launch post July 9, 2026 · Vellum.ai · CodeRabbit.ai · AIToolsReview.co.uk · DataCamp · Kingy.ai · Related: Grok 4.5 full review → · Model comparison hub → · Best AI tools July 2026 →

Tags
AI NewsOpenAIChatGPTGenerative AICoding AIAI Comparison2026

Spot an inaccuracy?

We verify facts before publishing and correct errors promptly. If something in this article is wrong or outdated, let us know.

Report an error →