GPT-5.6 Sol Pricing: API Cost vs Self-Hosted LLMs (2026)

GPT-5.6 Sol pricing is $5 per million input tokens and $30 per million output tokens, the same rate it launched with. What moved is everything under it: on July 30, 2026, three weeks after general availability, OpenAI cut Terra's price 20% and Luna's price 80%, funded in part by Sol rewriting its own production inference kernels (Digital Today). That's the real story right now: the flagship tier stood still while the two cheaper tiers got dramatically cheaper, and the open-weight models racing to match Terra-tier quality, GLM-5.2 and Kimi K3 among them, now do it at real self-hosting cost instead of Sol's still-expensive per-token rate.

This post breaks down the full Luna/Terra/Sol pricing structure including the caching and long-context mechanics most calculators skip, works out the daily token volume where self-hosting actually beats Sol, and covers what you give up on data control by routing through OpenAI's API instead of running the weights yourself.

GPT-5.6 Sol Pricing vs Terra and Luna: The Three Capability Tiers

GPT-5.6 shipped as three separate models, not three settings on one model. OpenAI treats the version number as the generation and Sol, Terra, and Luna as "durable capability tiers that can advance on their own cadence" (Vellum), meaning a future Sol could ship stronger without the whole family renaming around it.

The rollout itself was staged: a limited preview to a small set of trusted partners on June 26, 2026, followed by general availability in ChatGPT, Codex, and the API on July 9, 2026 (ExplainX). Current pricing per million tokens, after the July 30 cut:

All three tiers share a 1,050,000-token context window, so the tier you pick is a cost and quality decision, not a context-length one (eesel). On raw capability, the gap between tiers is smaller than the 25x price spread between Luna and Sol suggests: SWE-Bench Pro scores run 64.6% for Sol, 63.4% for Terra, and 62.7% for Luna, a 1.9-point spread across the whole family (llm-boss.com). Sol still leads on the harder frontier benchmarks, GPQA Diamond at 94.6%, tied for the top of that particular table, but Terra and Luna are closer to Sol than their price tags imply.

What Changed on July 30: Sol Rewrote Its Own Inference Stack to Fund the Cut

Direct answer: on July 30, 2026, OpenAI cut Luna's price 80% and Terra's price 20% while leaving Sol at $5/$30, and it credited the cut to Sol itself: OpenAI had Sol working inside Codex on its own production GPU kernels, rewriting them in Triton and Gluon (OpenAI's open-source GPU programming languages), cutting end-to-end serving costs 20% and lifting token-generation efficiency 15% through better draft-model performance in speculative decoding (Digital Today).

Sam Altman framed the strategy plainly: "we want to offer the best price/intelligence tradeoff at every level" (The Rundown AI). In practice that means the efficiency gains got routed almost entirely into the tiers competing hardest against cheap open-weight and rival-API alternatives, Luna and Terra, while Sol's price held because it isn't the tier under the most competitive pressure. If your workload already runs on Sol for its frontier quality, the July 30 cut doesn't touch your bill at all. If you were running Terra or Luna, or considering them, the math changed materially in three weeks.

Cache Reads, Cache Writes, and the Long-Context Surcharge

Prompt caching is where the headline $5/$30 number stops describing what you actually pay. GPT-5.6 replaced the older implicit caching model with explicit cache breakpoints: a cache write costs 1.25x the standard input rate with a 30-minute minimum cache life, and a cache read (hit) gets a 90% discount off standard input. For Sol, that's a $6.25/M write and a $0.50/M read; for Terra, $2.50/M and $0.20/M; for Luna, $0.25/M and $0.02/M (eesel). Writing a prefix costs more than sending it once uncached, but every reuse after that first write nets a real saving, so any workload that repeats a system prompt or tool schema across turns should be hitting cache reads, not paying standard input on every call.

The other mechanic worth budgeting for: requests with more than 272,000 input tokens get priced at 2x the input rate and 1.5x the output rate for the entire request, not just the tokens past the threshold (eesel). A 500K-token Sol request isn't $5/$30: it's effectively $10/$45 for that call. If your workload occasionally spikes into long-context territory, model that surcharge explicitly rather than assuming a flat per-token rate scales linearly with context length.

Cost Per Million Tokens at Each Tier (an LLM API Pricing Comparison)

Direct answer: as of August 2026, GPT-5.6 Sol is one of the more expensive flagship rates in an LLM API pricing comparison at $5.00/M input and $30.00/M output, well above Claude Opus 4.8's $5/$25 and roughly double Grok 4.5's $2/$6. Terra, at $2/$12 after the July 30 cut, now undercuts several rival mid-tier models outright, and Luna's $0.20/$1.20 puts it in the same bracket as budget specialist APIs rather than flagship pricing.

For the full cross-provider table, including Gemini, DeepSeek, and Mistral rates, see our LLM API pricing comparison. This post goes deeper on GPT-5.6 specifically: the tier mechanics, the self-hosting breakeven math, and the privacy tradeoffs of routing through OpenAI's API.

Blended Rate by Input/Output Split for Sol, Terra, and Luna

Blended cost per million tokens = (input share x input price) + (output share x output price). Most production workloads run more input than output. At an 80/20 split:

Sol:   (0.8 x $5.00) + (0.2 x $30.00) = $4.00 + $6.00  = $10.00/M
Terra: (0.8 x $2.00) + (0.2 x $12.00) = $1.60 + $2.40  = $4.00/M
Luna:  (0.8 x $0.20) + (0.2 x $1.20)  = $0.16 + $0.24  = $0.40/M

At a more output-heavy 60/40 split, all three climb:

Sol:   (0.6 x $5.00) + (0.4 x $30.00) = $3.00 + $12.00 = $15.00/M
Terra: (0.6 x $2.00) + (0.4 x $12.00) = $1.20 + $4.80  = $6.00/M
Luna:  (0.6 x $0.20) + (0.4 x $1.20)  = $0.12 + $0.48  = $0.60/M

Both ignore caching. If a meaningful share of your input hits the 90%-off cache rate instead of standard input, your real blended cost on any tier drops well below these numbers, and every crossover point below moves out further in the API's favor. Run these formulas against your own production logs, not a generic assumption; the input:output ratio alone can swing the blended rate by 3-5x.

The Daily Token Volume Where Self-Hosting Beats Sol

Direct answer: at an 80/20 split, Sol blends to $10.00/M tokens uncached. Neither DeepSeek V4-Flash nor GLM-5.2 fits on a single GPU, each is a large MoE model whose full parameter count has to sit resident in VRAM, so the real comparison is against a 4-to-8-GPU cluster. On a budget spot cluster, breakeven lands around 20-24M tokens/day. On a production on-demand cluster, it moves out to roughly 40-81M tokens/day depending on which model and context length you need.

Self-hosted cost per token comes down to cluster price per hour divided by throughput, so the cluster is a fixed cost regardless of how many tokens you push through it. Live Spheron GPU pricing, fetched 9 Aug 2026:

Multiplying those per-GPU rates by the minimum cluster size each model needs gives the daily infrastructure cost:

At the $10.00/M blended Sol rate (80/20, uncached), those cluster costs translate into daily token crossovers:

Worth calling out: those same crossovers move dramatically against self-hosting if you're actually running Terra rather than Sol. Terra's post-cut $4.00/M blended rate means the same $810/day production cluster doesn't break even until roughly 202M tokens/day, nearly 2.5x further out than against Sol. Self-hosting only clears the bar quickly against the expensive tier; against a $2/M-input tier that already got a 20% cut, the case for renting GPUs instead of calling the API needs genuinely heavy, sustained volume.

Pricing fluctuates based on GPU availability. The prices above are based on 09 Aug 2026 and may have changed. Check current GPU pricing → for live rates.

Where Self-Hosted Open Models Now Match Terra-Tier Quality

Cost is only half of what decides this. The reason Terra and Luna's price cuts matter as much as they do is that open-weight quality has closed most of the gap that used to make paying for Sol's tier an easy call for mid-range work.

GLM-5.2 and DeepSeek V4-Flash on the Intelligence Index

On the Artificial Analysis Intelligence Index, GPT-5.6 Terra (max) scores 57, sitting well above the 33 median for similarly-priced reasoning models (Artificial Analysis). Two open-weight models now sit right next to it. Kimi K3 (max) scores 60, actually edging past Terra outright, and well above the 26 median for its own open-weight peer group (Artificial Analysis). GLM-5.2 (max) scores 53, a few points behind Terra but similarly well clear of its own 26 peer median (Artificial Analysis). DeepSeek V4-Flash 0731 (max), the lighter-hardware option of the three, scores 52 on the same index, also well above its peer median of 26 (Artificial Analysis).

That's a materially different picture than it was even a generation ago: you're no longer trading away most of the model's intelligence to get out from under a per-token API bill. Sol keeps a real edge at the very top, on the Artificial Analysis Coding Agent Index, Sol (max) running in Codex scores 80, the top mark on that benchmark, at a per-task cost roughly 40% below Claude Fable 5 (max) and roughly 10% below Claude Opus 4.8 (max) (Artificial Analysis). If your workload genuinely needs that tier of coding-agent performance, none of the open-weight options here replace it. But for the broad middle of production inference work, classification, RAG, summarization, general agentic tool use, Terra-tier quality is now available on hardware you control.

GPU Cluster Requirements and Cost Per Day

Neither GLM-5.2 nor DeepSeek V4-Flash is a single-GPU model, which is the most common misconception in MoE serving: only the active parameter count drives compute per token, but the router can send any token to any expert, so every expert's weights have to sit resident in VRAM regardless of how few are active on a given pass.

DeepSeek V4-Flash's 284B total parameters need roughly 284GB of VRAM at FP8, fitting 4x H100 SXM5 (320GB total) for short context up to 32K tokens, or 4x H200 SXM5 (564GB total) if you need the full 1M-token window with real KV cache headroom; the DeepSeek V4-Flash deployment guide covers the vLLM setup and the CSA/HCA sparse-attention flags that make 1M-token serving viable at all. GLM-5.2's 744B total parameters need at least 4x H200 at AWQ INT4 as a floor, or the full 8x H200 (or 8x B200) FP8 cluster for production serving at its 1M-token context; the GLM-5.2 deployment guide has the exact tensor-parallel and expert-parallel flags for both tiers. For the general VRAM math across other model sizes and quantization formats, our VRAM tier guide is a useful companion reference before sizing a cluster for either model.

Privacy Tradeoffs of Routing Through OpenAI's API

Every prompt sent to Sol, Terra, or Luna leaves your infrastructure and lands on OpenAI's servers. What happens to it there, and how much control you actually have over that, is a separate question from cost, and it's worth answering before committing to either path.

Default 30-Day Retention vs Zero Data Retention (Enterprise-Only)

Direct answer: by default, OpenAI generates abuse-monitoring logs for all API usage and retains them for up to 30 days, then deletes them unless longer retention is legally required or reasonably necessary to protect the service (OpenAI developer docs). Zero Data Retention (ZDR) is not a setting available to standard accounts. It's an approved, negotiated arrangement: eligible customers get their content excluded from abuse-monitoring logs entirely by applying for ZDR or Modified Abuse Monitoring controls, subject to OpenAI's review and additional requirements, and the path to eligibility runs through OpenAI's sales team, not a dashboard toggle (OpenAI developer docs).

That distinction matters if you've been assuming "we can request ZDR" is a formality. It isn't automatic, and it isn't offered on a self-serve pay-as-you-go plan the way some competitors structure their enterprise tiers. If your compliance requirement is "customer prompts never get logged, full stop," you need the ZDR approval in hand before you architect around it, not after.

Self-hosting sidesteps the retention question entirely for the data itself: prompts and completions never leave infrastructure you control. It doesn't sidestep everything else: you still own patching, access control, and audit logging on whatever GPU instances you run, and "self-hosted" only means something for compliance if the underlying infrastructure is actually isolated. On Spheron, instances run with SSH root access, no shared GPU tenancy, and no prompt or completion logging on the provider side, which is the baseline worth confirming before treating self-hosting as an automatic compliance win. Spheron's docs cover instance provisioning and SSH setup if you're standing up the self-hosted side of this comparison for the first time.

Decision Framework: When Sol Is Still Worth It

If you're weighing this same tradeoff against a different provider, our Claude Opus 4.8 API pricing vs self-hosted LLMs breakdown runs the same framework at a comparable Sol-tier price point, and Grok 4.5 API pricing vs self-hosted LLMs covers a cheaper API where the self-hosting case is harder to make. OpenAI's separate GPT-6 line has its own GPT-6 vs self-hosted LLMs analysis if that's the model you're actually comparing against. Don't conflate the two: they're different generations with different pricing. And if you're already committed to leaving the OpenAI API rather than just running the numbers, our migration guide from OpenAI to self-hosted LLMs covers the vLLM cutover plan and the tool-calling gaps worth testing before you move production traffic.

If your GPT-5.6 Sol traffic is clearing tens of millions of tokens a day and prompt caching isn't bringing the blended cost down enough, it's worth benchmarking GLM-5.2 or DeepSeek V4-Flash against your own eval set on a rented H100 or H200 cluster.

Check H100 availability → | Spheron H200 instances → | Check current GPU pricing →