October 2026. When this deep dive first published in July, GPT-5.6 was one flagship with three names attached and a middle tier that looked like the deal of the year. The lineup is unchanged; the price card is not. As of the October 9, 2026 llm-stats snapshot, Sol leads that board's composite at 0.590, at $5 in / $30 out per million tokens, with Terra at 0.550 ($2/$12) and Luna at 0.510 ($0.20/$1.20). Two of the three tiers got cheaper — Luna spectacularly so — and the October relay scores all three tiers together on one board, which turns OpenAI's July "near-peak" claim for the cheap tiers into something checkable. The launch benchmark record has not been re-measured and still reads as measured, the weak spot is still the weak spot, and the top of the text leaderboard is now a two-board argument this article will not settle by averaging. What changed, what did not, and which tier deserves which job.
The family, unchanged since July
GPT-5.6 has been globally available since July 9-13, after a June 26 preview to trusted partners and a US government security review required by the June 2 executive order. Knowledge cutoff February 16, 2026; context window 1.05M tokens; 128K max output. The naming is still the part people get wrong: 5.6 is the generation, and the celestial body is the tier. Sol is the flagship, Terra the balanced daily driver, Luna the volume tier, and each iterates independently, so the next release is as likely to be a Terra 5.6.1 as a monolithic 5.7.
The effort dial remains the product's real interface: five reasoning_effort levels in ChatGPT (Instant, Medium, High, Extra High, Pro) plus Ultra, the mode that runs four parallel sub-agents on the hardest tasks. Medium and Ultra can differ by 5-10x in tokens for the same query, which made the dial the biggest cost knob in July and makes it a bigger one now that the tiers themselves span 25x. Positioning holds too. Sol for tasks where a failure costs real money. Terra for most of what most people do. Luna for the ten-million-call invoice.
What the October card did to the price math
Input and output, per million tokens:
| Tier | October 9, 2026 | July 2026 | |---|---|---| | GPT-5.6 Sol | $5 / $30 | $5 / $30 | | GPT-5.6 Terra | $2 / $12 | $2.50 / $15 | | GPT-5.6 Luna | $0.20 / $1.20 | $1 / $6 |
The right-hand column is July pricing, since revised. It stays in the table so the change stays auditable, and it should not be quoted anywhere else. Sol held. Terra came down about 20% on both sides of the token. Luna fell 80%, from $1/$6 to $0.20/$1.20, and with it the spread between flagship and cheap tier widened from 5x to 25x on input. Dropping from Sol to Terra saved $2.50 per million input tokens in July; the same drop saves $3 now, which is the quiet story of this card. OpenAI cut the middle and the floor and left the flagship's price exactly where it was.
The fine print survived the repricing. Cache writes bill at 1.25x input, and GPT-5.6 is the first OpenAI generation to charge for them at all; at launch, one developer reported Luna conversations costing roughly 96% more than GPT-5.4-mini, almost entirely from cache-write accounting. Any single request over 272K tokens bills the whole request at 2x input and 1.5x output, so the 1.05M window is not free at its top end. Ultra stays gated to Pro/Enterprise in Work and Plus+ in Codex, as does Sol's Pro effort tier, and the Fast speed tier runs at 1.5x rate with faster quota burn. A card this cheap deserves a footnote read.
The three tiers, scored on one board
| Tier | llm-stats composite (2026-10-09) | Price in / out per MTok | |---|---|---| | GPT-5.6 Sol | 0.590 | $5 / $30 | | GPT-5.6 Terra | 0.550 | $2 / $12 | | GPT-5.6 Luna | 0.510 | $0.20 / $1.20 |
The composite runs on its own 0-to-1 scale, so those three numbers compare to each other and to nothing else in this article; do not read them against the ~60-point Artificial Analysis index below. What the board shows is a tightly packed family. Terra sits 0.040 behind Sol at 40% of the input price, and Luna sits 0.080 behind the flagship at 4% of it. That closeness is the headline of the October snapshot. In July, the near-peak framing for Terra and Luna was a vendor claim you could only take on faith; this relay scores all three tiers together, and it reads like the vendor was, if anything, understating it. Twenty-five times the price buys 0.080 points here. Whether that is a deal depends entirely on how many of your tasks were Luna tasks all along.
The launch benchmark record still stands
Benchmarks move slower than price cards. Nothing in the October snapshot re-measures them, so the July launch numbers remain the only published per-benchmark record for this family, presented with their vintage attached (Sol unless noted):
- Terminal-Bench 2.1 (command-line agent work): 88.8% in Max, 91.9% in Ultra. At launch that led Mythos 5 (~88.0%) and Fable 5 (84.3%).
- Agents' Last Exam (long-horizon workflows across 55 domains): 53.6% in code mode, the only model over 50 at the time, +13.1 over Fable 5 (40.5%) and +8.4 over Opus 4.8 (45.2%).
- BrowseComp (web research): 90.4%.
- CyberGym (security): 84.5%, edging Mythos 5 (83.8%) on roughly a third of the output tokens. DeepSWE (72.7% vs Fable 5's 69.7%) and GeneBench v1 (30% vs GPT-5.5's 22%) showed the same shape.
The through-line survives the quarter: the scores pile up in long-horizon, tool-using work where finishing matters more than any single answer, and the +54% token-efficiency claim over GPT-5.5 kept showing up as the same score for fewer tokens. The weak spot survives it too. SWE-bench Pro read 64.6% at launch, well behind Mythos 5 (80.3%) and Opus 4.8 (69.2%), and nothing in the October snapshot suggests the trade was unwound. If your workload is strict repo bug-fixing scored hard, the July advice holds: Claude's top tier wins that one.
Where the October boards disagree
The awkward part of an otherwise clean quarter: the top of the text board is a dispute. The Artificial Analysis Intelligence Index v4.3, on its September 30 snapshot, puts Claude Opus 5.5 first. Nine days later, the October 9 llm-stats relay puts GPT-5.6 Sol first at 0.590. Different boards, different scales, snapshots nine days apart — and no dated source that settles the question between them. For calibration, this article's July edition had Sol at #2 on the AA index, about a point behind Claude Fable 5; Anthropic's fall releases reshuffled that order, and the measuring sticks changed mid-argument. Whoever is selling you a settled number one this fall is selling you a subscription. The rule from July still works: quote every rank with its board and its date, never average two calibers into a synthetic crown, and treat any "#1" that arrives without both as decoration. The whole reshuffle, board by board, gets the quarterly treatment in the Q4 model landscape.
Everyone else repriced too
OpenAI cut into a market that was already cutting. Anthropic shipped a three-model refresh in fifteen days: Opus 5.5 on September 22, claimed to reach Fable 5.1's level on most work at 40% lower running cost; Sonnet 5.5 on September 28, claimed 30% faster and 30% cheaper; Haiku 5.5 on October 7, claimed fastest and cheapest small model in the lineup. Every one of those is a vendor percentage with the denominator unpublished, and the Claude 5.5 family deep dive grades that paperwork line by line.
The floor moved further out. A September 22 report on Artificial Analysis's open-source board has Xiaomi's MiMo-V2.6-Pro taking the open-source and domestic top spot at 46 points and a reported $0.13 per task, passing Kimi K3 and Qwen3.8 Max, and lists GLM-5.3-Flash (320B parameters, $0.15/$0.50), Gemini 3.7 Flash ($0.75/$3.75), Grok 4.5 ($2/$6), and MiniMax M2.7 (205K context, $0.30/$1.20) on the same board. In July, this article called Luna, at $1/$6, the cheapest serious agent-tier model in OpenAI's lineup, and that was true. In October, $1/$6 is mid-budget: several entries on the open-source board had already crossed under a dollar before Luna's own 80% cut landed. Dimes are the entry fee now.
Which tier gets which job
Verdict, per tier, as of October 9:
- Sol ($5/$30): agent and terminal work, browser research, multi-step workflows, anywhere a failed task costs more than the token delta. The launch record is the reason, and the 0.590 says nothing moved.
- Terra ($2/$12): the default for most teams. 0.040 off the flagship at 40% of the input price is, in my book, the best ratio on the card.
- Luna ($0.20/$1.20): classification, extraction, routing, summarizing, any lane where volume dwarfs stakes. At a fifth of July's price it is the cheapest way into this family, and the 0.510 says it stayed in the family.
- Skip the family for SWE-bench Pro-shaped workloads, where Mythos 5 and Fable-class Claude tiers led at launch, and for native multimodal at scale, where Gemini 3.5 Pro was already the purpose-built tool when this article first ran.
How x-rush fits in
One disclosure before this last part: I evaluate models for x-rush.ai, and nothing above is an availability claim. On the platform you never pick GPT-5.6 by name. You open a text feature, describe the job, and each task routes to a model that fits it, with provider names kept out of the product vocabulary on purpose. What carries over from a quarter like this one is the billing shape: text generation runs 0.5 credits a request, pooled chat assistance runs 0.5 credits per ten successful replies, engines unnamed and priced identically. A quarter that cut a tier's price by 80% and split the crown across two boards changes none of that. Whether any given lab is in the rotation is an implementation detail the product does not promise, and this review is not a promise either.
How to use it
If you want GPT-5.6 itself, that is the API and ChatGPT's tiers, token-billed, effort dial in hand. If you would rather describe the job than manage a dial, AI Chat on x-rush takes the brief and quotes the credit price; the engine question stays internal.
Price the job before you pick the tier.
Or let the job pick its own model, and read the price cards for sport.