AI Writing & Research

GPT-5.6 Explained: OpenAI's Sol, Terra and Luna Models

OpenAI launched GPT-5.6 on 9 July 2026 as three tiers—Sol, Terra and Luna—with new pricing and an agent-coordinating 'ultra' mode. Here's what changed.

Editorial noteThis news analysis is based on the linked primary sources. Performance and product claims are attributed to the announcing vendor unless the article explicitly says they were independently tested.

OpenAI released GPT-5.6 on 9 July 2026, and for the first time split its flagship line into three separate, permanently maintained models rather than one general-purpose release [1]. The family comprises Sol, described as the flagship; Terra, pitched as "a balanced model for everyday work"; and Luna, "the most cost-efficient model" of the three [1]. OpenAI says the three are "durable capability tiers that can advance on their own cadence," meaning each will be updated on its own schedule rather than being replaced wholesale by the next version number [1].

The launch also introduced an "ultra" setting, which coordinates multiple agents working in parallel on a single demanding task rather than running one model thread. OpenAI has since revised pricing twice: on 30 July 2026 and again on 21 August 2026, according to the company's own announcement page and its newsroom index [1][2].

What OpenAI actually announced

Each tier targets a different budget and workload rather than a different skill set. Sol is aimed at the hardest reasoning, coding and research tasks; Terra sits in the middle for day-to-day drafting and analysis; Luna is built for high-volume, low-complexity work such as classification or short-form generation, where cost per call matters more than peak capability.

Launch pricing, before the cuts

At launch, OpenAI listed the following per-million-token API prices [1]:

  • Sol: $5 input / $30 output
  • Terra: $2.50 input / $15 output
  • Luna: $1 input / $6 output

The "ultra" setting

Rather than a fourth model, "ultra" is a mode that OpenAI says coordinates several agents in parallel on one task, intended for problems where a single reasoning pass isn't sufficient. OpenAI frames this as central to GPT-5.6's positioning around agentic, multi-step work rather than single-turn chat.

The performance claims — and who is measuring them

OpenAI's own figures, published on its GPT-5.6 announcement page, credit Sol with a score of 80 on the Artificial Analysis Coding Agent Index, which the company says it reaches "using less than half the output tokens" and at roughly a third of the cost of a rival model it names as "Claude Fable 5" [1]. On a benchmark OpenAI calls Agents' Last Exam, the announcement's narrative reports a 53.6% score and a 13.1-point lead over that comparison model [1]. The company also reports a 73.5% score on a cybersecurity benchmark it calls ExploitBench, up from 47.9% for GPT-5.5, and a 28.7% score on a science benchmark it calls GeneBench Pro [1]. These are vendor-reported results rather than tests reproduced by Next AI Compare.

Cross-vendor benchmark comparisons deserve extra care. The material we reviewed does not fully explain every comparison-model setting, and Next AI Compare has not rerun the tests. Treat the figures as directional vendor evidence until a neutral third party reproduces them under comparable conditions.

Pricing has already moved twice

GPT-5.6's list prices are not the ones OpenAI is currently charging. On 30 July 2026, OpenAI published "Advancing the price-performance frontier with GPT 5.6," cutting Luna's price by 80% and Terra's by 20%, according to the headline and date listed on OpenAI's own newsroom index [2]. The company's GPT-5.6 page states that on 21 August 2026 it also reduced Sol's API and credit pricing by "over 20%" [1]. Anyone comparing current GPT-5.6 costs against other frontier models should check OpenAI's live pricing page rather than the July launch figures above, which are now out of date.

Why this matters

Splitting a flagship model into three tiers, each on its own update schedule, is a change in how OpenAI plans to compete — not just a version bump. It lets the company chase frontier benchmark scores with Sol while separately optimising Luna purely on cost, rather than forcing every user onto whichever trade-off the single "best" model happens to make. Two price cuts inside six weeks of launch also suggest OpenAI is treating per-token cost as a competitive lever in real time, likely in response to rival releases such as Anthropic's Claude Opus 5 (24 July 2026) and xAI's Grok 4.6 (12 August 2026).

Who should care

Developers building on the OpenAI API need to pick a tier deliberately rather than defaulting to "the newest model," since Sol, Terra and Luna are meant to coexist rather than supersede one another. Businesses with high-volume, low-complexity AI workloads — support ticket triage, basic classification, short copy — are the most likely to benefit from Luna's cost cuts specifically. Teams already computing cost-per-task for AI vendor comparisons should treat OpenAI's July launch prices as obsolete and re-check current rates before making a decision.

Practical implications for buyers and users

For a ChatGPT subscriber, the practical change is mostly invisible — model selection inside the product is largely automatic. For API customers and businesses evaluating vendor lock-in, the three-tier structure means a workload can potentially move from Sol to Terra to Luna as it matures and its accuracy requirements become clearer, without changing provider. Anyone budgeting against OpenAI's published benchmark claims should treat them as a starting shortlist for their own testing, not a substitute for it — particularly the cross-vendor comparisons, which use figures OpenAI attributes to competitors' own published system cards rather than tests OpenAI ran itself.

Limitations, availability and unresolved questions

OpenAI says GPT-5.6 rolled out globally starting 9 July 2026 across ChatGPT, Codex and the API, reaching full availability within 24 hours [1]. What isn't public, at least in the material we reviewed, is a detailed methodology for the benchmark comparisons against named rivals, or how "ultra" mode is priced relative to standard usage of the same tier. OpenAI also has not published a retirement or deprecation timeline for GPT-5.5, so it's unclear how long the two generations will run in parallel.

Verdict

GPT-5.6 is a real structural change — three permanently distinct tiers rather than one model with a version number — and the two rapid price cuts since launch show OpenAI is treating pricing as an active competitive weapon rather than a fixed list. The performance claims are worth taking seriously as directional signals, but they come from OpenAI itself, using benchmarks it selected and, in several cases, comparison figures for competitors that OpenAI did not independently verify. Anyone choosing between Sol, Terra and Luna should match the tier to the workload's actual complexity and volume rather than assuming Sol is the right default.

The current AI Writing & Research shortlist

Where this sits in the wider market: our current shortlist for AI Writing & Research, what each tool is best at and the main caution to check before committing.

ToolBest forCurrent positionImportant caution
ChatGPT
Best all-rounder
General writing, analysis and multimodal workGPT-5.6 combines strong reasoning with files, images, tools and broad workflow support. It is the safest starting point when one assistant must cover many jobs.Teams should define data-handling rules and verify important claims.
Claude
Long-form pick
Editorial work, complex documents and careful reasoningClaude’s current Opus and Sonnet 5 family is built for sustained professional and agentic work, with a strong reputation for readable long-form output.The highest-capability tiers can be unnecessary for routine copy.
Gemini
Google ecosystem
Workspace users and multimodal source materialGemini 3.7 Flash, documented in August 2026, is the current Flash release, connecting reasoning, multimodal inputs and Google’s productivity ecosystem.Feature availability varies by Workspace plan and region.
Perplexity
Research pick
Fast web research and cited discoveryPerplexity is useful when the first requirement is finding and comparing live web sources rather than drafting from memory.A citation does not guarantee that the source supports every sentence; open the evidence.
Jasper
Brand governance
Marketing teams with repeatable brand workflowsJasper focuses on governed marketing content, brand context and campaign production rather than being a general-purpose chatbot.Plans, limits and model availability change frequently; confirm the current vendor page before purchasing.
Copy.ai
GTM workflows
Sales and marketing process automationCopy.ai has evolved from a copy generator into a go-to-market workflow platform for repeatable content and sales operations.Plans, limits and model availability change frequently; confirm the current vendor page before purchasing.
Writesonic
AI visibility
SEO content and answer-engine monitoringWritesonic combines assisted content production with tooling aimed at search and AI-answer visibility.Plans, limits and model availability change frequently; confirm the current vendor page before purchasing.
Grammarly
Editing layer
Everyday rewriting, tone and quality controlGrammarly works best as an editing and communication layer across existing applications rather than as the only writing system.Plans, limits and model availability change frequently; confirm the current vendor page before purchasing.
Notion AI
Knowledge workspace
Teams whose documents and projects already live in NotionNotion AI is strongest when it can work inside an existing team knowledge base instead of requiring constant copying between tools.Plans, limits and model availability change frequently; confirm the current vendor page before purchasing.
KoalaWriter
SEO drafts
Structured long-form drafts and niche publishingKoalaWriter remains a focused option for producing structured, search-aware drafts quickly.Human research, original experience and fact-checking are still required before publishing.

Sources and verification notes

Primary product documentation checked for this update: