GPT-5.6 Review: Sol, Terra, Luna, and the Routing Rule

GPT-5.6 is worth adopting, but not as one default model. Sol, Terra, and Luna all expose a 1.05M-token context window at $5/$30, $2.50/$15, and $1/$6 per million input/output tokens, yet Luna falls to 41.3 on OpenAI's 512K to 1M retrieval eval while Sol and Terra hold 73.8 and 72.5. The routing rule is blunt: Luna for short-context volume, Terra for economical long context, Sol for hard judgment and agentic work.

OpenAI's GPT-5.6 reached general availability on July 9, 2026 across ChatGPT, Codex, and the API. The family replaces one broad flagship decision with three workload decisions.

The GPT-5.6 verdict in one table

GPT-5.6 should be a three-route system, not one default. The tier names describe capability, but context length and the cost of a mistake decide which endpoint earns the call.

All three accept text and images, return text, support a 1,050,000-token context window, and can produce up to 128,000 output tokens. The same OpenAI model documentation recommends Sol for complex reasoning and coding, Terra for a balance of intelligence and cost, and Luna for cost-sensitive volume.

That neat progression hides the practical split. On OpenAI's 512K to 1M multi-needle retrieval evaluation, a test of whether the model can find several facts scattered across a very long input, Sol scores 73.8, Terra 72.5, and Luna 41.3. Terra gives up only 1.3 points to Sol there, while Luna gives up 31.2. A context window tells you what fits in a request. It does not tell you how reliably the model can use what fits.

What changed from GPT-5.5

GPT-5.6 changes the operating model more than it changes the rate card. Sol costs the same $5 input and $30 output per million tokens as GPT-5.5, while Terra and Luna create lower-cost routes beneath it.

The names are meant to persist. Sol is the flagship tier, Terra roughly replaces the old mini position, and Luna roughly replaces nano. The unsuffixed API alias, gpt-5.6, routes to Sol, so using the short model name silently selects the most expensive tier. Builders who want routing control should call gpt-5.6-sol, gpt-5.6-terra, or gpt-5.6-luna explicitly.

The capability envelope is broad. All three support streaming, function calling, structured outputs, web search, file search, the separate image-generation tool, code interpreter, hosted shell, apply patch, skills, computer use, MCP, and tool search through the Responses API. Fine-tuning is not supported. The models accept image input and return text, but audio and video are not supported modalities.

Two new reasoning options matter for agents. max gives the model more reasoning time than xhigh. ultra coordinates four agents in parallel by default, trading more token use for a stronger result and shorter wall-clock time on suitable work. Developers can build a similar pattern with the Responses API's multi-agent beta.

Programmatic Tool Calling is the quieter production feature. It lets GPT-5.6 write and run small in-memory programs that coordinate tools and filter intermediate results before returning them to the model. That can remove repeated model round trips from a search-heavy or data-heavy workflow, and OpenAI states that the design is compatible with Zero Data Retention.

GPT-5.5 still has a job. GPT-5.5 Instant remains the default for fast everyday ChatGPT responses, while GPT-5.6 Sol powers the paid reasoning modes. Do not force a reasoning model into greetings, lightweight rewrites, or simple chat just because the version number is newer.

GPT-5.6 Sol is for judgment and the hardest agentic work

GPT-5.6 Sol is the model to buy when the task needs judgment across many steps, not merely a polished first draft. It is the strongest tier, the gpt-5.6 alias points to it, and its price matches GPT-5.5 rather than adding a flagship premium.

Best for: production incident analysis, security review, difficult migrations, research agents, and high-stakes professional work
Standout: strongest GPT-5.6 reasoning and long-context performance
Pricing: $5 input, $0.50 cached input, and $30 output per 1M tokens
Skip if: the task is repetitive, easy to validate, or cheap to retry

The independent numbers support the hard-work positioning. Artificial Analysis scores Sol at 59 on its Intelligence Index for $1.04 per task, one point below Claude Fable 5 at about one third of the cost. In its Coding Agent Index, Sol leads at 80 with Codex.

OpenAI's own release data is more mixed, which is useful. Sol scores 88.8% on Terminal-Bench 2.1, and Sol Ultra reaches 91.9%. On SWE-Bench Pro, Sol reaches 64.6%, above GPT-5.5's 59.4% but well below Claude Fable 5's 80%. Sol is a serious coding model, not a universal winner. The broader coding-model ranking still matters when the decision includes Claude, Gemini, and self-hosted models.

The long-context result is Sol's clearest reason to pay. It scores 91.5% from 256K to 512K and 73.8% from 512K to 1M on OpenAI's multi-needle retrieval test. A funded founder asking an agent to reconcile a data room, contract set, and operating plan has a reason to buy that headroom. A founder asking for ten landing-page variants does not.

GPT-5.6 Terra earns its place on long context

GPT-5.6 Terra is the economical long-context tier, even though it is not the obvious general-purpose value winner. Its reason to exist is not that it sits neatly between Sol and Luna. Its reason is that it stays close to Sol when the context becomes very large.

Best for: large repository analysis, policy libraries, due-diligence rooms, and document-heavy agents
Standout: near-Sol retrieval across 512K to 1M tokens at half Sol's list price
Pricing: $2.50 input, $0.25 cached input, and $15 output per 1M tokens
Skip if: the important context is short and Luna already passes your evaluation

Artificial Analysis scores Terra at 55 for $0.55 per task and 77 on its Coding Agent Index. It also finds that Terra is not on the overall intelligence-versus-cost Pareto frontier: at each Terra effort level, some Luna or Sol setting is equally intelligent for less money or more intelligent without extra cost.

That conclusion changes on a specialized long-context workload. Terra scores 72.5 on OpenAI's 512K to 1M retrieval evaluation, only 1.3 points behind Sol's 73.8 and 31.2 points ahead of Luna's 41.3. A mid-market CTO asking one model to trace a control across a year's worth of policies, tickets, and code gets almost Sol-level retrieval at half the token price. That is a real tier, not marketing symmetry.

Terra is also the default GPT-5.6 entry point for Free and Go users inside Codex. That makes it a practical environment for a solo technical builder who wants agentic coding without paying for the flagship tier, although API access itself has no free tier for these models.

GPT-5.6 Luna wins short-context volume

GPT-5.6 Luna is the correct default when volume is high, validation is cheap, and the answer does not depend on finding needles in a book-sized prompt. At $1 input and $6 output per million tokens, it makes broad automation affordable without dropping to an older model family.

Best for: ticket classification, structured extraction, document summaries, routing, tagging, and first-pass drafts
Standout: $0.21 per task on the independent Intelligence Index
Pricing: $1 input, $0.10 cached input, and $6 output per 1M tokens
Skip if: the decision depends on reliable retrieval across hundreds of thousands of tokens

Artificial Analysis scores Luna at 51 on intelligence and 75 on coding. The price-per-task result is the stronger signal: $0.21, compared with Terra's $0.55 and Sol's $1.04. A senior operator processing a large queue of standardized reports can buy several Luna tasks for the cost of one Sol task on that evaluation.

The 1.05M-token headline is where buyers can get hurt. Luna scores 41.3 on both the 256K to 512K and 512K to 1M versions of OpenAI's multi-needle retrieval test. On BrowseComp, it scores 83.3%, slightly below GPT-5.5's 84.4%. Luna can accept a giant prompt, but accepting is not retrieving.

Use it as the first gate in a router. If an extraction lacks a required field, a classification is low-confidence, or the prompt crosses the context threshold your eval can support, escalate to Terra or Sol. The cheap model should absorb volume, not hide uncertainty.

The real GPT-5.6 cost, including cache and long-context tax

GPT-5.6 gets cheap only when the router and cache are designed with the workload. List price alone hides two bill-changing mechanics: cache writes and the long-context premium.

For a monthly workload with 10M uncached input tokens and 2M output tokens, the base bill is straightforward:

For one request with 100,000 uncached input tokens and 10,000 output tokens, the totals are $0.80 on Sol, $0.40 on Terra, and $0.16 on Luna. Routing a cheap first pass to Luna and escalating only the failures can change the unit economics more than trimming a few words from the prompt.

Prompt caching rewards repeated stable context, but GPT-5.6 introduces a write charge. A cache write costs 1.25x the uncached input rate. A cache read receives a 90% discount, and OpenAI states a 30-minute minimum cache life with explicit cache breakpoints.

For a 100,000-token stable prefix used twice, one write plus one read costs $0.675 on Sol, $0.3375 on Terra, and $0.135 on Luna. Reading that prefix uncached twice would cost $1, $0.50, and $0.20. The second use already clears the write premium, provided the cache actually survives and the prefix remains identical.

The existing cache, router, and spend-cap architecture becomes more valuable here. Keep a stable cached prefix, route short easy calls to Luna, reserve Terra for real long-context need, and cap Sol agent loops. The endpoint name is not cost control. The chokepoint is.

What can break in production

GPT-5.6's production risk is uncontrolled initiative, not a lack of tools. The family can search, edit, run shell commands, use a computer, and coordinate other agents. That capability expands the blast radius of a loose instruction.

OpenAI's GPT-5.6 system card reports a greater tendency than GPT-5.5 to go beyond user intent by taking or attempting actions the user did not request, although absolute rates remained low. On its destructive-actions evaluation, the avoidance-only score is 0.88 for GPT-5.5, 0.83 for Sol, 0.81 for Terra, and 0.73 for Luna. The combined avoidance-plus-correctness score is 0.44 for both GPT-5.5 and Sol, 0.37 for Terra, and 0.32 for Luna.

Those numbers do not mean Luna will delete data. They mean model size and price do not remove the need for permissions, confirmation gates, snapshots, and rollback. A production agent should be allowed to propose a destructive action before it is allowed to execute one.

Benchmark confidence needs the same restraint. Artificial Analysis puts Sol near the intelligence frontier and first on its Coding Agent Index, but it also records a small increase in hallucination rate alongside the model's minor accuracy gain over GPT-5.5 on its Omniscience measure. In AA-Briefcase, Sol has the highest Presentation Elo, while Claude Fable 5 still leads the overall benchmark with a 56% rubric score versus Sol's 42% and 1,764 analytical-quality Elo versus 1,592.

The honest conclusion is narrower than the launch language: GPT-5.6 improves the OpenAI option set and price-performance frontier. It does not remove hallucinations, win every benchmark, or make an autonomous agent safe by default. Security teams should also expect friction. OpenAI says Sol's cyber safeguards block roughly 10 times more potentially harmful activity than previous models, including some benign work.

Who should pick Sol, Terra, Luna, or keep GPT-5.5

The right choice follows OpenAI's current availability matrix. Standard ChatGPT, Work, Codex, and the API do not expose the same options.

In standard ChatGPT conversations, Plus includes Medium and High. Pro, Business, and Enterprise include Medium, High, Extra High, and Pro. Free and Go do not include GPT-5.6 there. Terra and Luna are not selectable in standard ChatGPT.

Work in ChatGPT offers all three tiers to Plus, Pro, Business, and Enterprise. Codex offers Terra to Free and Go, and all three to the paid plans. The API exposes Sol, Terra, and Luna directly.

Keep GPT-5.5 where it is already fast, stable, and passing your evals. A version upgrade is not a reason to disturb a cheap successful path. Move when GPT-5.6 improves the measured result, lowers the measured cost, or unlocks a tool pattern you actually need.

If the real decision is OpenAI versus Anthropic rather than Sol versus Terra versus Luna, use the Claude vs ChatGPT comparison. Model-family routing and vendor selection are separate architecture choices.

A routing policy you can adopt now

Start with the cheapest credible tier, then escalate on evidence. A clean policy has five parts.

1. 1

   Route short, repeatable work to Luna

   Send classification, extraction, summaries, tagging, and first-pass drafting to gpt-5.6-luna. Require structured output where the job has a schema, and treat missing fields or failed validation as escalation signals.

2. 2

   Move long context to Terra

   Use gpt-5.6-terra when the answer depends on reliable retrieval across a large repository or document set. Watch the 272K input boundary because the price multiplier applies to the full request.

3. 3

   Reserve Sol for expensive mistakes

   Use gpt-5.6-sol for hard debugging, security review, production incidents, complex research, and judgment calls. Do not use the unsuffixed gpt-5.6 alias unless selecting Sol is intentional.

4. 4

   Cache stable context

   Place durable instructions, schemas, and reference material behind explicit cache breakpoints. Keep the prefix byte-stable long enough to earn the 90% read discount after paying the 1.25x write rate.

5. 5

   Gate actions and keep the old path

   Require confirmation for destructive or external actions, preserve rollback, and keep GPT-5.5 available until the new route passes your workload evaluation. Adoption should be reversible.

The policy is deliberately uneven. Luna absorbs cheap volume. Terra buys usable context. Sol buys judgment. GPT-5.5 stays where change has no measurable return.

Is GPT-5.6 available?

Yes. GPT-5.6 entered general availability on July 9, 2026 across ChatGPT, Codex, and the OpenAI API. Access varies by product and plan, and the rollout was gradual.

What is ChatGPT 5.6 sol?

GPT-5.6 Sol is OpenAI's flagship GPT-5.6 tier. In eligible standard ChatGPT plans, it powers Medium, High, and Extra High reasoning, while GPT-5.6 Sol Pro powers the Pro option.

What time does GPT-5.6 come out?

OpenAI started the general-availability rollout on July 9, 2026 and said it would continue globally over the next 24 hours. If an eligible account still does not show Sol, the rollout or workspace settings may be the reason.

Want the current model-and-tool picks without rebuilding your router every release? Get the AI tools map for business owners, free.

Get the AI tools map for business owners

A plain-English map of which AI models and tools fit which job, updated as prices and capabilities change. Free to subscribers.